Sustained use of generative AI in organizations

Evidence dossier: From scattered GenAI experiments to sustained organizational adoption

Research snapshot: October 3, 2026

The evidence does not support treating “organizational adoption of generative AI” as a single state, nor as a reliably linear maturity ladder. The literature and current statistics are measuring several different phenomena under the same label: employee exposure, individual use, frequency, employer provision, team diffusion, workflow embedding, organizational governance, persistence, and realized outcomes. A firm can be high on one dimension and low on another—for example, widespread unsanctioned employee use with little workflow integration, or centrally licensed and governed AI with low active usage. 1

Across the strongest evidence, the recurring mechanism is not simply “more use.” Sustained useful adoption appears to depend on a joint fit among task, AI capability, worker expertise, verification cost, organizational permission/support, and the surrounding workflow. The same technology can substantially raise performance on one task and lower it on another; employer encouragement and tool provision are strongly associated with use, while generic training by itself is a much weaker predictor; and task-level productivity gains do not automatically propagate into lower labor costs, higher earnings, shorter hours, or firm-level performance. 2

The most defensible synthesis is therefore multidimensional rather than stage-based: measure reach, frequency/persistence, depth of workflow integration, institutional support, governance, and outcomes separately.

Scope, evidence and definition map

What “adoption” is currently made to mean

Current studies operationalize adoption at radically different levels. The resulting percentages are often incomparable before any substantive disagreement about behavior is considered. The U.S. Census Bureau asks whether a business used AI in a recent two-week period; Bick, Blandin, and Deming ask workers about recent GenAI use; Eurostat asks whether an enterprise used at least one of several AI technologies; the Federal Reserve small-business survey allows “the business or its employees” to count as current AI use and separately asks whether that use is experimental, partially integrated, or fully integrated. 3

A useful definition map is therefore:

Construct	What would count	What it does not establish
Awareness	Employee knows GenAI exists or knows an approved tool is available.	Any actual use.
Experimentation	Trying a tool, pilot, proof of concept, one-off prompting.	Repeated use, task fit, value or persistence.
Occasional individual use	Sporadic drafting, search, coding, summarization, analysis, etc.	Habitual use or organizational adoption.
Repeated individual use	Recurring use over days/weeks for work tasks.	Team diffusion or workflow change.
Team-level use	Multiple members of a working group use AI or coordinate work through it.	Formal process integration or enterprise support.
Workflow integration	AI becomes a regular step, interface, automated component or handoff in a process rather than an optional side tool.	That the process performs better overall.
Organizationally supported use	Employer provides or approves tools, access, support, training, data connections or policies.	High active utilization; provision can precede adoption.
Governed use	Use is bounded by policies, access controls, data rules, review requirements, monitoring or accountability.	Useful use; governance can coexist with very low usage.
Routinized/sustained use	The behavior persists and becomes an ordinary way particular work is done.	Positive economic outcome.
Value realization	Measured improvement in quality, speed, throughput, cost, customer outcomes, revenue or another target metric.	Broad organizational penetration. A narrow use case can create value.
Work redesign	Tasks, roles, handoffs, decision rights, staffing or processes change because AI is present.	Necessarily more AI use; redesign can reduce unnecessary interactions with AI.

This map is a research synthesis, not a validated universal stage model. The need for separation is directly visible in empirical measurement. In the 2026 Federal Reserve Small Business Credit Survey report, 46% of sampled employer firms said the business or employees currently used AI, yet among these users roughly half characterized themselves as experimenting, 44% as partially integrated, and only 7% as fully integrated. “Use” and “integration” were thus explicitly different constructs in the same instrument. 4

Likewise, Bick, Blandin, and Deming distinguish any recent work use from frequency and intensity: in late 2024, 27% of employed U.S. respondents had used GenAI for work at least once in the previous week, but only 10% used it every workday; their estimates implied that only roughly 1–7% of total work hours were being assisted by GenAI. That is a very different phenomenon from saying that 27% of workers had “adopted AI.” 1

Why a single progression is misleading

At least five empirically plausible paths coexist:

Bottom-up shadow diffusion: individuals discover public tools → repeated personal use → peer diffusion → perhaps later employer approval. Microsoft/LinkedIn’s 2024 provider survey of roughly 31,000 knowledge workers famously reported widespread “bring your own AI,” illustrating how use can precede institutionalization. Because this is a vendor-sponsored knowledge-worker survey rather than a representative business census, it is better evidence for the existence of this path than for its population prevalence. 5

Top-down provision: the organization licenses or exposes a tool before many employees have found a valuable use for it. Availability and licenses therefore cannot be treated as active adoption.

Bounded use-case institutionalization: a narrow workflow becomes deeply embedded and valuable while most of the organization remains unaffected. The customer-support deployment studied by Brynjolfsson, Li, and Raymond is closer to this model: a specific AI assistant was integrated into customer-service work and increased issues resolved per hour by 15% on average, with sharply heterogeneous effects by experience and skill. 2

Governance-first deployment: regulated or risk-sensitive organizations can establish approval, access and review controls before broad use. The European regulatory environment reinforces the distinction between deployment and governance: the EU AI Act is a risk-based legal framework rather than an adoption metric, and Eurostat separately finds legal uncertainty and privacy concerns among organizations that considered but had not implemented AI. 6

Experiment without institutionalization: individuals or teams obtain temporary task benefits that never become a stable organizational process. Experimental studies at BCG and Procter & Gamble demonstrate real task/team effects without demonstrating enterprise-wide routinization. 7

The stages in the research question therefore make sense primarily as dimensions or observable states, not mandatory milestones.

The measurement hierarchy that matters

For reusable research, it is useful to treat commonly reported adoption metrics as increasingly demanding evidence claims:

Measure	Legitimate inference	Common over-interpretation
Tool available / license purchased	Organization has enabled possible use.	“Employees have adopted AI.”
Someone in company uses AI	Presence somewhere.	“The company uses AI operationally.”
Self-reported individual use	Respondent says AI enters some work.	Frequency, quality or value.
Recent/frequent use	Habit/persistence is more plausible.	Workflow integration.
Task-level telemetry	Observable use on specific tasks.	Organizational diffusion.
Active users / seats	Breadth and some frequency.	Meaningful work transformation.
Team diffusion	Collective exposure.	Process redesign.
Workflow penetration	AI is embedded in recurring process steps.	Positive outcomes.
Sustained longitudinal use	Behavior persists over time.	Value realization.
Task outcomes	Local productivity/quality effect.	Firm performance.
Financial/organizational outcomes	Stronger evidence of value capture.	That AI itself caused the outcome absent a causal design.

This hierarchy explains a large part of the apparent confusion in the 2023–2026 evidence base. 3

Current landscape

Individual work use is widespread, but intensity remains much lower than headline adoption rates

The strongest probability-based U.S. worker evidence shows rapid diffusion but a large gap between “has used” and intensive incorporation. Bick, Blandin, and Deming’s nationally representative surveys found that by late 2024 about 45% of U.S. adults aged 18–64 had used GenAI, while 27% of employed respondents reported work use in the previous week; 10% used it every workday and another 17% on some workdays. The authors estimate only 1–7% of work hours were assisted and respondents’ reported time savings amounted to about 1.4% of total work hours. 1

A later cross-country worker study reported that in early 2026 about 43% of U.S. workers reported using GenAI for their jobs, versus roughly 26–36% across Germany, the United Kingdom, France, Italy, Sweden and the Netherlands. Younger, college-educated employees, workers in larger firms and workers in information and communication technology were more likely to use it. Industry, occupation and firm size explained more of the U.S.–European gap than demographics alone. 8

Those two U.S. percentages should not automatically be read as a clean time series: the surveys and reported adoption constructs are not identical. That is exactly the measurement problem this dossier highlights. 1

Firm-level adoption is much lower under stricter business measures

The Census Bureau’s Business Trends and Outlook Survey is nationally representative of U.S. employer businesses outside farming. Between December 14, 2025 and May 3, 2026, overall business AI use under its recent-use measure hovered around 17–20%, while 20–23% expected to use AI during the following six months. The instrument asks whether the business used AI in the previous two weeks. 3

Crucially, the Census changed its wording in November 2025. Earlier versions asked about AI “in producing goods and services,” explicitly excluding simple activities such as drafting an email; the newer version asks about AI “in any business function.” The series therefore contains a definition discontinuity that can mechanically alter an apparent adoption trend. 3

Size matters strongly in that source: by the period ending May 3, 2026, roughly 37% of firms with at least 250 employees and 32% of firms with 100–249 employees reported AI use—substantially above the economy-wide rate. 3

European enterprise data show similarly strong size, sector and geography gradients

Eurostat reports that 19.95% of EU enterprises with at least 10 people employed used at least one AI technology in 2025. The rate was 17% for small enterprises, 30.36% for medium enterprises and 55.03% for large enterprises. These are broad AI figures, not GenAI-only figures. 9

Sector differences are enormous: 62.52% of information-and-communication enterprises used at least one included AI technology, versus 40.43% in professional, scientific and technical activities and 10.79% in construction. Country rates ranged from 42.03% in Denmark, 37.82% in Finland and 35.04% in Sweden to 5.21% in Romania, 8.36% in Poland and 8.55% in Bulgaria. 9

Eurostat also permits a closer look at generative functions. In 2025, 11.75% of enterprises reported text-mining technologies, 9.55% technologies generating pictures/video/audio and 8.76% technologies generating written or spoken language or programming code. Thus the 19.95% headline cannot legitimately be described as “20% of EU firms use GenAI.” 9

Small-business surveys illustrate how broad wording can produce much higher estimates

The Federal Reserve’s 2026 Report on Employer Firms, based on 6,525 small employer firms responding to the Small Business Credit Survey in September–November 2025, found 46% saying that their business or employees currently used AI and another 15% planning to do so within 12 months. But this is a nationwide convenience sample, not a probability sample, and its respondent-level construct is broader than the Census two-week business-use measure. 4

Among its AI-using firms, writing and marketing were reported by 83%, individual productivity by 61%, and planning/analysis by 51%. Yet approximately half were still “experimenting,” 44% were partially integrated and only 7% described AI as fully integrated into business processes. 4

That combination—high “current use,” much lower “full integration”—is one of the clearest empirical reasons not to collapse experimentation, employee use and organizational adoption into one variable. 4

Mechanism map

The evidence is better understood as interacting mechanism clusters than as an unranked list of drivers.

Mechanism cluster	What the evidence supports	Important interaction	Evidence status
Task fit × model capability	AI can produce large gains on tasks within its effective capability region and meaningful harm on superficially similar tasks outside it. In the BCG experiment, AI users completed 12.2% more within-frontier tasks and did them 25.1% faster; on a deliberately outside-frontier task, AI users were 19 percentage points less likely to reach the correct answer. 7	“Usefulness” is not intrinsic to the tool; it depends on task/model fit. A workflow containing both kinds of task can have positive and negative effects simultaneously.	Strong causal evidence at task level. Limited generalizability to all jobs.
Task fit × expertise	Customer-support AI raised issues/hour by 15% on average, but less experienced/lower-skilled workers gained more, while the most experienced workers had small speed gains and small quality declines. 2	AI can transfer expert-like practices to novices while adding less—or even imposing costs—for experts who already possess relevant tacit knowledge.	Strong field evidence in one occupational setting.
Expertise × context × verification burden	In METR’s RCT, 16 experienced open-source developers completed 246 real tasks in repositories they knew well. Allowing early-2025 AI increased completion time by 19%, although developers had predicted a 24% speedup and afterward still believed AI had saved roughly 20%. 10	Familiarity with the codebase reduced the relative value of generated assistance, while reviewing, steering and correcting AI could consume the apparent production gains. Perceived usefulness therefore need not equal measured productivity.	Causal but narrow; point-in-time models, small sample, experienced developers.
Output quality × trust calibration	On BCG’s outside-frontier task, AI reduced objective correctness while still improving the subjective coherence/persuasiveness of responses. 7	Fluent output can simultaneously raise perceived usability and raise the verification problem; high-quality presentation is not a reliable signal of correctness.	Strong causal evidence for the experimental task.
Employer permission/support × access	In early-2026 cross-country data, employer encouragement and provision of AI tools were among the strongest correlates of work adoption and statistically accounted for a substantial part of U.S.–European differences. Formal AI training was much less predictive once these enabling conditions were considered. 8	Skills cannot produce work adoption if employees lack legitimate access or believe use is discouraged. Conversely, licenses without task relevance need not create useful adoption.	Strong association, not causal identification.
Management environment × support	A one-standard-deviation increase in a management-practice index was associated with roughly 9.6 percentage points higher AI adoption in the same cross-country study, but the additional association largely disappeared after employer encouragement and tool provision were included. 8	“Good management” may operate through concrete enabling behaviors rather than through an abstract cultural effect.	Cross-sectional; mediation is suggestive, not causal.
Informal diffusion × governance boundary	Provider evidence shows that substantial employee use can arise through personal tools before an employer has standardized access; Microsoft/LinkedIn’s 2024 knowledge-worker survey reported widespread bring-your-own-AI behavior. 5	Restricting approved access can suppress observable sanctioned usage while shifting behavior into shadow use rather than eliminating it.	Provider survey; useful signal, weak population benchmark.
Risk/legal/privacy × willingness to institutionalize	Among EU enterprises that had considered but not adopted AI, Eurostat reports lack of relevant expertise as the most common obstacle (70.89%), followed by unclear legal consequences (52.52%) and data-protection/privacy concerns (48.83%); only 20.68% cited lack of usefulness. 9	A technically attractive use case may remain individual or experimental when data cannot safely move into the tool or accountability is unclear.	Representative descriptive evidence; no causality.
Usability/friction × task frequency	Repeated use is more plausible when AI is available at the point of work; the customer-support study embedded recommendations into agents’ actual interface rather than requiring a separate experimental workflow. 2	Small per-use friction compounds on high-frequency tasks; conversely, deep integration can amplify both useful and harmful model behavior.	Mechanistically plausible; causal contribution of integration itself not separately isolated.
Skill/AI literacy × task-specific judgment	A short prompt-engineering overview in the BCG experiment did not generate a statistically significant increase in task completion relative to GPT access alone, while skill in navigating which tasks AI could handle mattered to outcomes. 7	Generic prompting competence is different from the harder organizational capability of choosing tasks, supplying context, and detecting failure.	Experimental evidence against treating generic training as sufficient.
Firm size × organizational capability/economics	Census and Eurostat independently show substantially higher AI use in larger firms. 3	Scale can make fixed costs for procurement, governance, data integration and specialist talent easier to absorb, but the surveys do not identify which mechanism causes the size gradient.	Robust descriptive relationship; mechanism unresolved.
Sector/occupation × task composition	ICT, information/communication and professional/scientific work exhibit much higher adoption than construction and many other sectors; younger and college-educated workers also use GenAI more often. 8	Much of what can look like a “culture” or “country” effect may actually reflect different mixes of AI-suitable tasks, occupations, industries and firm sizes.	Strong descriptive evidence.

What is less well established than practitioner discourse suggests

Peer effects are plausible and informal diffusion is plainly observable, but high-quality causal estimates of within-firm peer spillovers remain scarce in the public GenAI literature reviewed here. Firm climate is associated with adoption, but broad terms such as “culture” usually bundle permission, peer norms, management signals, incentives and perceived job risk together. 1

Generic AI literacy should not be equated with effective use. Current evidence more strongly supports task-specific judgment—knowing what to delegate, what context to supply and when/how to verify—than a claim that more generic training monotonically increases value. 7

Data quality and internal-data availability are heavily emphasized in practitioner accounts but are still poorly isolated as causal adoption determinants. They plausibly become more important as organizations move from public-chatbot experimentation to applications requiring proprietary context, retrieval, process integration or automation; the current representative surveys more directly establish expertise, privacy and legal uncertainty than a clean independent “data quality effect.” 9

Cost and ROI likewise become conceptually more important when firms move from free individual experimentation to enterprise licenses, integration and controls, but public data identifying the elasticity of sustained adoption to price or a specific ROI threshold are thin. In the Federal Reserve small-business survey, prominent practical obstacles included output accuracy and adapting tools to business needs among users, and selecting appropriate tools plus implementation/training time among planned adopters. 4

A compact interaction model

The best-supported relationship can be summarized as:

Potential benefit from AI capability
× fit with the specific task
× worker ability to delegate and verify
× legitimate access / managerial permission
× workflow accessibility and context
− verification, correction and coordination costs
− risk/privacy/legal costs
= local net usefulness

Local net usefulness must then survive repetition over time, team/process interdependencies, and organizational value capture before it becomes sustained organizational adoption.

That equation is an interpretation of the evidence, not an estimated structural model. The experimental literature establishes the first three terms especially strongly; representative surveys establish access, management, firm size, sector and risk correlates; evidence on conversion from local gains to firm performance remains much weaker. 2

Adoption versus value; failure and abandonment

Use can create genuine task value

The experimental evidence leaves little doubt that GenAI can improve meaningful work outcomes under favorable task conditions.

In the 5,172-agent customer-service field study, access to an AI conversational assistant increased issues resolved per hour by 15% on average. Less experienced and lower-skilled workers improved both speed and quality; the authors also found evidence consistent with learning and improved English fluency. 2

In the preregistered BCG study of 758 consultants, AI improved speed, task completion and assessed quality across 18 tasks chosen to lie within GPT-4’s capability frontier. 7

A later field experiment at Procter & Gamble found that individuals working with AI produced solutions of quality comparable to two-person human teams in the studied product-development setting, and that AI broadened participants’ ability to contribute outside their core functional expertise. This is evidence that useful value can emerge at the individual/team level without proving broad organization-wide transformation. 11

Use can also reduce value

The opposite result is equally real. The same BCG study found a 19-percentage-point reduction in correctness on the outside-frontier problem even though AI-assisted answers remained persuasive. 7

METR’s experienced-developer RCT produced a 19% slowdown, directly contradicting both participant expectations and widespread productivity narratives for that particular context. The result is especially important because participants were completing real issues in repositories they knew, not benchmark exercises. 10

The substantive lesson is therefore not “AI improves productivity” or “AI harms productivity”; it is that effect size and even effect sign depend on task, worker, model, context and measurement. 2

Broad use can coexist with little aggregate measurable effect

Bick, Blandin and Deming estimate meaningful but still modest work-hour penetration despite rapid individual diffusion: only about 1–7% of work hours were assisted in their late-2024 data, and respondents’ self-reported time savings summed to about 1.4% of total work hours. This reconciles widespread awareness/use with limited aggregate productivity consequences in the early period. 1

Humlum and Vestergaard’s linked Danish employer/worker survey and administrative-record research similarly finds a disconnect between rapid chatbot diffusion and conventional labor-market outcomes. Their 2026 revision describes substantial changes in tasks and work organization while finding no clear corresponding changes in worker earnings or recorded hours over the early post-ChatGPT period. 12

That is not evidence that AI creates no value. Time savings can be absorbed into greater output, higher quality, new tasks or slack rather than wages or hours; organizational adjustment can also lag technical adoption. The Danish evidence shows that use, work reorganization and conventional economic outcomes can move on different timelines. 13

Self-reported value should not be confused with measured value

In the Federal Reserve small-business sample, 71% of AI-using firms reported increased productivity, 39% improved product/service quality and 31% higher sales, while the vast majority reported no reduction in labor costs. These are respondent assessments among current AI users, not randomized or accounting-based treatment effects. 4

The METR result demonstrates why the distinction matters: developers believed AI had made them faster even after completing the experiment, while objective timing showed the opposite. 14

Thus surveys asking “has AI increased your productivity?” and studies measuring task completion time are measuring different constructs and can legitimately produce different answers.

Workflow change is a separate conversion step

A recurring pattern in the evidence is that the easiest early uses—writing, summarizing, marketing, brainstorming or personal productivity—can be overlaid on an existing workflow without altering the workflow itself. The Federal Reserve survey’s combination of widespread writing/marketing use and low full-integration rates is consistent with this pattern, though it does not itself causally prove that workflows remained unchanged. 4

This distinction matters because a person doing the same process with an extra AI interaction may generate:

gross task savings → minus prompting → minus checking → minus correction → minus handoff/coordination → net workflow effect.

METR provides direct evidence that those downstream costs can overturn the apparent benefit of generation in a complex real-world task. 10

Apparently successful adoption can coexist with the requested failure modes

Phenomenon	State of evidence
Little measurable value despite substantial use	Supported. Danish evidence shows rapid adoption/reorganization without clear early earnings/hours effects; Bick et al. show broad use but limited total work-hour penetration. 13
Verification/rework burden	Supported in bounded settings. METR finds net slowdown among experienced developers; BCG shows attractive but objectively wrong AI-supported reasoning outside the capability frontier. 15
Employee workarounds / shadow AI	Supported descriptively. Microsoft/LinkedIn provider research documented extensive use of personally sourced AI among knowledge workers. 5
Fragmented individual practices	Supported indirectly. High employee use can coexist with low formal integration; small firms report much experimentation and partial integration. 4
Unused paid licenses	Unresolved in strong public evidence. Seat purchase, activation and task-level value are rarely available together in independent datasets.
Duplicated work	Partially supported as rework, not cleanly measured as duplication. Review/correction is observable in developer work; organization-wide duplicated effort is not well quantified. 10
Skill degradation	Unresolved longitudinally in workplaces. Existing evidence establishes learning in some settings and dependence/critical-thinking concerns in others, but durable deskilling has not been convincingly measured across organizations. The customer-support study actually finds evidence of learning, illustrating the possibility of augmentation rather than degradation. 2
Increased workload	Evidence exists mainly in self-report/practitioner surveys; causal organizational evidence is limited. Time saved need not translate into reduced workload because organizations can raise output expectations or create new AI-related tasks. Danish evidence documents new/reorganized tasks without reductions in recorded hours. 13
Unchanged underlying workflows	Highly plausible and observable through low integration measures, but rarely measured directly. “Using AI” usually does not reveal whether process architecture changed. 4

Why adoption stalls, stays superficial or disappears

The evidence points to several different failure processes rather than one generic “pilot problem.”

No repeatable task advantage. Experimentation discovers that AI does not improve the relevant task or that the frontier shifts unpredictably across steps. This is the strongest causally supported failure mechanism. 7

Gross benefit is consumed by verification. Generation becomes cheaper but checking, correcting and integrating outputs does not; the bottleneck simply moves. 7

The individual benefit never becomes a process benefit. Employees use AI for personal drafting or ideation, while handoffs, approvals, systems and decision rights remain unchanged. The Federal Reserve integration distribution is consistent with a large population in this intermediate state. 4

Legitimate access is missing. Employer encouragement and approved tool provision are strongly associated with adoption; where they are absent, employees can either abstain or move to unsanctioned tools. 8

Risk dominates expected benefit. Among EU firms that considered AI but did not adopt, expertise gaps, unclear legal consequences and privacy/data-protection concerns are much more common explanations than simply perceiving AI as useless. 9

Capability and organizational change proceed at different speeds. Model performance can change faster than workflows, policies, evaluation methods and employee habits. METR explicitly describes its 2025 result as a point-in-time result because tool capability was evolving rapidly. 10

Evidence on actual abandonment is surprisingly weak. Most surveys ask current use or future intention rather than following the same employees, workflows, licenses or projects through initiation, repeated use and cessation. Claims about very high “pilot failure” or abandonment rates therefore deserve more skepticism than they often receive; the publicly verifiable evidence is much stronger on barriers and shallow integration than on a universal abandonment rate.

Contradictions and measurement reconciliation

Many headline contradictions dissolve after the unit, denominator and definition are made explicit.

Apparently conflicting findings	Reconciliation
Census: roughly 17–20% of U.S. employer businesses use AI vs. Federal Reserve small-business survey: 46% currently use AI. 3	Not the same measure. Census is representative and asks whether the business used AI in a recent two-week window; the Fed survey is a convenience sample and allows either the business or employees to constitute current use. The Fed instrument also captures many firms still experimenting.
Eurostat: 19.95% of EU enterprises use AI vs. far higher worker-use estimates. 9	Enterprise-level use and worker self-use have different denominators. One employee using a public tool need not imply that an enterprise reports formal AI use; Eurostat also covers AI broadly rather than GenAI only.
GenAI produces 15–25%+ productivity gains in major experiments vs. METR finds a 19% slowdown. 2	A genuine contextual conflict, not merely bad measurement. Customer support, consulting tasks and mature-codebase engineering have different task structure, worker expertise, feedback loops and verification costs. “AI effect” is heterogeneous.
Less experienced workers benefit most in customer support vs. elite consultants also benefit strongly on many tasks. 2	Not contradictory. Prior expertise interacts with task/model fit. Experts can gain on tasks where AI supplies useful parallel capability but gain little or suffer when their own tacit/contextual knowledge already dominates the model.
Workers/firms report productivity gains vs. measured hours, earnings or task time sometimes show little gain or deterioration. 4	Different outcomes and methods. Perception, task throughput, quality, labor costs, hours and earnings are not interchangeable. Selection into use also biases cross-sectional self-reports.
Management quality is strongly associated with adoption vs. formal training is not the dominant predictor. 8	The management association appears to operate substantially through concrete encouragement and tool provision; broad management quality or training may be proxies unless the enabling mechanism is measured.
AI improves answer quality vs. AI increases error risk. 7	Both occurred in the same BCG experiment. The task frontier was jagged: performance improved within it and correctness deteriorated outside it.
Governance inhibits adoption vs. governance enables adoption. 9	Both are plausible. Legal/privacy uncertainty can prevent deployment; clear approved access can make employees more willing/able to use AI. Current evidence is stronger on each association separately than on the causal effect of “more governance.”

One particularly important statistical contradiction is created inside the Census series itself: changing the question from AI used in producing goods/services to AI used in any business function changes what qualifies as adoption. A discontinuity caused by survey wording can easily be mistaken for a behavioral acceleration. 3

The general rule for interpreting any adoption statistic is therefore to recover at least:

unit of analysis + population + AI definition + task scope + recall window + frequency threshold + sanctioned/unsanctioned status + integration threshold + outcome measure.

Without those fields, “X% adoption” is not a sufficiently specified fact.

Claim-to-source table

The table below records the most load-bearing claims rather than every source inspected. “Supported” means the cited evidence directly supports the narrowly worded claim; “conditional” means it holds for a specified population/context or the inference is constrained; “unresolved” means the available evidence is materially inadequate.

Exact claim	Evidence type	Population/context	Source date	Inspected	Direct URL / DOI	Assessment	Main limitation
In late 2024, 27% of employed U.S. respondents used GenAI for work in the previous week, 10% every workday; only about 1–7% of work hours were AI-assisted. 1	Nationally representative repeated survey; published peer-reviewed paper	U.S. adults 18–64 / employed subset	Published online Jan. 20, 2026	2026-10-03	https://doi.org/10.1287/mnsc.2025.02523	Supported	Self-reported usage/time; technology evolved after measurement.
Census business AI use was roughly 17–20% over Dec. 2025–May 2026 under a recent-use business measure. 3	Official representative business survey	U.S. employer businesses excluding farms	May 26, 2026	2026-10-03	U.S. Census Bureau, Business Trends and Outlook Survey (official source accessible through citation)	Supported	Questionnaire changed in Nov. 2025; broad AI rather than GenAI-only.
The Census question changed in Nov. 2025 from AI used in producing goods/services to AI in any business function, creating a measurement discontinuity. 3	Official survey documentation	U.S. employer businesses	2026	2026-10-03	U.S. Census Bureau BTOS AI-use documentation	Supported	Does not quantify exactly how much of subsequent change is wording-driven.
In the 2026 Federal Reserve small-business report, 46% said their business or employees currently used AI; among users roughly half were experimenting, 44% partially integrated and 7% fully integrated. 4	Nationwide convenience survey	6,525 U.S. small employer firms; fielded Sep.–Nov. 2025	2026	2026-10-03	Federal Reserve Banks, 2026 Report on Employer Firms	Conditional	Convenience—not probability—sample; broad “business or employees” definition.
19.95% of EU enterprises with 10+ workers used at least one AI technology in 2025; 55.03% of large enterprises did so. 9	Official harmonized enterprise statistics	EU enterprises, 10+ employed/self-employed	Data extracted Dec. 2025	2026-10-03	https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Use_of_artificial_intelligence_in_enterprises	Supported	Measure is broad AI, not specifically GenAI.
Among EU enterprises that considered but had not adopted AI, lack of expertise, unclear legal consequences and privacy/data-protection concerns were leading reported barriers. 9	Official enterprise survey	EU enterprises that considered AI but did not use it	2025	2026-10-03	https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Use_of_artificial_intelligence_in_enterprises	Supported descriptively	Self-reported reasons; cannot establish causal effects.
Access to a GenAI assistant increased customer-support issues resolved/hour by 15% on average; benefits were larger for less experienced/lower-skilled agents. 2	Staggered workplace deployment / field empirical study	5,172 customer-support agents	Feb. 4, 2025	2026-10-03	https://doi.org/10.1093/qje/qjae044	Supported	One occupation/company setting; specific assistant and period.
On 18 BCG tasks inside GPT-4’s capability frontier, AI users completed 12.2% more tasks and worked 25.1% faster. 7	Preregistered randomized field experiment	758 BCG consultants	Published online Mar. 11, 2026	2026-10-03	https://doi.org/10.1287/orsc.2025.21838	Supported	Elite knowledge-worker sample; experimental tasks and GPT-4-era technology.
On an outside-frontier BCG task, AI users were about 19 percentage points less likely to be correct, even while AI improved subjective coherence. 7	Randomized experiment	BCG consultants	Mar. 11, 2026	2026-10-03	https://doi.org/10.1287/orsc.2025.21838	Supported	Only one designated outside-frontier task.
Early-2025 AI tools made experienced developers working in familiar mature repositories 19% slower despite users believing AI had made them faster. 10	Randomized controlled trial	16 experienced open-source developers, 246 tasks	Jul. 10–12, 2025	2026-10-03	https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ ; https://doi.org/10.48550/arXiv.2507.09089	Conditional	Small, specialized sample; early-2025 models; authors explicitly call it a point-in-time estimate.
Employer encouragement and employer-provided tools are among the strongest correlates of worker GenAI adoption across the U.S. and six European countries. 8	Cross-country worker survey	Workers in U.S., Germany, UK, France, Italy, Sweden, Netherlands	Apr. 14, 2026 analysis of early-2026 survey	2026-10-03	Federal Reserve Bank of St. Louis analysis of Bick et al. cross-country study	Conditional	Cross-sectional associations; authors warn against causal interpretation.
Much of the cross-country association between management quality and AI adoption is accounted for by employer encouragement and tool provision. 8	Cross-sectional multivariate analysis	Same cross-country worker population	2026	2026-10-03	Federal Reserve Bank of St. Louis analysis	Conditional	Statistical accounting is not causal mediation.
AI can reproduce some collaborative benefits: in the P&G field experiment, individuals with AI produced solutions comparable in quality to two-person teams and broadened contributions beyond core expertise. 11	Organizational field experiment	Procter & Gamble professionals / product-development tasks	2026 publication	2026-10-03	https://doi.org/10.1287/orsc.2025.20702	Conditional	Specific innovation-task design; does not demonstrate firm-wide adoption or financial impact.
Danish evidence shows rapid GenAI diffusion and work/task reorganization can coexist with little detectable early change in conventional earnings or hours. 12	Linked worker/employer surveys + administrative records	Denmark, AI-exposed occupations/workplaces	Revised 2026	2026-10-03	https://www.nber.org/papers/w33777 ; https://doi.org/10.3386/w33777	Supported for early period	Earnings/hours may lag; absence of those effects is not absence of quality/output value.
Enterprise legal uncertainty and privacy concerns are empirically associated with non-adoption consideration in the EU, while the EU AI Act creates an expanding formal governance environment. 9	Official statistics + regulation	EU enterprises	2025–2026	2026-10-03	https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai	Supported/conditional	Cannot infer that the AI Act itself reduced or increased adoption from these data.
“Use,” “integration,” “governance” and “value realization” should not be treated as synonymous constructs. 4	Cross-study synthesis	Multiple populations	Through 2026	2026-10-03	Multiple sources above	Supported as interpretation	A conceptual synthesis rather than one validated scale.
A universal sequential GenAI-adoption maturity ladder is not currently established by empirical evidence. 4	Evidence synthesis	Firms/workers/teams across studies	Through 2026	2026-10-03	Multiple sources above	Unresolved / no support for universality	Absence of a validated universal ladder does not exclude useful sector-specific maturity models.
High paid-license counts reliably predict workflow integration or value.	No convincing independent evidence identified	Enterprise GenAI	Through 2026	2026-10-03	—	Unresolved	Independent seat-level longitudinal utilization and outcome data are rarely public.
Sustained workplace GenAI use causes long-term deskilling.	Insufficient longitudinal workplace evidence identified	Knowledge work	Through 2026	2026-10-03	—	Unresolved	Short-run dependence/learning evidence cannot establish durable human-capital effects.

Unresolved questions and research frontier

Questions the current evidence cannot answer well

Persistence is undermeasured. Most datasets are repeated cross-sections or short experiments. We know much more about whether someone used AI recently than whether the same person, team and workflow continues using it six, twelve or twenty-four months later. Bick et al. improve frequency measurement, but broad longitudinal employee/workflow panels remain rare. 1

Actual abandonment rates are not known with high confidence. Public discourse often cites pilot-failure statistics, but representative evidence following identified GenAI projects from initiation through cessation, renewal or institutionalization is scarce. Current official surveys measure adoption status and obstacles much better than abandonment trajectories. 3

License utilization is largely a proprietary-data problem. Independent researchers rarely observe purchased seats, monthly/daily active use, task composition, prompt/output telemetry, renewal decisions and economic outcomes for the same organization. Consequently, “unused paid licenses” is plausible but cannot yet be assigned a reliable cross-industry prevalence from the strongest public evidence.

Team diffusion is a major missing middle. Research is comparatively rich at the individual/task level and increasingly rich at the enterprise-survey level, but much thinner on how use moves between coworkers, teams and functions. The P&G experiment demonstrates that AI changes the economics of individual-versus-team work, but not how such patterns diffuse longitudinally inside organizations. 11

Governance’s causal effect is unknown. Legal uncertainty and privacy concerns inhibit contemplated adoption, while approved tools and managerial encouragement correlate with greater use. What remains unresolved is the shape of the trade-off: which controls create trust and sanctioned access, which create prohibitive friction, and where the optimum lies by task risk. 9

There is no settled measure of workflow penetration. “Percent of employees using AI” is not equivalent to “percent of workflows materially touched by AI,” much less “percent of value-creating process steps redesigned around AI.” Official statistics have not converged on a workflow-level denominator. 3

Verification economics are poorly aggregated. We have compelling examples where review cost overturns generation speed and others where assistance raises throughput. What is missing is a general method for measuring gross AI time saved versus review, correction, exception handling, escalation and downstream defect costs. 2

Human-capital effects are unresolved. Customer-support evidence suggests AI can teach less experienced workers; concerns about deskilling imply the opposite path when AI replaces rather than scaffolds reasoning. Longitudinal evidence measuring human performance after AI is removed is too sparse to determine when sustained use builds versus erodes expertise. 2

The relationship between local productivity and firm performance remains weakly identified. Task experiments produce clear causal estimates, while firm surveys usually produce associations or self-reported outcomes. Bridging these levels is one of the central unanswered empirical problems. 2

Cessation may be rational rather than failure. When a use case is outside the model frontier, carries excessive verification cost or creates legal/data exposure, lower use or abandonment can represent good organizational adaptation. Existing maturity narratives often assume monotonic increases in use are desirable; the experimental evidence gives no basis for that assumption. 7

Developments that could materially change the picture

Agentic and computer-operating systems change the unit of adoption. The relevant unit is moving from “employee prompts chatbot” toward “human delegates a multi-step task or process to an AI agent.” Contemporary tools already expose workflow- and computer-operating capabilities rather than only chat interfaces. Future surveys will therefore need to distinguish human AI use from tasks executed autonomously or semi-autonomously by systems. 16

Model capability is moving faster than much of the evidence cycle. METR explicitly notes that its early-2025 developer result is a point-in-time estimate; the Organization Science frontier paper similarly stresses that the location of the capability frontier changes as models improve. Conclusions about specific task fit have unusually short half-lives even when the underlying mechanism—fit and verification matter—may remain stable. 10

Embedded enterprise context may alter verification economics. Current experiments range from generic frontier models to tools embedded in actual support workflows. Better access to organizational knowledge, structured data, permissions and business systems may increase usefulness, but it can also increase the consequences of incorrect autonomous action. The causal evidence on this transition remains immature. 2

Governance is becoming observable rather than merely voluntary. In Europe, the AI Act and related guidance are creating explicit organizational obligations and stronger incentives to document how AI systems are used and overseen. This may eventually generate better evidence separating sanctioned, governed deployment from shadow use, although current data do not yet demonstrate its net effect on adoption. 6

Longitudinal administrative and telemetry studies are the highest-value next evidence class. Danish linked survey/administrative work demonstrates what becomes visible when reported adoption is connected to actual labor-market outcomes; comparable studies linking enterprise licenses, activity logs, workflow metrics, HR data, quality outcomes and financial performance would resolve many of the current contradictions. 13

Measurement standards will matter as much as better adoption percentages. The research frontier is not simply obtaining a newer estimate of “how many companies use AI.” The more consequential advance would be a measurement system that separately records: sanctioned access, unique active users, frequency, tasks, AI contribution to task time, human verification effort, workflow penetration, persistence, exception rates, quality, and downstream business outcomes. The current evidence base shows why each layer is necessary: headline adoption can rise while intensity stays low, integration remains partial, objective productivity falls for some tasks, or firm-level outcomes remain unchanged. 1

The most robust conclusion from the evidence through October 3, 2026 is therefore not a maturity sequence. It is a separation principle: organizational GenAI adoption should be decomposed into reach, persistence, task fit, workflow depth, institutional support, governance and realized outcomes. Existing evidence supports several pathways between those dimensions, including sustained shadow use, top-down provision with limited integration, narrow but high-value workflow adoption, broad individual use with little aggregate effect, and deliberate non-use where verification or risk costs exceed benefits. Treating all of them as points on one “adoption” scale obscures more than it explains. 4

What Empirical Research Establishes About Sustained Organizational Adoption of Generative AI

Research snapshot: October 3, 2026

Bottom line and evidentiary boundary

The empirical literature is much stronger on short-run use of generative AI by individuals performing bounded tasks than it is on sustained organizational adoption, assimilation, or enterprise value. That distinction is not semantic. It changes what can legitimately be claimed.

There is now credible causal evidence that access to generative AI can raise measured performance in customer support, professional writing, software development, consulting, and product ideation. But the effects vary materially by task and worker, can be zero or negative in some settings, and are usually observed over minutes, days, weeks, or months rather than over the multi-year periods required to establish organizational transformation. 1

By contrast, the evidence base for what causes sustained adoption is surprisingly thin. Representative worker and firm studies identify strong associations with occupational exposure, education, age, gender, organizational access, firm size, digital infrastructure, workforce composition, employer encouragement, and training provision. Yet most of these relationships are observational. One unusually informative randomized information experiment changed workers’ beliefs about ChatGPT substantially but did not produce a detectable increase in actual use two weeks later, demonstrating directly that favorable expectations or intentions need not translate into behavior. 2

The best current evidence on organizational context comes from Denmark. Employer encouragement, training, and deployment of internal AI tools are strongly associated with employee uptake; quasi-experimental analyses using coworkers’ reports of employer encouragement also imply that organizational policy can cause substantially greater worker adoption. Yet the same study finds that widespread chatbot use and average self-reported time savings of about 2.8% of working hours had no detectable effect on workers’ earnings or recorded hours, with confidence intervals ruling out effects larger than about 1%. Organizational facilitation therefore appears capable of increasing use, but increased use cannot automatically be equated with measured economic value. 3

Firm-level evidence is only beginning to become credible. A 2026 Danish study using nationally representative firm surveys and linked administrative records finds that roughly half of surveyed firms reported some generative-AI adoption in 2024, with adoption strongly associated with digitalization and workforce composition; importantly, the predictors of generative AI differed from those of earlier “core AI” technologies. A separate 2026 BIS working paper, covering more than 12,000 firms and using an instrumental-variable design, estimates that AI in general increased short-run labor productivity by about 4%, without reducing employment, but it does not isolate generative AI and therefore cannot establish the firm-productivity effect of GenAI specifically. 4

The most defensible synthesis is therefore:

Question	What the evidence currently establishes
Does GenAI sometimes improve individual task performance?	Yes, causally and repeatedly, but not universally. Effects depend strongly on the worker, task, model, and workflow. 1
Do positive attitudes or intentions establish adoption?	No. Intention is a distinct construct, and experiments show belief/intention changes can fail to alter subsequent use. 5
Do access and organizational encouragement increase employee use?	Yes, with moderate-to-strong evidence, including quasi-causal evidence for employer encouragement. 3
Does training cause sustained workplace adoption?	Not established. Training correlates with use, but clean causal evidence isolating training is scarce; a light information intervention changed beliefs without changing actual use. 2
Does more individual GenAI use cause firm productivity growth?	Not yet established. Task-level causal effects cannot be aggregated mechanically to firms. Generic-AI evidence is more encouraging but is not GenAI-specific. 6
Does GenAI improve organizational innovation?	Promising task-level causal evidence, notably at P&G, but downstream commercialization or firm-level innovation rates have not been established. 7
Is workflow redesign necessary for substantial organizational value?	Plausible but not adequately tested as a necessity claim. Existing evidence is consistent with complementary organizational change mattering, but no strong study cleanly randomizes workflow redesign versus simple provision. 3
Do we have strong evidence on sustained assimilation/infusion across whole organizations?	No. This remains the central empirical gap.

This synthesis deliberately treats behavior, implementation, and value as separate dependent variables. Much apparent disagreement in the literature disappears once that separation is made.

Construct map: what “adoption” actually means

“Adoption” is used so loosely in the GenAI literature that it can denote everything from knowing ChatGPT exists to having transformed an enterprise workflow. These are not equivalent stages.

Construct	What is actually measured	What can legitimately be inferred	State of GenAI evidence
Awareness	Recognition or knowledge that a tool exists	Opportunity to consider adoption, nothing more	High awareness is documented in exposed occupations; in the Danish ChatGPT study, almost all sampled workers knew of ChatGPT. 2
Attitude	Favorability, perceived usefulness, enjoyment, trust, anxiety	Psychological evaluation	TAM/UTAUT-derived studies commonly explain attitudes and intentions, but these should not be treated as use. Actual exposure can itself change attitudes, creating reverse causality in cross-sectional studies. 8
Intention to use	Stated intention to try/use in future	Planned behavior, not behavior	Particularly common in TAM/UTAUT GenAI studies. UTAUT itself explicitly distinguishes behavioral intention from usage. A randomized information experiment changed intention without changing subsequent actual use. 5
Trial / ever use	“Have you ever used ChatGPT/GenAI?”	At least one exposure	Danish and U.S. representative surveys measure this. It says little about persistence or integration. 2
Initial workplace use	Ever used for one’s job / recent use	Behavioral adoption at the individual level	In the U.S. representative survey, 28% of employed respondents reported having used GenAI for work in August 2024. 9
Frequency of use	Days used per week, daily use, monthly use	Repeated behavioral exposure	The same U.S. study found 24.2% used GenAI for work during the preceding week and 10.6% on every workday; this is substantially stronger evidence than “ever used.” 9
Intensity of use	Minutes, prompts, tasks, accepted suggestions, adherence	Depth of usage conditional on adoption	Bick et al. measure time spent; coding and customer-support studies observe richer behavioral telemetry. Intensity is not equivalent to beneficial use. 9
Continued use	Usage observed at repeated later dates	Persistence beyond trial	Rarely measured cleanly. The three-week developer diary experiment and longer customer-service rollout get closer; most acceptance surveys do not. 10
Task-level integration	AI occupies a recurring step inside a real workflow	Operational incorporation into a task	Strongest in customer support, where the assistant generated real-time recommendations inside the live service process, and in coding field experiments. 11
Routinization	AI use becomes stable, normal, repeated practice rather than optional experimentation	Institutionalized individual/team practice	Seldom measured explicitly. Long-running embedded deployments are suggestive, but the GenAI literature lacks mature measures of routinization.
Team diffusion	Use spreads among interdependent coworkers or across a work group	Social diffusion beyond isolated individual use	Direct evidence is limited. The P&G experiment studies AI-assisted teams but not diffusion over time; Danish workplace data capture shared employer context, not a randomized diffusion process. 7
Organizational adoption	Firm reports using/deploying AI; internal model exists; purchase or policy decision	Some firm-level adoption	Census and Danish firm surveys offer this level, but definitions range from “uses AI in producing goods/services” to particular technology classes. 12
Assimilation / infusion	Breadth across units, depth inside processes, interconnected workflow redesign, governance and complementary investments	Technology has penetrated organizational routines	Poorly measured in current GenAI research. Binary firm adoption should not be interpreted as assimilation.
Business value / performance	Productivity, revenue, profitability, innovation, quality, wages, employment, customer outcomes	Consequence of adoption	This is an outcome, not an adoption measure. Short-run task performance is relatively well studied; GenAI-specific firm performance remains largely unresolved. 11

Intention is often being used as “adoption”

This is a significant literature-quality problem. Technology-acceptance studies often use “adoption,” “acceptance,” and “behavioral intention” interchangeably even though the underlying models do not. The original UTAUT separates intention and use behavior and was validated longitudinally across organizations over six months rather than treating the two as identical. 8

This problem persists in recent GenAI research. For example, recent TAM/UTAUT-style studies model behavioral intention, willingness, or readiness as their principal “adoption” outcome, sometimes alongside self-reported use; one 2025 study of 742 Saudi students modeled intention and self-reported usage separately. Such studies can test psychological acceptance mechanisms, but they are not evidence about sustained workplace adoption or organizational assimilation. 13

The behavioral gap is empirically consequential rather than merely theoretical. Humlum and Vestergaard randomized workers to information about expert-estimated task-level ChatGPT time savings. The treatment reduced errors in workers’ beliefs by about 14.9% and moved intentions modestly, yet two weeks later there was no statistically detectable increase in actual ChatGPT use. Favorable expectancy was therefore insufficient to produce behavioral adoption over the observed interval. 2

Firm surveys create a second construct problem. The U.S. Census BTOS measure asked whether firms were using AI in producing goods or services and included, but did not isolate, generative AI. Its reported AI-use rate rose from 3.7% to 5.4% between September 2023 and February 2024. By contrast, the 2024 Danish survey explicitly separated GenAI from machine learning, NLP, machine vision, and related technologies and found roughly 50% of surveyed firms reporting GenAI use. These numbers are not contradictory adoption estimates: they use different populations, definitions, survey instruments, and technology boundaries. 12

Theory map and evidence on antecedents

No major existing theory adequately spans the full sequence individual awareness → trial → continued use → team diffusion → organizational assimilation → business value. Theories that are relatively good at explaining one part are often being stretched into another.

Framework	Natural level of analysis	Main thing it predicts	Fit to modern GenAI	Empirical verdict
Technology Acceptance Model	Individual	Perceived usefulness/ease → attitude/intention/use	Developed for comparatively stable information systems, not general-purpose conversational systems whose outputs and affordances are discovered through use	Still useful for explaining evaluations and intentions, but inadequate by itself for sustained organizational adoption. GenAI evidence is dominated by cross-sectional self-report and intention.
UTAUT	Primarily individual within organizational context	Performance expectancy, effort expectancy, social influence and facilitating conditions → intention/use	Better organizational context than TAM, but originally developed for conventional IT introductions	Several propositions fit GenAI descriptively—especially performance expectancy, facilitating conditions and social context—but current GenAI evidence is much stronger for intention than for assimilation or value. Original UTAUT itself used repeated measures and distinguished intentions from usage. 8
Diffusion of Innovations	Individual, network and organization	Relative advantage, compatibility, complexity, trialability, observability; diffusion over time	Conceptually useful because GenAI adoption visibly involves experimentation, peer diffusion and reinvention	Strong conceptual fit for diffusion/routinization, but few GenAI studies directly test the full diffusion process longitudinally. It should not be cited as direct empirical proof that champions or opinion leaders cause GenAI diffusion.
Technology–Organization–Environment	Organization	How technology, organizational capability and environment condition organizational adoption	Better suited to firm adoption than TAM/UTAUT	Current firm-level results support several TOE-like correlates—size, digitalization, human capital and technology-specific capability—but evidence is mainly predictive/observational rather than causal. 14
Social cognitive approaches	Individual/team	Self-efficacy, observational learning, social modeling	Potentially useful for AI literacy, confidence and peer learning	Plausible, but direct causal GenAI tests of self-efficacy → sustained workplace behavior remain thin.
Sociotechnical systems	Team/organization	Joint optimization of technology, tasks, roles and social system	Arguably better suited than acceptance models to explaining workflow-level value	Strong explanatory relevance, but as of 2026 it functions more as a mechanism lens than as a well-validated predictive GenAI-adoption model. The strongest experiments manipulate AI access, not sociotechnical redesign. 1
Job demands/resources	Worker/job	Well-being, demands, resources, strain, engagement	More relevant to consequences than adoption itself	Useful for hypotheses about workload/stress, but causal GenAI evidence on long-run worker well-being remains limited.
Implementation/change theories	Team/organization	Implementation climate, leadership, routines, complementary change	Highly relevant to assimilation	Direct GenAI tests are scarce. Assertions that sponsorship, champions or change capacity are “essential” currently exceed the causal evidence.
Affordance perspectives	Task/team/organization	Which actions a technology enables in a particular work context	Particularly suitable to a general-purpose technology whose value changes sharply by task	The “jagged frontier,” customer-support, P&G, and coding experiments strongly support the substantive idea that effects depend on task–technology fit, even where studies do not formally test affordance theory. 1

Systematic reviews of organizational AI adoption generally find recurring technological, organizational, and environmental correlates and often combine TOE with diffusion approaches. But those reviews mostly aggregate pre-GenAI or generic-AI studies, many of which use cross-sectional survey designs. They are useful for hypothesis generation; they do not establish that the same antecedents cause ChatGPT-style generative-AI assimilation. A 2026 SME review synthesized 68 peer-reviewed AI studies, illustrating the size of this broader literature but also the transfer problem. 15

Evidence map of antecedents

Performance expectancy / perceived usefulness. This is among the most reliable predictors of attitude and intention. Evidence that it causes actual sustained use is much weaker. The Danish randomized information experiment is particularly diagnostic: changing expected productivity substantially did not change actual use after two weeks. Performance expectancy therefore matters psychologically but is not sufficient behaviorally. 5

Ease of use / effort expectancy. These predict intention in acceptance-model research, but little high-quality workplace GenAI evidence establishes an independent causal effect on continued use. Modern conversational systems also complicate the construct: opening a chatbot is easy, but obtaining, evaluating and integrating a trustworthy output may be difficult. The relevant “effort” may therefore be task-specific verification and integration rather than interface usability. The large variation in coding outcomes despite similarly accessible tools is consistent with that interpretation. 16

Task fit. This has some of the strongest empirical support as a boundary condition on outcomes, although evidence that it predicts adoption itself is more observational. Consultants performed substantially better on tasks within GPT-4’s competence frontier yet became less likely to reach the right answer on a deliberately chosen task outside that frontier. Experienced open-source developers working in mature codebases were 19% slower when early-2025 AI tools were allowed, whereas three corporate coding RCTs found a 26.08% increase in completed tasks on average. The technology is therefore not characterized by a stable “GenAI productivity effect”; fit between model, task, prior knowledge, verification requirements and workflow matters materially. 17

Skills, education and AI literacy. Education and occupational skill composition predict who uses GenAI, but “AI literacy causes value” is not yet established. U.S. representative survey data show much greater workplace use among college-educated workers; European evidence likewise finds skills and nonroutine cognitive job content associated with adoption. In Danish firms, workforce composition predicts technologies differently: STEM-educated workforces are more strongly associated with established/core AI, whereas non-STEM university-educated workforces are associated with GenAI. 9

Crucially, higher prior skill does not imply greater marginal benefit. In customer support, the largest gains accrued to inexperienced and lower-performing workers, with the lowest-skill quintile gaining around 36% in resolutions per hour while the highest-skilled workers saw little productivity gain and small quality declines. Less experienced developers also had higher uptake and larger gains in the three-company Copilot RCT. 11

Trust and perceived risk. These reliably correlate with willingness and attitudes in survey research, but causal effects on sustained workplace use are poorly identified. A three-week randomized workplace study of developers found actual exposure made the tools seem more useful and enjoyable but did not alter perceived trustworthiness of AI-generated code. Thus experience with GenAI can improve attitudes without necessarily resolving trust judgments. 10

Social influence and peers. There is suggestive evidence but less than industry discourse implies. UTAUT predicts social influence, and Danish workplace evidence shows employer-level policies have large effects on adoption. However, evidence that informal peer use itself causes diffusion is scarce; the strongest Danish design instruments employee-reported encouragement using coworkers’ reports of the same workplace policy, which is better interpreted as evidence about shared employer policy than as a clean peer-effect experiment. 18

Facilitating conditions, access and restrictions. Evidence here is comparatively strong. Randomizing access produces actual AI use in field experiments, although compliance is imperfect. Danish workers frequently identified employer restrictions as a barrier, while employers that encouraged chatbot use, deployed internal tools or offered training had much higher uptake. Employer encouragement was associated with adoption rising from roughly 47% to 83% and daily use from about 8% to 21%; because these raw comparisons are observational, the more credible evidence comes from the coworker-policy instrument, which also produces a large estimated effect. 3

Management support. There is therefore credible evidence that organizational encouragement can matter, but the stronger claim that senior-leadership sponsorship is essential is not tested. Workers can and do adopt public GenAI without employer sponsorship, while encouragement raises the probability and intensity of use. “Helpful cause” and “necessary condition” should not be confused. 19

Culture, incentives, governance and change capacity. These are frequent explanatory variables in organizational-AI surveys and reviews, but most direct GenAI evidence remains cross-sectional or qualitative. There is insufficient causal evidence to rank their independent effects once access, occupation, task and worker selection are accounted for. 15

Firm size and digital infrastructure. The evidence that these predict firm adoption is considerably stronger. In the 2026 Danish firm study, digitalization was significantly associated with generative-AI adoption, and firm characteristics were measured, where possible, before observed adoption. The authors emphasize that firm capacity predicts adoption more consistently than prior productivity. U.S. Census evidence for AI broadly also historically finds size gradients, although GenAI may have lower fixed-cost barriers than earlier AI. 20

Data readiness. “Data maturity is required before adopting GenAI” is too strong. Digital infrastructure predicts firm adoption, but public or hosted foundation models allow meaningful GenAI use without firms first training models on proprietary datasets. Pulito and colleagues themselves interpret the approximately 50% GenAI adoption rate as potentially reflecting lower barriers than traditional/core AI. Data infrastructure may matter greatly for proprietary integration and scaling, but existing evidence does not establish it as a prerequisite for initial or useful GenAI adoption. 4

Interventions and causal evidence

The intervention literature is considerably smaller than the correlational antecedent literature.

Intervention	Best evidence	What it establishes	What it does not establish
Tool provision / randomized access	Writing, software-development, consulting and P&G experiments; corporate Copilot RCTs. 21	Access can causally change use and performance on selected tasks.	That employees continue using the system once experimentation ends; that the firm assimilates it; that firm productivity rises.
Performance information	Danish randomized information treatment. 2	Expert information causally improved workers’ beliefs and modestly affected intentions.	It did not measurably increase actual use two weeks later.
Training	Employer training is strongly associated with adoption in Denmark; P&G AI-treated workers all received a one-hour session. 3	Training is compatible with higher use; structured onboarding can accompany successful task experiments.	Clean causal effect of training itself. P&G had no AI-without-training counterfactual. In fact, only 38% used the recommended prompts, and those who did not use them performed similarly to those who did. 7
Managerial encouragement	Danish workplace-policy IV. 3	Relatively strong quasi-causal evidence that encouragement raises employee adoption and reported benefits/new tasks.	That encouragement raises earnings, recorded hours, or firm profits; the same study finds no measurable labor-market effects. 3
Policy / restrictions	Worker reports and employer-policy variation in Denmark. 2	Employer rules meaningfully shape opportunity to use AI.	The optimal governance regime, or whether stricter governance increases trust enough to offset reduced ease of use.
Workflow redesign	No clean factorial workplace experiment located that randomizes AI access independently of process redesign.	Existing studies show embedded tools and complementary organization matter.	The widely stated proposition that redesign is necessary for substantial value.
Coaching	No strong causal GenAI workplace evidence located.	—	Effect on sustained adoption remains unknown.
Champions	No strong GenAI-specific causal diffusion study located.	—	Classic innovation-diffusion logic cannot be treated as direct GenAI evidence.
Peer learning	Social context is associated with adoption; no strong network-randomized GenAI diffusion experiment located.	—	Causal peer multiplier remains uncertain.
Incentives	Organizations are increasingly experimenting with usage mandates and incentives, but credible outcome evaluations remain scarce.	—	Whether incentives produce useful integration rather than superficial log-ins.

The distinction between intervention exposure and intervention mechanism matters. P&G participants received both GPT-4 access and an hour of training, so the study establishes the effect of the treatment package relative to no AI, not the marginal effect of training. Similarly, the customer-support study evaluates an AI system embedded in a specific workflow; it cannot tell us whether the result would have occurred with a generic chatbot and no integration. 7

The Danish information experiment is particularly important because it directly challenges an implicit model behind many “AI literacy” programs: correct beliefs → intention → use. The first arrow moved; the final behavioral outcome did not. Workers who thought ChatGPT could halve task time still often did not intend to use it, with training needs and employer restrictions much more commonly reported as barriers than fear of redundancy. 2

Those reported barriers should themselves not be overinterpreted causally. Saying “I need training” is not equivalent to showing that randomly assigned training will induce sustained behavior. This is precisely the kind of common-method and self-report error that the adoption literature frequently risks.

Adoption, outcomes, mechanisms and heterogeneous effects

Productivity and quality

The proposition that “GenAI increases productivity” is supported only after adding for some tasks, workers and implementations.

Noy and Zhang’s randomized experiment with 453 college-educated professionals found that ChatGPT users completed professional writing tasks about 40% faster and produced outputs rated about 18% higher in quality. This is clean causal evidence for bounded writing work, not evidence about whole jobs or firms. 21

Peng and colleagues’ controlled GitHub Copilot experiment found developers assigned Copilot completed a JavaScript programming task 55.8% faster. Again, the outcome was one relatively self-contained coding task rather than production work in an established codebase. 22

Cui and colleagues substantially improve external validity. Across randomized field experiments at Microsoft, Accenture and a Fortune 100 company involving 4,867 developers, access to a coding assistant increased completed tasks by an estimated 26.08% when the experiments were pooled, with larger adoption and gains among less-experienced developers. Individual company experiments were noisy, so the pooled estimate is more precise than each site estimate. 16

Yet Becker and colleagues’ 2025 randomized trial found the opposite in a very different software context: 16 experienced open-source developers completed 246 real tasks in mature projects with about five years of prior project experience, and allowing frontier early-2025 AI tools increased completion time by an estimated 19%. Developers themselves believed AI was making them faster. This is a small sample, but because tasks were randomized within developers and projects, it is important causal evidence that experienced workers in familiar, high-context codebases can face verification and interaction costs exceeding generation gains. 23

This contradiction became even more informative in 2026: METR reported preliminary later results with point estimates moving toward speedups—about 18% among returning developers and 4% among new recruits—but confidence intervals included zero. Rapid model/tool evolution can therefore change treatment effects over time; “AI productivity” is partly a moving technology-treatment construct. 24

Brynjolfsson, Li and Raymond provide the strongest evidence in a recurring operational workflow. In a staggered deployment to 5,172 customer-support agents, AI access increased resolutions per hour by 15% on average. The lowest-skill workers gained substantially more, while the highest-skilled workers had little productivity improvement and some small declines in quality. The study also found improvements in language fluency, customer behavior, and reduced requests for managerial escalation. Its principal limitation is concentration in one occupation, workflow and family of firms. 11

The BCG experiment further demonstrates the jagged frontier mechanism. Consultants using GPT-4 performed better and faster on tasks deliberately chosen to be inside the model’s competence frontier; on a task outside it, AI users were materially less likely to arrive at the correct answer. The implication is not merely “hallucinations are bad.” It is that AI changes the production function differently across tasks, and workers cannot always identify when the model has crossed from complement into liability. 17

Innovation and team outcomes

The P&G field experiment is unusually strong evidence on organizationally relevant innovation. Between May and July 2024, 791 professionals were randomized among individuals or two-person cross-functional teams, with or without GPT-4, while solving real product-development problems. Individuals with AI improved solution quality by roughly 0.37 standard deviations relative to individuals without AI and reached quality similar to traditional two-person teams. AI also reduced the usual divide between commercial and R&D perspectives. 7

The mechanism appears to have been stronger idea generation and expertise integration, not uniformly better judgment. Teams without AI correctly selected their own best idea about 50% of the time versus roughly 37% in AI-enabled conditions. This is an important example of GenAI improving the distribution of generated alternatives while not improving—and potentially weakening—evaluation. 7

The experiment’s downstream organizational-value inference is nevertheless limited. It lasted one working day, involved “flash-team”-like collaborations, and did not observe eventual product development, commercialization or financial returns. The proposals were embedded in the real P&G innovation pipeline, but the measured outcome was early-stage proposal quality rather than realized innovation revenue. 7

Learning and skill formation

There is credible evidence for one mechanism: GenAI can transmit patterns associated with high-performing workers to less-experienced workers. In customer support, lower-skill and novice agents gained most; language quality improved, and the authors find evidence that some performance gains persisted in ways consistent with worker learning rather than pure contemporaneous dependence on suggestions. 11

But this does not establish long-run human capital accumulation. The literature does not yet robustly answer whether years of AI assistance ultimately make workers more capable unaided, deskill them through reduced practice, shift skill toward verification/orchestration, or produce some combination. The P&G authors explicitly identify whether boundary spanning produces real knowledge growth or only temporary access to expertise as an unresolved research question. 7

A plausible dual mechanism therefore remains open: AI may accelerate novice learning when it exposes them to expert-like patterns, while simultaneously reducing practice or independent generation among experts. The customer-support evidence that top performers sometimes see quality declines while continuing to follow suggestions is consistent with the latter possibility but does not prove long-run deskilling. 11

Satisfaction, stress and workload

Evidence on worker experience is considerably weaker than evidence on task output. In customer support, AI made customers more polite, reduced requests for managers, and was associated with lower attrition, especially among newer workers. These are behavioral workplace outcomes rather than generic attitude surveys, making them relatively valuable evidence. 11

At P&G, AI-treated professionals reported more positive and fewer negative emotions during the experimental task. A three-week developer study similarly found increased perceptions of usefulness and enjoyment, but qualitative responses ranged from enthusiasm to heightened pressure to stay current. These are meaningful signals but do not establish long-run effects on stress, burnout or workload. 7

There is especially little causal evidence that hours “saved” become lower workload. Workers may instead complete more tasks, raise output expectations, review AI output, or have the saved time absorbed by other work. Danish users reported reallocating most saved time to other job tasks, which helps explain why self-reported time savings need not appear as shorter recorded hours. 3

Employment and organizational performance

For generative AI specifically, current labor-market evidence does not establish large displacement effects. Humlum and Vestergaard’s Danish administrative-data study finds no significant effects on earnings or recorded hours across the studied exposed occupations, including among workers whose employers encouraged adoption. The study also found no clear differential effects for workers reporting greater AI benefits or newly created tasks. 3

That result can coexist perfectly well with large task-level productivity effects. A 15% productivity gain on one workflow is not a 15% increase in firm output, because a job contains multiple tasks, AI use is partial, demand may adjust, bottlenecks may move elsewhere, workers reallocate saved time, and organizations may change staffing only gradually. The customer-support authors explicitly caution against translating their task-production result directly into long-run labor demand. 11

The strongest current firm-level causal evidence is for AI broadly, not GenAI. Aldasoro and coauthors’ 2026 BIS working paper links more than 12,000 EU and U.S. firms and instruments European-firm adoption with U.S.-peer adoption. They estimate a 4% short-run labor-productivity increase, no short-run employment reduction, and greater gains among medium and large firms; they also report complementarity with software, data and workforce training. Because the technology measure combines AI types, this evidence should inform priors about organizational complementarity but should not be reported as a 4% GenAI effect. 6

The mechanisms currently supported

Across the stronger studies, five mechanisms have more empirical support than generic “AI makes people productive” stories.

First, knowledge compression and diffusion: AI can expose novices to response patterns or knowledge characteristic of stronger workers, producing disproportionately large novice gains. 11

Second, generation-cost reduction: writing, coding and ideation become faster where generating an initial answer is an important bottleneck. 21

Third, expertise boundary spanning: P&G professionals using AI generated solutions less constrained by their original functional specialization. 7

Fourth, task-contingent complementarity: when AI competence aligns with a task, output can improve dramatically; when verification/context requirements dominate or the model is outside its frontier, performance can deteriorate. 17

Fifth, organizational enabling conditions: policies, access and embedded deployment affect whether the technical capability is actually used. The Denmark studies show large gaps between potential usefulness and realized behavior, and much higher use where employers encourage or support AI. 2

What has not been demonstrated is a general causal chain of the form:

training → sustained use → workflow redesign → organizational assimilation → firm productivity → financial performance.

Individual links have evidence; the complete chain does not.

Workflow redesign, level-of-analysis errors, contradictions and claim audit

Does organizational value require workflow redesign?

The strongest defensible verdict is plausible but not adequately tested.

There are good theoretical reasons to expect complementary process changes, and historical IT research makes that expectation reasonable. Current GenAI evidence is also consistent with it: employer-led implementation raises adoption; self-reported time savings often fail to appear as earnings/hours changes; and generic-AI firm evidence finds larger productivity effects alongside complementary investments. 3

But the empirical case for necessity is much weaker than professional discourse implies.

The customer-support study generated a 15% measured productivity gain because AI was embedded directly in the chat workflow. This could reasonably be called a micro-workflow change, so it cannot identify whether “AI alone” or “workflow redesign” generated value. 11

Conversely, P&G largely inserted GPT-4 into an existing early-stage product-development routine and produced substantial one-day quality gains without redesigning the entire organizational process. That weakens any claim that broad transformation is necessary for task-level value, although it says nothing about long-run enterprise value. 7

The Denmark labor-market evidence is compatible with the complementary-investment hypothesis: adoption and reported savings rose markedly while measured earnings/hours remained unchanged. But the study does not randomize workflow redesign and therefore cannot show that redesign would have converted the savings into business value. 3

A 2026 European preprint using more than 36,600 workers similarly reports no detectable effect of early GenAI adoption on worker-reported technology-related task restructuring in a shift-share design. This argues against assuming rapid transformation automatically follows adoption, but it does not show whether deliberate redesign would increase value. 25

The correct professional claim is therefore narrower:

Existing evidence suggests organizational complements and integration matter, but research has not established workflow redesign as a necessary or sufficient condition for substantial GenAI value.

Level-of-analysis problems

Several common inferences are empirically invalid:

Observed evidence	Invalid leap	Why
Individual writing or coding is faster	“Firm productivity rose”	Jobs contain other tasks; bottlenecks, quality control, demand and coordination may absorb gains. 21
Employee intends to use AI	“AI has been adopted”	Intention and actual behavior diverge experimentally. 2
Employee uses ChatGPT	“The organization adopted GenAI”	Employee use can be unauthorized, personal, isolated or invisible to the employer. 9
Firm reports any AI use	“GenAI is assimilated”	Binary adoption does not measure breadth, routinization or workflow integration. 12
Workers report time savings	“Productivity increased by that amount”	Self-reported counterfactual time estimates are not measured output; Danish administrative outcomes remained unchanged. 3
Executives report company benefits	“Employees became more productive”	Respondent level and outcome level differ, with potential common-method and optimism bias.
AI-treated employees perform better	“Using AI more caused the gain”	Randomized access can identify an intent-to-treat effect, but voluntary adherence within treatment is endogenous. Brynjolfsson et al. explicitly caution that adherence–performance correlations can reflect selection. 11
Teams with AI perform well for one day	“AI has transformed teamwork”	Experimental task composition is not longitudinal diffusion or stable team restructuring. 7
Larger firms adopt more	“Size causes adoption”	Resources, workforce, digital infrastructure and sector are correlated with size. 20

Claim audit

Common organizational claim	Verdict	What the evidence warrants
“Training drives adoption.”	Partially/conditionally supported	Employer training correlates strongly with uptake, and workers frequently cite training needs. But training is highly selected; no major clean RCT establishes its effect on sustained use. A lighter information intervention changed beliefs but not actual behavior. 3
“Leadership sponsorship is essential.”	Plausible but not adequately tested	Employer encouragement has a large quasi-causal effect on employee use. But employees also adopt without sponsorship, and no study establishes senior sponsorship as a necessary condition. 3
“Champions improve diffusion.”	Plausible but not adequately tested	Classic diffusion logic supports the hypothesis, but strong GenAI-specific champion experiments are lacking.
“AI literacy increases value.”	Partially/conditionally supported	Skills/education predict use, but lower-skilled workers sometimes gain more from AI than experts. Training and prompt knowledge are not cleanly isolated from access, task and selection. At P&G, use of the supplied recommended prompts did not predict better performance among AI-treated participants. 26
“Employees resist AI primarily because they fear losing their jobs.”	Contradicted in the strongest directly relevant evidence, with external-validity limits	In Denmark, among workers who saw large potential time savings but did not plan to use ChatGPT, training needs and employer restrictions were much more common reasons; redundancy/dependence fears were cited by fewer than 10%. This cannot automatically be generalized worldwide. 2
“Workflow redesign is necessary for value.”	Plausible but not adequately tested	Complementarity is plausible and indirect evidence is consistent with it. No clean GenAI study demonstrates necessity. 3
“Wider adoption increases productivity.”	Mixed evidence	More access/use raises output in several causal task studies, but some settings show no gain or negative effects, and organizational productivity is not mechanically implied by use. 27
“Younger workers adopt AI more easily.”	Partially/conditionally supported	Younger workers consistently report more workplace use in representative surveys, but “more easily” implies a mechanism not identified by age correlations. Occupation, exposure, education and career stage confound age. 28
“Firms need high data maturity before using GenAI.”	Currently unsupported as a necessity claim	Digitalization predicts organizational adoption, but GenAI diffusion is much broader than traditional AI, and Danish authors explicitly point to lower GenAI barriers. Proprietary/embedded deployment may require data readiness; initial useful use plainly does not always. 4
“Enterprise AI tools reduce shadow AI.”	Currently unsupported	The literature establishes that employer restrictions and provision affect use, but no strong causal study was found showing that enterprise-tool deployment displaces unauthorized external use rather than adding another channel.
“Governance increases trust and adoption.”	Plausible but not adequately tested	Governance could increase perceived safety while simultaneously adding friction or prohibiting use. Current evidence does not identify the net causal effect of governance design on trust, continued use and value.
“More AI use means more business value.”	Contradicted as a general proposition	Use and value are separate constructs. Denmark shows extensive use and reported savings without detectable earnings/hours effects; high-skill customer agents can adhere heavily while receiving little benefit or experiencing quality losses. 3

Why studies appear to contradict one another

Many “contradictions” are differences in construct rather than true effect reversals.

High worker adoption versus low firm adoption. U.S. workers reported substantial personal workplace use while Census firm surveys reported much lower company-level AI use. The former asks workers whether they themselves used GenAI; the latter asks establishments whether AI is used in producing goods or services. Shadow use, different denominators and different definitions naturally produce different estimates. 29

Large experimental productivity effects versus negligible aggregate labor-market effects. Noy and Zhang, Brynjolfsson et al., Cui et al. and P&G measure tasks close to where AI operates. Humlum and Vestergaard measure earnings and hours after all other organizational and market adjustments. Both can be true: local task efficiency can be diluted in whole-job and labor-market aggregates. 21

AI helps developers versus AI slows developers. The evidence genuinely differs, but the populations and tasks differ dramatically. Corporate RCTs include thousands of developers using completion tools across work; METR studies 16 highly experienced maintainers working in familiar, mature repositories with strong quality requirements. The later METR update also suggests the effect may shift as tools improve. This looks more like treatment-effect heterogeneity and technology-time dependence than simple replication failure. 30

Training appears important versus information does not change use. The constructs differ. Firm-provided training is a richer intervention and is selectively allocated; the randomized Danish treatment provided information about productivity, not hands-on capability building. The evidence therefore rejects “information about benefits is enough,” not “all training is useless.” 2

Experts benefit less than novices, yet expertise sometimes complements AI. Customer support and corporate coding show larger gains among novices, while complex tasks may require substantial domain expertise to evaluate AI output. These are not inconsistent: GenAI can compress routine expertise while increasing the value of judgment at the frontier. P&G likewise finds AI improves generation while human teams retain an advantage in selecting their best idea. 1

Study table, evidence gaps and research frontier

The table below emphasizes studies most informative for behavior, intervention, mechanisms or outcomes, rather than the much larger literature of cross-sectional acceptance surveys.

Citation / status	Design, population, N, setting	Technology, exposure, comparator	Outcome and duration	Main finding	Limitations and transferability
Noy & Zhang, 2023, Science; peer-reviewed 21	Randomized experiment; college-educated professionals; N=453; professional writing tasks	ChatGPT access vs no ChatGPT	Completion time and externally rated quality; one experimental session	About 40% faster completion and 18% higher quality	Bounded writing tasks; recruited experiment rather than organizational workflow; cannot infer sustained adoption or firm value
Peng et al., 2023; arXiv/preprint 22	Controlled experiment; software developers; approximately N=95; programming task	GitHub Copilot vs no Copilot	Time to implement HTTP server; one task	AI group finished 55.8% faster	Artificial/self-contained coding task; early Copilot; weak evidence for production-code environments
Dell’Acqua et al., “Jagged Technological Frontier”; working paper, Organization Science forthcoming as of snapshot 17	Preregistered experiment; 758 BCG consultants	GPT-4, GPT-4 plus prompting overview, or no AI	Consultant tasks inside/outside GPT-4 capability frontier; short experimental period	Large speed/quality gains inside frontier; materially worse correctness on an outside-frontier task	High-skill consultants; chosen tasks; short duration; valuable evidence for task-fit mechanism, not sustained adoption
Brynjolfsson, Li & Raymond, 2025, QJE; peer-reviewed 11	Staggered workplace deployment; 5,172 customer-support agents; Fortune 500 software setting	GPT-3-based conversational assistant embedded in support workflow vs not-yet/never treated	Resolutions/hour, quality, language, customer behavior, attrition; repeated agent-month observations	15% average productivity gain; much larger for novice/low-skill workers; little gain and some quality reduction among top workers; evidence of learning	One occupation and workflow; rollout not fully randomized, though event studies/IV/pilot support identification; no long-run firm-profit estimate
Cui et al., 2025; preprint 16	Three company-run RCTs; Microsoft, Accenture, Fortune 100 company; N=4,867 developers	AI coding assistant access vs control	Completed development tasks; real workplace period	Pooled effect +26.08% completed tasks; less-experienced developers had higher adoption and gains	Individual experiments noisy; compliance varied; specific coding assistant and developer population
Becker et al., 2025; arXiv/METR preprint 23	Within-developer randomized trial; 16 experienced developers, 246 tasks; mature open-source projects	Early-2025 Cursor/Claude-family AI allowed vs prohibited	Real task completion time	AI increased completion time by 19%; developers incorrectly believed it saved time	Very small developer sample; unusually experienced workers/projects; nevertheless strong task-level randomization
Dell’Acqua et al., “Cybernetic Teammate,” 2026, Organization Science; peer-reviewed 7	Preregistered field experiment at P&G; 791 randomized professionals; 776 complete post-task surveys	Individual/team × GPT-4/no GPT-4; AI groups received one-hour training	Real product-innovation proposals, expertise integration, emotions; one-day workshop	Individuals+AI about 0.37 SD above individual control in solution quality and similar to unaided teams; AI bridged functional silos; human teams stronger at selecting their best ideas	One-day flash-team setting; no commercialization outcome; cannot isolate training from AI access
Humlum & Vestergaard, 2024/2025 adoption study; working-paper version and subsequent peer-reviewed publication 31	Large survey of workers in 11 exposed occupations; about 18,000 completed responses in key 2023 sample; randomized information experiment and two-week follow-up	ChatGPT information treatment vs control	Awareness, ever use, work use, intention, beliefs and follow-up actual use	Information moved beliefs by about 14.9% toward expert benchmarks but did not significantly increase actual use after two weeks; large gender gap; training needs/restrictions common barriers	Denmark; behavioral use is self-reported; short experimental follow-up
Humlum & Vestergaard, 2025; working paper 3	Two surveys linked to Danish administrative data; roughly 25,000 valid linked workers per main 2024 round, 7,000 workplaces, 11 occupations	Chatbots including employer tools; variation in employer encouragement/training/access	Adoption, reported time savings, new tasks, earnings and hours; 2023–2024 plus administrative outcomes	Employer initiatives substantially increase adoption; average user reports 2.8% of work hours saved; no detectable effect on earnings/hours, with effects >~1% ruled out	Encouragement IV is stronger than raw training associations; Denmark; early diffusion period; labor outcomes not firm profit
Bick, Blandin & Deming, 2024; NBER working paper 9	Nationally representative U.S. Real-Time Population Survey; >5,000 adults	Self-selected use of ChatGPT/Gemini/etc.	Ever use, last-week frequency, intensity and task category; cross-sectional	28% of employed respondents had used GenAI for work; 24.2% in previous week; 10.6% every workday; use higher among younger and more-educated workers	Excellent adoption prevalence; no causal antecedent or productivity inference; worker use ≠ firm adoption
Bonney et al., 2024; Census/NBER working paper — generic AI 12	U.S. Business Trends and Outlook Survey; repeated representative business samples	Broad AI definition including GenAI	Firm-level current and expected use; Sep 2023–Feb 2024	Reported firm AI use rose from 3.7% to 5.4%	Does not isolate GenAI; binary firm use does not measure depth/assimilation
Henseke, 2026; arXiv preprint 25	European Working Conditions Survey; >36,600 workers in 35 countries; observational plus shift-share outcome design	GenAI workplace exposure/use	Adoption and worker-reported task restructuring; 2024 snapshot	Adoption strongly patterned by exposure, skills, job content, worker voice and national context; no detectable early effect on technology-related task restructuring	Preprint; primarily cross-sectional antecedents; self-reported behavior; shift-share assumptions
Pulito, Pytlikova, Schroeder & Lodefalk, 2026; working paper 20	Two nationally representative Danish firm surveys (2023/24) linked to worker/admin records; GenAI predictor models use up to roughly 1,837 firms, broader analyses larger	Core AI vs GenAI vs other technologies	Binary firm adoption and function of use	About 50% weighted GenAI adoption in 2024 survey; digitalization strongly predicts GenAI; workforce predictors differ from core AI; prior firm capacity matters more than performance	Observational; contemporaneous digitalization measures partly create reverse-causality concern; unusually digital Denmark; “adoption” remains binary
Landini et al., 2026; peer-reviewed systematic review — generic AI 15	Systematic review of 68 peer-reviewed SME AI studies	Multiple AI technologies and frameworks	Adoption antecedents and performance associations	TOE/DOI-style technological, organizational and environmental factors recur across studies	Underlying literature heterogeneous and often cross-sectional; limited ability to infer GenAI-specific causality
Aldasoro et al., 2026; BIS working paper — generic AI 6	Matched EIBIS–ORBIS data; >12,000 nonfinancial EU/US firms; IV using U.S.-peer adoption to instrument EU adoption	Broad AI adoption	Labor productivity, wages, innovation and employment; short run	Estimated ~4% labor-productivity increase, no short-run employment loss; gains concentrated in medium/large firms; complementarity with software/data/training	Not GenAI-specific; validity rests on IV assumptions; therefore informative about organizational AI but not a GenAI causal estimate

Causal evidence versus correlational evidence

At present, the strongest causal or quasi-causal claims are fairly narrow:

Randomized access can improve—and occasionally worsen—specific task performance. 21

An embedded generative assistant can improve recurring customer-support productivity and disproportionately assist novices, with several robustness strategies supporting a causal interpretation. 11

GenAI can change the quality and composition of innovation output in a real organization during a randomized one-day product-development exercise. 7

Information about expected productivity can causally change beliefs without changing subsequent actual use. 2

Employer encouragement appears causally capable of increasing worker adoption under the assumptions of the Danish coworker-policy instrument, but this increased use has not translated into detectable earnings or hours effects over the observed horizon. 3

For most other purported antecedents—trust, culture, leadership, training, AI literacy, governance, data readiness, perceived risk, change capacity and champions—the GenAI-specific evidence is predominantly correlational.

What we still do not know

The most important missing evidence is not another cross-sectional “what predicts willingness to use ChatGPT?” survey. The frontier questions require different research designs.

Sustained adoption. We need worker-level telemetry panels following the same employees for at least 12–24 months and distinguishing trial, frequency, intensity, task breadth, discontinuance and routinization. These should be linked to policy changes and job characteristics. Current surveys provide snapshots more often than trajectories.

Organizational assimilation. Research needs organization-level measures of penetration across functions, integration into core systems, process dependence, governance, team interdependence and whether use survives management/model changes. A firm checkbox saying “uses GenAI” is insufficient.

Training. The appropriate design is a cluster or individual factorial RCT separating access × hands-on training × coaching × managerial encouragement, with behavioral telemetry and delayed follow-ups. Current studies usually bundle these components.

Workflow redesign. The crucial experiment would randomize comparable organizational units into at least three arms: no AI, AI access within the old workflow, and AI plus deliberately redesigned workflow. Outcomes should include throughput, quality, error/rework, customer value, labor hours and financial performance over a long enough period for learning and adjustment. Without this design, the “redesign is necessary” proposition remains largely unfalsified rhetoric.

Champions and peer diffusion. Network-randomized rollouts could seed trained users or champions in randomly selected teams and observe spillovers to untreated coworkers. Current observational workplace clustering cannot separate peer influence from common management, tasks or selection.

Governance and shadow AI. Staggered policy experiments could compare unrestricted public-tool access, prohibited access, and approved enterprise alternatives while measuring actual network/tool telemetry, not reported compliance. That is what would be required to determine whether enterprise tools actually displace shadow use.

AI literacy and skill formation. Longitudinal experiments need both AI-assisted and unaided assessments. A worker who performs better with AI has not necessarily become more skilled. Withdrawal tests months later could distinguish genuine learning from temporary technological scaffolding.

Firm value. GenAI rollouts should be linked to accounting, operational and customer data using randomized or plausibly exogenous staggered adoption. Firm-level productivity, unit costs, sales, margins, innovation success and employment need to be measured directly rather than inferred from employee time-savings surveys.

Equilibrium effects. Even good firm experiments cannot by themselves determine sector-wide employment effects, because productivity can alter output prices, demand, entry, outsourcing and occupation composition. Current GenAI labor-market studies are still observing a very early diffusion period. 32

Research frontier versus claims already robust enough to make

Several recent results are promising enough to shape the research frontier but not yet robust enough for universal professional advice.

Pulito and colleagues’ 2026 findings suggest GenAI is organizationally different from earlier AI: its adoption rate is much higher and its human-capital correlates differ, warning against transferring older machine-learning adoption models mechanically. 4

Henseke’s European evidence suggests workplace exposure alone does not determine adoption; worker skills, job content, employee voice and national context appear to shape the exposure-to-use conversion. The absence of detectable early task restructuring is equally important because it shows transformation should not be presumed from access. Publication as a preprint and remaining identification limitations mean this should be treated as promising rather than definitive. 25

The P&G experiment moves the literature beyond isolated individual productivity into team composition, expertise boundaries and innovation, but its one-day duration prevents treating it as evidence of permanent organizational restructuring. 7

The 2025–2026 developer evidence shows just how quickly boundary conditions can change. Large corporate RCTs find substantial gains; experienced open-source developers initially become slower; later METR estimates begin moving toward speedups as tools and behavior evolve. Any timeless claim that “AI raises developer productivity by X%” is therefore scientifically indefensible. 30

The generic-AI BIS firm study is an important frontier result because it begins to connect actual organizational adoption to firm-level productivity using a quasi-experimental design. But until GenAI can be separated from other AI technologies, its estimated 4% effect should remain outside claims specifically about generative AI. 6

What is already robust enough to claim is narrower:

GenAI can materially improve performance on some knowledge-work tasks. The average effect is highly heterogeneous, and negative effects occur. 11

Lower-performing or less-experienced workers often receive larger gains in the settings studied, but this is not universal and should not be generalized to every job. 11

Task fit and the worker’s ability to evaluate AI output are central boundary conditions. 17

Access, employer policy and facilitating conditions materially affect whether workers use GenAI. 3

Positive beliefs and intentions are not sufficient evidence of actual adoption. 2

Firm adoption and individual adoption are different constructs, and both are different from assimilation and business value. 12

Reported time savings should not be equated with measured productivity or financial returns. 3

And, most importantly, the empirical literature as of October 2026 does not yet establish a general model of sustained organizational GenAI adoption or a validated causal chain from common adoption practices—training, champions, sponsorship, governance or workflow redesign—to assimilation and organizational performance. The evidence is already substantial on what happens when particular workers use particular models on particular tasks; it remains immature on how an organization converts that capability into durable, scalable economic value.

From Scattered GenAI Use to Reliable Daily Work: A Practitioner Evidence Dossier

Research snapshot: October 3, 2026

Executive finding and evidence base

The operational evidence does not support a simple linear story in which a company buys an AI tool, trains people, and steadily “adopts” it. The more defensible model is a sequence of three transformations:

private experimentation → shared practice → governed workflow.

Each transformation solves a different problem. Private experimentation discovers useful tasks. Shared practice makes individual discoveries transferable. Governed workflow makes them dependable enough to become part of how work is actually produced, reviewed, recorded, and maintained. The final state is not “high AI usage”; it is a portfolio of work practices in which AI has an explicit role, acceptable failure modes, an owner, observable outcomes, and a way to change when models, data, tools, or policies change. This synthesis is consistent with surveys showing widespread individual use but much less workflow redesign, with field experiments showing strong task-specific effects, and with engineering guidance emphasizing evaluation, permissions, and operational simplicity. 1

That distinction matters because usage has already outrun organizational integration. Microsoft and LinkedIn found in 2024 that 75% of surveyed knowledge workers were using AI at work, 78% of AI users were bringing their own AI tools, and 52% were reluctant to admit using AI for important tasks. 1 Two years later, PwC’s September 2026 global workforce survey found 64% of workers reporting some AI use during the previous year but only 22% using GenAI daily; among PwC’s 56% “engine room” majority, daily GenAI use was only 11%, versus 51% among its 14% “front-runner” cohort. 2 These are different samples and definitions, so the percentages should not be compared as a time series. They do, however, converge on an operational pattern: access and occasional use spread much faster than normalized, repeated use across the workforce. 1

The enterprise-level evidence shows the same gap. McKinsey’s 2025 survey reported that 71% of respondents said their organizations regularly used GenAI in at least one function, while explicitly warning that its broader “AI adoption” measure spans everything from a few employees experimenting to business processes being fully redesigned. Fewer than one in five respondents said their organizations tracked well-defined KPIs for GenAI solutions. 3 A BearingPoint study reported by Reuters on October 1, 2026 similarly found that fewer than one-third of surveyed organizations had progressed beyond pilots despite widespread claims of positive financial results; regulatory constraints and legacy-system integration were among the cited barriers. 4

The evidence also says not to equate adoption with benefit. In a large customer-support field study covering more than 5,000 agents, an embedded AI assistant increased issues resolved per hour by roughly 15% on average, with larger gains among less experienced workers and much smaller gains among top performers. 5 In the BCG/HBS experiment, consultants did substantially better on tasks inside the model’s capability frontier but worse on a business-analysis task outside it; training and warnings did not eliminate that failure. 6 And METR’s 2025 study of experienced open-source developers found that the then-current AI tools increased completion time on its selected real tasks by 19%, even while participants believed they were faster; later METR work found that newer tools, changing user behavior, and selection effects made subsequent estimates much less certain. 7

The practitioner’s unit of adoption should therefore be the workflow, not the user or license. A useful operational representation is:

trigger → input → authorized context → AI operation → validation → human decision → downstream action → system of record → exception/escalation path → measured outcome → owner

That is a synthesis rather than a published standard, but it follows from production-evaluation guidance, workflow/agent engineering practice, permission-aware enterprise AI architecture, and the empirical distinction between mere use and redesigned work. 8

The evidence base here intentionally mixes source classes. Controlled and field studies are strongest for task-level effects; workforce surveys illuminate behavior but are vulnerable to self-report and vendor framing; enterprise telemetry is valuable for actual usage but represents customers of a particular product; practitioner engineering guidance is strong on implementation mechanics but is not causal evidence; community discussions are useful only as hypothesis-generating evidence. No firsthand experience is claimed.

Adoption lifecycle observed in practice

The proposed lifecycle—experimentation → enablement → workflow fit → repeated use → team normalization → operational integration → maintenance/change—is directionally right, but real-world evidence suggests two corrections.

First, enablement is not really one stage. Permission, access, learning, examples, and coaching recur throughout the lifecycle. Second, workflow fit must usually be discovered before extensive enablement, not after it. Training people generically and then asking them to find value reverses the more successful pattern visible in field studies and practitioner guidance: find work with a credible AI advantage, let people practice on it, and then standardize what proves useful. 9

A more operational lifecycle is therefore:

Observed state	What has actually changed	Transition requirement	Typical stall
Exploration	Individuals try chatbots/copilots on convenient tasks, often with personal tools and informal prompts.	Psychological permission, low-friction access, safe experimentation boundaries.	People either hide usage or never start because they do not know what is allowed.
Legitimized experimentation	Approved accounts and basic rules make use discussable and observable.	Enterprise access, identity, data rules, usable policy, examples close to real work.	Secure tool exists but feels inferior or irrelevant; workers continue using personal alternatives.
Task fit	Workers discover tasks where AI predictably helps enough to justify the interaction/review cost.	Task-specific practice, examples, enough domain context, explicit “good output” criteria.	AI is impressive in demos but not reliably better for recurring work.
Repeated personal use	A worker returns to the same AI-assisted task without being reminded.	Low interaction overhead, reusable instructions/context, confidence from prior successful outcomes.	Novelty decays; users forget the tool or find prompting/review slower than doing the work themselves.
Team practice	The method survives transfer between people.	Shared examples, templates/workspaces, role-specific guidance, review norms and peer learning.	Knowledge remains in one enthusiast’s prompt history or personal technique.
Operational workflow	AI is connected to the real trigger, data, systems, review and downstream action.	Integration, permissions, exception paths, observability, owner and acceptance criteria.	Local AI output is faster but creates manual copying, review queues, security friction or downstream rework.
Managed production	Performance and failures are observable enough to adjust review and automation.	Evals, sampling, outcome measures, incidents/feedback, change control.	Management measures activity rather than quality or end-to-end outcomes.
Maintained portfolio	Workflows are changed, migrated, consolidated or retired as technology and work change.	Named ownership, regression checks, knowledge freshness, permission review and pruning.	Prompts, agents and playbooks accumulate without anyone knowing whether they still work.

The first two states are visible particularly clearly in the evidence on shadow AI. Microsoft found 78% BYOAI among surveyed workplace AI users in early 2024; it also found that more than half were reluctant to reveal important-task use. 1 Slack later found 48% of surveyed desk workers uncomfortable telling their manager that they had used AI for common workplace tasks, with concerns that doing so would be interpreted as cheating, laziness, or incompetence. 10 These findings make an important operational point: a restrictive policy does not necessarily produce less AI use. It can instead produce less observable AI use.

This creates a paradox at the experimentation stage. Bottom-up use is valuable precisely because workers know tiny, role-specific pain points that a central AI team will never enumerate. Microsoft’s “power users,” for example, were much more likely than other users to experiment, try different prompts, ask before a task whether AI could help, and redesign workflows. 1 Mitchell Hashimoto’s 2026 account of his own AI adoption—clearly practitioner experience rather than organizational evidence—similarly describes an initial period of inefficiency before stable patterns emerge, followed by progressively more delegated work. 11

But the exact properties that make bottom-up experimentation generative make it fragile: personal accounts, prompt histories, ad hoc source material, tacit verification routines, undocumented workarounds, and results that are not attached to a system of record. The organization may not know which model touched which data, why an output was trusted, or whether the same procedure works for another employee. Microsoft’s permission architecture for enterprise Copilot illustrates the opposite operational model: AI access is tied to the user’s existing role-based permissions rather than creating an entirely separate knowledge-access layer. 12 That still does not solve pre-existing over-permissioning; it makes the quality of the underlying access model more consequential.

The first major stall is therefore “use without legitimacy.” The intervention is not a fifty-page policy. It is a usable answer to: Which tools may I use, with which classes of data, for which kinds of work, and what must a human still own? Google’s UK “AI Works” pilots, reported by Reuters in 2025, found that explicit permission plus a relatively small amount of hands-on training materially increased repeated use months later, particularly among initially hesitant groups. 9 This is unusually useful evidence because it points to permission itself as an adoption mechanism rather than treating training purely as transfer of technical skill.

The second major stall is “use without task fit.” The BCG experiment is the clearest warning against assuming that greater familiarity fixes this. Consultants working with GPT-4 performed markedly better on some creative/analytical tasks but worse on a task outside the model’s effective capability frontier, despite warnings about limitations. 6 This is why a reusable practitioner program should curate tasks, not merely prompts.

The third major stall is “personal success without transmissibility.” A successful personal interaction is not yet a team capability. Transfer requires enough of the hidden method to be externalized: examples, input requirements, context, expected output shape, review checks, and failure cases. Slack’s small internal microlearning experiment with 25 selected participants is weak causal evidence, but it is instructive field evidence: participants said seeing colleagues’ real examples and failures was especially useful, and weekly reported AI use rose from 24% to 72% during the program. 13 Because the cohort was tiny, self-selected, uncontrolled, and measured largely by self-report, it should support experimentation with peer learning—not claims that peer learning is universally effective.

The fourth stall is “team habit without process integration.” McKinsey’s survey is important here because its definition of adoption explicitly spans experimentation through redesigned workflows, making headline adoption numbers intrinsically ambiguous. The survey found that the organizations reporting the strongest financial effects were disproportionately those doing things such as workflow redesign and KPI tracking rather than simply distributing tools. 3 PwC’s 2026 survey reaches the same operational conclusion from a workforce perspective: rising daily usage is not itself enough, and the stronger performers in its associated analysis were more likely to redesign workflows and provide ongoing role-based learning. 2

The final stall is “production without maintenance.” This is increasingly important because the substrate itself changes. OpenAI’s current evaluation documentation, for example, says its existing Evals platform is being transitioned, becoming read-only on October 31, 2026 and scheduled to shut down on November 30, 2026. 14 Anthropic explicitly warns on its still-useful December 2024 agent-engineering article that much of the surrounding tooling landscape described there has since changed. 8 Even good implementation material therefore has a half-life.

The operating system: bottom-up discovery, top-down rails, and workflow integration

The evidence favors a hybrid operating model: centralize the things that must be coherent and decentralize the discovery of useful work.

McKinsey’s survey provides direct evidence that this is already a common organizational pattern. Respondents tended to describe risk/compliance and data governance as centralized, whereas AI adoption and technology talent were more often organized in hybrid or partially centralized models. 3 That arrangement makes operational sense: employees and functional teams own task knowledge; platform/security teams own infrastructure and cross-company risk.

A useful division is:

Bottom-up should discover	Top-down should provide
Where AI removes real friction	Approved provider/tool portfolio
Which role-specific tasks recur	SSO, identity lifecycle and access control
What context actually improves output	Data classification and permitted-use rules
What “good” looks like in the role	Logging, incident and evaluation infrastructure
Which prompts/examples survive repeated use	Procurement, legal/privacy/security review
Where review becomes expensive	Shared integration components and knowledge access
Which practices peers copy voluntarily	Minimum review and escalation requirements
Where local workflow redesign is worth doing	Ownership model and portfolio-level observability

This is not a recommendation for laissez-faire experimentation. BYOAI evidence shows that unconstrained bottom-up adoption fragments tools and hides data flows. 1 Nor is it an argument for one centrally mandated UI. PwC’s 2026 workforce evidence shows very large differences in actual AI access and use between employee cohorts, while empirical task studies show correspondingly large differences in benefit by task and skill level. 2 A universal tool policy can therefore be administratively elegant while operationally wrong.

The strongest operating patterns found in practice are these:

Mechanism	What it operationalizes	Evidence status
Shared worked examples	Turns an abstract capability into a recognizable work task.	Supported by small field evidence and observational surveys; stronger than generic prompt advice but not broadly causal. 13
Reusable instructions / prompts / workspaces	Reduces the cost of reconstructing task context every time.	Strong practitioner convention; enterprise telemetry increasingly shows heavier users employing reusable instructions, apps and shared context, but business impact is not independently established. 15
Role-specific guidance	Maps AI to real outputs, data and acceptance criteria in one function.	Survey correlations favor role/use-case training over generic exposure; causality remains uncertain. 1
Guided hands-on practice	Helps workers cross the initial inefficiency barrier.	Some field evidence: Google’s pilot showed sustained post-training use; Slack’s microlearning experiment showed large self-reported usage changes in a tiny selected sample. 9
Peer channel / community / office hours	Makes failures, examples and workarounds socially visible.	Mostly practitioner convention, with small-field support for peer sharing. 13
Champions	Creates local translators between a central platform and role-specific work.	Widely used practitioner convention; Microsoft’s survey associates power-user environments with stronger leadership support and training, but does not isolate champions causally. 1
Workflow documentation	Captures trigger, data, procedure, checks and exceptions rather than just prompt text.	Strong engineering rationale; direct organization-wide causal evidence is sparse. 8
Evaluation set + acceptance criteria	Makes quality testable across model/prompt changes.	Strong official engineering guidance and standard production practice; broad workplace RCT evidence is limited. 14
Feedback/error channel	Converts production failures into new tests and workflow changes.	Strong production-evaluation convention and part of surveyed scaling practices. 14

The key distinction is between a prompt asset and a workflow asset. A prompt says what to ask. A workflow asset records what causes the work to happen, where the input comes from, what information the model may see, what must be true about the output, who reviews it, what happens next, and what to do when the model cannot complete the task. A company can have hundreds of good prompts and still have almost no operational AI.

Tooling and access should follow the maturity of the work. The practical progression is not necessarily toward maximum technical sophistication:

Form	Best fit	Main advantage	Main failure mode
Ad hoc chat	Discovery, brainstorming, irregular knowledge work	Almost zero setup; maximum exploratory freedom	Non-repeatable context, invisible method, manual transfer
Reusable prompt / project / workspace	Recurring individual or team tasks	Cheap repeatability and shared context	Stale instructions; still dependent on user judgment and manual process
Native copilot	Work already concentrated inside an application	AI appears where the input and action already live	Limited customization; value constrained by host workflow and underlying data hygiene
Deterministic workflow with an AI step	Stable trigger/action with one fuzzy transformation	Keeps deterministic logic deterministic and isolates AI uncertainty	Bad handoff design or inadequate exception handling
Custom AI workflow	Recurrent, valuable work needing proprietary context or system integration	Fine control over data, UI, validation and records	Integration/maintenance cost can exceed local time savings
Agentic workflow	Open-ended multi-step tasks in which the path cannot be specified cheaply in advance	Can select tools and adapt execution dynamically	Higher latency/cost, much larger failure surface, permission risk and harder evaluation

Anthropic’s field-informed engineering guidance makes essentially this complexity argument: workflows use predefined code paths; agents dynamically determine their own process; organizations should start with the simplest solution that works because agentic complexity trades additional latency and cost for flexibility. For many applications, it argues, a single model interaction with good retrieval/context is sufficient. 8 OpenAI’s 2026 enterprise telemetry describes the opposite side of the maturity curve: heavier-use organizations increasingly employ reusable instructions, connected company data/actions, and more agent-like delegation, while explicitly noting that agents require clear rules over data access, permitted actions, and human review. 15 These views are more complementary than contradictory: agency is useful when variability in the work justifies variability in execution.

For recurring business work, the default should therefore usually be deterministic orchestration around a bounded AI operation, not an autonomous agent. This is a recommendation derived from the sources rather than a directly tested universal rule. It preserves predictable triggers, permissions, records, and downstream actions while confining nondeterminism to the part that actually needs language/model capability. 8

Knowledge retrieval requires particular caution. Enterprise systems increasingly ground models in internal sources, but an AI layer does not magically repair access control. Microsoft’s current Copilot architecture states that Copilot accesses information according to the requesting user’s existing permissions. 12 Operationally, that means old oversharing and stale permissions become AI problems too: search and synthesis can make information already technically accessible much easier to discover. Identity lifecycle, knowledge ownership, and permission hygiene therefore belong inside the adoption operating model, not in a separate security workstream.

A sensible multi-provider posture is similarly neither “one model forever” nor “everyone uses anything.” The evidence does not establish an optimal provider count. The practitioner compromise is an approved portfolio with common identity, data rules, procurement, evaluation and possibly a routing/gateway layer, while allowing task owners to choose among approved models when material quality/cost differences justify it. That conclusion is an interpretation of the rapid provider/tool change documented across current platform guidance, not a demonstrated causal best practice. 8

Quality, enablement, and observability

Repeated use requires a very particular kind of trust: not belief that the model is correct, but confidence that the workflow will catch mistakes cheaply enough.

That distinction is visible in experimental evidence. The BCG/HBS work demonstrated that workers can obtain large gains where the task matches model capabilities while becoming worse on superficially similar work beyond that capability frontier. 6 PwC’s 2026 survey likewise found that frequent users were not necessarily naïve about AI: 31% of daily AI users named accuracy or quality concerns as a top barrier, higher than the overall workforce rate. 2 Trust and skepticism can therefore rise together.

A practical review regime should be risk- and task-specific, not “humans review AI” as a blanket rule.

Situation	Sensible review pattern	Why
New workflow; unknown error distribution	Full human review plus error logging	You do not yet know what sampling would miss.
Objective output with deterministic checks	Automated validation first; human review on failures/edge cases	Machines can cheaply test schema, calculations, code, identifiers or business rules.
Stable, lower-risk content with measurable historical quality	Statistical sampling plus targeted review	Full review may destroy economics once error distribution is known.
Factual/decision-support output	Require source evidence and inspect claims with consequential impact	Fluency is not evidence.
High-stakes or irreversible action	Human approval immediately before the consequential action	The review should sit at the decision boundary, not somewhere earlier in the conversation.
Agent with external tools	Permission boundaries, action logs, approval gates for consequential tools, explicit handoff/escalation	Agentic systems introduce nondeterminism in both output and execution path.

These are practitioner recommendations supported by evaluation engineering practice rather than controlled trials comparing every review regime. OpenAI’s current evaluation guidance recommends task-specific tests, representative production data, continuous evaluation, production-log mining, human calibration of automated graders, and evaluation wherever nondeterminism enters a workflow or agent. 14 It also notes the core economic trade-off: human evaluation is high quality but slow and expensive. 14

This is where verification becomes hidden work. METR’s 2025 developer study is especially useful because its measured completion time included interaction with and review of AI-generated work; experienced developers were slower despite believing AI had made them faster. 7 METR’s 2026 follow-up is equally important because it warns against freezing that result in time: newer tools, workers’ increasing unwillingness to accept no-AI experimental conditions, parallel agent use and task-selection changes made clean measurement increasingly difficult. 16 The durable lesson is not “AI makes developers slower.” It is measure total work, including prompting, waiting, checking, fixing, integrating, and downstream review—not merely generation time.

A second version of the same problem occurs downstream. OpenAI’s 2026 enterprise analysis notes that AI can increase identification of security issues faster than organizations can prioritize and remediate them; more generated insight creates a queue somewhere else unless the downstream capacity changes too. 15 The same systems principle applies to emails, reports, analyses, tickets, code and sales research: local throughput is not productivity when another team becomes the filter.

Evaluation sets become necessary when a workflow matters enough to maintain. A useful minimum set should include ordinary examples, hard examples, known previous failures, permission-sensitive cases, edge cases, and examples where escalation is the correct answer. Acceptance criteria should be role-specific: factuality/source support for research, policy compliance for regulated communication, calculation correctness for finance, executable tests for code, resolution/CSAT criteria for service work, and so forth. OpenAI and Google’s production-evaluation guidance both emphasize that organization-specific evals should represent the actual use case rather than generic model benchmarks. 14

Confidence should therefore be calibrated through evidence and workflow behavior, not a model’s rhetorical certainty. In practice this means asking whether required evidence is present, whether deterministic constraints pass, whether the case resembles the evaluated distribution, and whether an explicit uncertainty/escalation rule fires. This is an operational inference from evaluation practice rather than a claim that one confidence-calibration technique has been validated across all workplace settings. 14

Training evidence is much weaker than the volume of training advice suggests. There is reasonable evidence for hands-on, contextual learning:

• Google’s UK workplace pilots found meaningful continued use three months after permission plus a few hours of training, including large increases among initially lower-adoption older women. 9

• Slack’s three-week microlearning experiment showed substantial pre/post increases in comfort and reported daily/weekly use, while participants highlighted seeing peers’ work as especially valuable; the tiny selected cohort and absence of a control group make the causal inference weak. 13

• Microsoft’s large workforce survey found power users more likely to report prompt training, role-specific training, use-case training and leadership encouragement, but those are correlations and could partly reflect more AI-ready organizations or users self-selecting into training. 1

• PwC’s 2026 survey reports that AI-heavy “front-runner” workers have far better access to learning resources than the majority “engine room” cohort, again an association rather than proof that training caused their greater usage. 2

By contrast, this research did not find strong evidence that a one-off generic “how to use GenAI” session reliably transfers into sustained daily workflow change. That is an evidence gap, not proof that such training never works.

The better-supported enablement design is consequently:

real task → worked example → supervised attempt → peer comparison → reusable artifact → independent repeated use → observed outcome.

Awareness training can precede it, but should not be confused with operational adoption. Training material also needs a maintenance owner because both product affordances and model behavior move quickly; current official engineering documentation itself contains deprecations and explicit warnings about older tooling guidance. 8

Observability should form a ladder from activity to outcomes.

Layer	Useful measures	What it actually tells you
Access	approved accounts, entitlement by role, SSO coverage	Whether people can use the approved environment
Activation	first use, first successful target task	Whether access became behavior
Habit	weekly active users, active weeks, repeat rate, actions on the same target workflow, lapse/abandonment	Whether AI became recurring work
Breadth	roles/functions/tasks, feature distribution	Where practice is spreading
Quality	eval pass rate, corrections, rejected outputs, source failures, escalations, incidents	Whether repeated use is dependable
Human effort	generation time + review + correction + coordination + exception handling	Whether the workflow reduces real labor rather than moving it
Process outcome	turnaround time, cases resolved, defects, conversion, response quality, customer outcome, queue size	Whether work improved
Economics	licenses/API/infrastructure + training + integration + review + maintenance + incident costs	Whether improvement is worth sustaining

Microsoft’s current Copilot analytics illustrate both the possibilities and traps. Its adoption report can distinguish “novice,” “habitual” and “power” usage based on frequency and consistency over weeks rather than simply counting assigned licenses. 17 Its impact reporting can also link usage with collaboration patterns. 18 But an “active user” can be defined as someone taking at least one action during a chosen period, which is far too weak to represent workflow adoption by itself. 18

Likewise, Microsoft’s “Copilot assisted hours” is explicitly an estimate derived from product actions and research-based multipliers, not direct observation of net hours saved; Microsoft’s documentation cautions that the methodology may evolve. 18 OpenAI’s 2026 enterprise telemetry makes a similar caveat about its own usage intensity measures: tokens are useful behavioral data but are an imperfect proxy for business value. 15

That yields a clean distinction:

Useful leading indicators: repeated task-specific use, consistency over time, penetration across the target population, corrections, error patterns, abandonment, incidents, escalation and human review effort.

Useful outcome indicators: cycle time, throughput adjusted for quality, rework, downstream queue burden, customer/employee outcome and total cost per completed unit of work.

Vanity or highly ambiguous indicators when used alone: licenses purchased, accounts provisioned, training completions, prompts/messages sent, token volume, number of agents built, number of use cases submitted, aggregate “hours saved” based only on employee estimates, and raw weekly-active-user counts.

McKinsey’s finding that KPI tracking had the strongest positive association with reported GenAI EBIT impact among the scaling practices it studied is notable here, but because the evidence is observational it does not prove KPI dashboards cause economic returns. 3

Maintenance model and failure-mode catalogue

GenAI adoption should be operated more like a product and workflow portfolio than a transformation campaign with a launch date. Models change, tools are deprecated, organizational data changes, permissions change, regulations change, people leave, and tasks themselves are redesigned. Current official documentation from multiple providers already demonstrates this pace of change. 8 In the EU, AI governance is also no longer a static future issue: the AI Act entered into force in 2024, AI-literacy provisions began applying in 2025, and most of the Act became applicable in August 2026, making stale policy and training material an immediate operational concern for affected organizations. 19

The minimum ownership structure is not necessarily a new bureaucracy. It is clarity about who owns which failure surface:

Asset	Named owner must answer
AI provider/platform portfolio	Which tools/models are approved, replaced or retired?
Identity/data access	Who can expose or act on which information?
Business workflow	Is this still worth running? What outcome is it responsible for?
Prompt/instructions/configuration	Does it still produce acceptable results after changes?
Evaluation set	Does it represent current work and recently observed failures?
Knowledge source	Is it current, authoritative and correctly permissioned?
Review/escalation process	Who catches failures, and where do ambiguous cases go?
Enablement material	Is guidance still accurate for today’s tools and policy?
Incident/error log	Are recurring failures becoming tests and workflow changes?
Community/champions	Is knowledge distributed, or does the system depend on one exhausted expert?

A sensible review rhythm is event-driven plus periodic. Model/provider/configuration changes should trigger regression evaluation; production incidents should immediately create candidate eval cases; permission and knowledge changes should trigger access/freshness checks. Monthly or quarterly portfolio reviews should examine sustained usage, quality, outcomes, downstream burden and duplication. The exact cadence has not been established by comparative evidence and should scale with workflow risk and change rate. Continuous evaluation itself is directly recommended in production evaluation guidance. 14

Not everything deserves maintenance. A workflow should be allowed to die when nobody repeatedly uses it after reasonable support, when its target work disappears, when another workflow supersedes it, when verification cost consumes the advantage, when quality remains below the acceptance threshold, or when nobody is prepared to own it. This is a practitioner recommendation, not an empirical threshold.

The following failure catalogue prioritizes mechanisms that recur across empirical work, company evidence and practitioner reports. Community items are deliberately treated only as anecdotal.

Failure mode	Symptom	Likely mechanism	Evidence	Detection	Possible mitigation	Uncertainty
Shadow AI / hidden usage	Official dashboard looks quiet while employees clearly use GenAI; personal accounts appear in browser/tooling data; employees avoid discussing use.	Approved options arrive late, policy is ambiguous, or employees fear judgment/replacement.	Microsoft: 78% BYOAI and 52% reluctant to disclose important-task use in its 2024 survey; Slack found 48% uncomfortable admitting use to managers. 1	Anonymous employee survey; SaaS/security telemetry where lawful; compare approved usage with self-reported task use; ask explicitly what people cannot do in the approved tool.	Give a competitive approved option; make permitted/prohibited uses concrete; preserve low-risk experimentation space; remove career stigma around disclosure.	Medium. Strong self-report evidence; exact prevalence changes rapidly and varies by firm.
License rollout mistaken for adoption	High seat allocation, low repeat use; leadership reports “X% deployed” with little evidence of recurring work.	Availability is treated as behavior.	Microsoft distinguishes licensed users from actual frequency/consistency and defines habitual/power usage over multiple weeks. 17	Cohort retention; active weeks; repeat use on target tasks; abandonment after first month.	Measure activation → repeated task use → outcome separately.	High confidence in mechanism; no universal threshold for “good” repeat usage.
Enthusiast plateau	A small group races ahead; the median worker barely changes behavior.	Self-selected early adopters have more curiosity, autonomy, time, skill or better role fit.	PwC 2026 finds 51% daily GenAI use among front-runners versus 11% among the core workforce; Microsoft finds distinct skeptic-to-power-user segments. 2	Usage distribution by role/team/seniority rather than company average; Gini-like concentration; share of target workers with sustained use.	Study power-user workflows, then test whether they transfer to non-enthusiasts; remove access/permission gaps; protect practice time.	High that inequality exists; lower on which intervention closes it.
Training without transfer	Good course ratings; prompt demos work in class; little sustained task-specific use afterward.	Generic learning lacks work context, practice, permission or manager reinforcement.	Role/use-case training correlates with power use in Microsoft data; Google field pilots showed sustained use after permission plus hands-on training; Slack’s tiny microlearning experiment suggests peer practice helps. 1	Measure 30/60/90-day target-task use and workflow outcomes, not course completion.	Replace generic-only training with role tasks, repeated practice, examples and coached transfer.	Medium-low. There is little strong comparative evidence directly pitting training designs against one another.
Prompt library graveyard	Hundreds of saved prompts, few known users, unclear authorship, contradictory instructions.	Prompt is stored without task ownership, context assumptions, versioning or outcome evidence.	Reusable prompts/playbooks are common practitioner practice; Microsoft itself recommends them for spreading power-user workflows, but this is guidance rather than causal evidence. 17	Prompt reuse, last-use date, owner, workflow association, evaluation status.	Promote only prompts attached to recurring workflows; version/evaluate high-value ones; archive unused artifacts aggressively.	Medium-low. Failure pattern is plausible and widely reported, but rigorous organizational prevalence data are scarce.
Jagged-frontier failure / overtrust	Strong performance on familiar tasks causes employees to trust AI on subtly different tasks where it performs badly.	Capability boundaries are hard to infer from fluency; experience on one task generalizes psychologically faster than model competence does.	BCG/HBS experiment found gains on in-frontier tasks but materially worse performance on an outside-frontier task. 6	Task-level quality, not aggregate satisfaction; adversarial/edge cases; compare AI/no-AI baseline by task family.	Define approved task envelopes; preserve escalation; include known failure cases in training/evals.	High for existence; boundary locations change by model.
Verification tax erases the gain	Creation looks faster but total cycle time stays flat or rises; reviewers complain about subtle errors.	AI reduces generation cost while increasing checking, debugging, reconciliation or integration effort.	METR measured a 19% slowdown for experienced OSS developers using 2025-era tools despite perceived speedups; later work warns the exact effect has changed and is difficult to estimate. 7	Time from trigger to accepted outcome; reviewer minutes; correction count; rejected output; downstream rework.	Constrain AI task, automate deterministic checks, use sampling only after error distribution is understood, improve context/evals—or stop using AI there.	High that review cost can dominate; low that any historical percentage generalizes to current models/tasks.
Local optimization creates downstream overload	One team produces more reports/code/leads/tickets than another team can review, prioritize or execute.	Generation throughput rises without downstream capacity or quality gates.	OpenAI’s enterprise analysis explicitly notes security-finding volume can create a remediation/prioritization bottleneck. 15	Queue age/size, reviewer load, reject rate, unresolved findings, end-to-end throughput rather than creator output.	Optimize the full value stream; rate-limit low-value generation; automate triage; change upstream acceptance criteria.	Medium-high as systems mechanism; evidence outside technical workflows remains less quantified.
Workflow embedding adds friction instead of removing it	Employees bypass the AI step, duplicate data entry, copy between systems, or create manual exceptions.	AI was inserted because it was possible, not because it removed a constrained step; trigger/input/action are poorly integrated.	Anthropic reports simpler composable patterns as more successful among teams it has worked with and recommends adding agentic complexity only when necessary. 8	Bypass rate; manual handoffs; exceptions; clicks/time added; user workarounds.	Remove the AI step unless it improves total workflow; integrate at a natural decision/transformation boundary.	Medium. Strong practitioner engineering evidence, sparse cross-company causal research.
Permissions and retrieval amplify old data hygiene problems	AI exposes unexpected internal material or produces answers from stale/wrong sources.	Retrieval respects existing access, but existing access or knowledge ownership is poor.	Microsoft states Copilot uses the requesting user’s existing authorized access. 12	Permission audits; source attribution; stale-content sampling; access-path incident analysis.	Clean permissions before broad retrieval; name knowledge owners; index authoritative sources; provide source links/citations.	High for architectural mechanism; actual incident rate is organization-specific.
Management measures activity instead of work	Prompts, tokens, WAU and “hours saved” rise while business outcomes are unknown.	Product telemetry is available earlier and more cheaply than causal outcome measurement.	OpenAI cautions that tokens are an imperfect value measure; Microsoft’s “assisted hours” is an estimate based on activity multipliers; McKinsey finds KPI tracking uncommon despite its association with higher reported value. 15	Put an outcome metric beside every adoption metric; audit whether estimated savings appear in capacity/cycle-time/quality.	Treat usage as diagnostic leading data only; instrument end-to-end workflow outcomes.	High. The difficulty is attribution, not the inadequacy of activity alone.
No maintenance owner / workflow sprawl	Duplicate bots/agents appear; nobody knows which prompt is canonical; behavior changes after model updates; documentation is stale.	Experiment artifacts graduate into production without product ownership or retirement rules.	Current provider documentation changes rapidly; Anthropic flags its own older tooling material as outdated and OpenAI currently has scheduled evaluation-platform deprecations. 8	Owner field; last eval date; last use; duplicates; model/version dependency; failed regression tests.	Workflow registry for production-grade assets; required owner; regression tests; archive/merge policies; event-driven reviews on provider changes.	High that change creates maintenance demand; low on optimal governance structure.
Review norms fail socially	Some workers paste outputs directly; others re-check every word; quality and economics vary wildly between people.	“Use judgment” provides no shared acceptance threshold.	McKinsey found wide variation in review behavior, including organizations where all generated content was reviewed and others where only a small share was reviewed. 20 Community reports from experienced developers likewise describe frustration reviewing AI-generated code that the submitter cannot explain; this is anecdotal, not prevalence evidence. 21	Reviewer disagreement; correction variance; incidents tied to unreviewed output; review time by person.	Define role/task-specific acceptance criteria and evidence requirements; train using pass/fail examples.	Medium-high for variation; exact optimal review depth depends on risk and error observability.
Champion dependence and burnout	Same few people answer every question, maintain examples and rescue bad workflows; program weakens when they leave.	“Community” becomes unpaid support infrastructure without formal capacity or asset ownership.	Champion programs are widespread practitioner convention, but rigorous evidence on burnout or optimal champion ratios is sparse; Microsoft’s evidence supports champion-style diffusion only indirectly. 1	Support concentration, unanswered questions, bus factor, champion time spent, asset ownership gaps.	Fund the function, distribute expertise, turn recurring answers into maintained artifacts, rotate champions and retire low-value requests.	Low-medium. Plausible and frequently reported operationally, weakly quantified.

The catalogue exposes a recurring theme: many GenAI failures are not model failures. They are failures of permissions, task selection, workflow design, review economics, social permission, measurement, ownership, or downstream capacity. Better models may reduce some of them while making others worse by increasing output volume and encouraging more automation.

Practitioner disagreements and claim-to-source map

Credible practitioners do not agree on several major operating choices. The disagreements are useful because they reveal where context, rather than universal doctrine, determines the answer.

Central control versus open experimentation. Security and governance logic pushes toward a small approved portfolio, enterprise identity and centrally governed data access. Shadow-AI evidence argues that if these controls make legitimate experimentation too cumbersome, employees simply migrate outside observable systems. 1 The practical disagreement is therefore over the width of the sandbox: some organizations prioritize containment; others prioritize discovery. The evidence supports neither unrestricted BYOAI nor a single rigid tool as universally superior.

Training first versus workflow redesign first. Microsoft and many adoption practitioners emphasize role training, champions and encouragement; their survey data show strong associations with power-user behavior. 1 McKinsey and PwC put considerably more weight on workflow redesign and business-process integration as the bridge to economic value. 3 These positions need not conflict. Training helps people discover and use capabilities; redesign is what captures the resulting capability in a repeatable operating process. The mistake is treating either as a substitute for the other.

Prompt craft versus context/integration. Early practitioner programs often focus on prompting. By 2026, heavier enterprise usage increasingly emphasizes reusable context, connections to company data/actions and delegated workflows. 15 The more operational interpretation is that prompt skill remains useful for exploration, while production value increasingly depends on giving the model the right context and tools and constraining the resulting action.

Copilot everywhere versus purpose-built workflows. Native copilots minimize adoption friction because employees stay inside familiar applications. Purpose-built workflows create better control over inputs, outputs, tests and actions but cost more to build and maintain. No broad evidence establishes a universal winner. Customer-support field evidence demonstrates that a tightly embedded assistant can yield measurable gains, while engineering guidance emphasizes that bespoke complexity should earn its keep. 22

Deterministic workflows versus agents. Anthropic argues explicitly for the simplest architecture possible and sees predefined workflows as preferable when tasks are well defined. 8 OpenAI’s 2026 enterprise evidence emphasizes a movement among heavier-use organizations toward greater delegation, tools, reusable context and agents. 15 Both can be true because they are answering different questions: “What architecture minimizes unnecessary complexity?” versus “What does advanced use increasingly look like?” The operational decision turns on whether valuable variability in task execution outweighs the extra evaluation, permission, latency and failure complexity.

Full human review versus sampling. Full review is attractive in regulated, novel or high-stakes workflows, and McKinsey reports that some organizations use it extensively. 20 But human evaluation is costly, and production-evaluation practice explicitly supports automated tests and scalable graders calibrated against human judgment. 14 Sampling becomes defensible only once the failure distribution is understood and the cost of missed failures is tolerable; there is no evidence-based universal sampling percentage.

Who adopts fastest. Some self-report surveys show executives and managers using GenAI more heavily than individual contributors; McKinsey, for example, reported 53% regular use among surveyed C-level executives versus 44% among midlevel managers. 3 Slack’s 2025 workforce survey likewise reported very large seniority differences in daily use. 23 But OpenAI’s 2026 administrative telemetry found early-career users sending more messages per week than executives six months after adoption in its enterprise data. 24 This should not be “resolved” by picking a favorite statistic. The sources measure different products, populations, periods and definitions; the disagreement is itself a warning that role adoption should be measured locally rather than inferred from generic benchmarks.

Activity as a useful signal versus a dangerous target. Product providers legitimately expose frequency, active weeks, features and token/action volumes because these measures reveal behavior. Microsoft has become increasingly sophisticated in distinguishing novice, habitual and power users. 17 Yet OpenAI explicitly cautions that usage volume is not business value, and Microsoft’s own time-savings measures are estimates. 15 The resolution is to use activity diagnostically and outcomes evaluatively.

The following claim-to-source map separates evidence type from interpretation.

Load-bearing claim	Principal source(s)	Source classification	Evidential weight / limitation
Individual workplace AI use can significantly precede official organizational enablement.	Microsoft/LinkedIn 2024 Work Trend Index: 78% BYOAI among surveyed AI users. 1	Company report	Large multinational survey; self-report and vendor sponsorship.
Employees may conceal AI use because of social/career concerns.	Microsoft/LinkedIn; Slack Workforce Lab. 1	Company report	Convergent surveys; cannot quantify any individual firm’s hidden use.
Adoption remains highly uneven even when aggregate use is high.	PwC Global Workforce Hopes & Fears 2026. 2	Company report	Nearly 50,000-worker global sample; cohort construction and outcomes are survey-based.
“Adoption” numbers often mix experimentation and embedded process use.	McKinsey State of AI 2025 explicitly notes this span. 3	Company report	Important measurement caveat; respondent-defined organizational state.
Workflow redesign and KPI tracking are associated with greater reported business value.	McKinsey 2025; PwC 2026. 3	Company report	Consistent observational association; cannot establish causality.
Embedded AI can materially improve real task performance, with heterogeneous gains by worker skill.	Customer-support field study, >5,000 agents. 5	Empirical study / field evidence	Strong real-work evidence; one occupation and implementation.
AI can make workers worse when they misjudge task/model fit.	BCG/HBS “jagged frontier” experiment. 6	Empirical study	Strong task-level experimental evidence; models/tasks have since changed.
Review and coordination cost can reverse apparent productivity gains.	METR 2025 developer field experiment and 2026 follow-up. 7	Empirical study / field evidence	Valuable direct timing; narrow expert-developer setting, rapidly changing tool generation.
Hands-on, permission-backed training can transfer into continued use.	Google AI Works pilots reported by Reuters. 9	Field evidence	Continued usage measured months later; vendor-connected intervention and limited contexts.
Peer microlearning is promising but not strongly proven.	Slack internal 25-person experiment. 13	Field evidence	Direct workplace experiment; tiny self-selected cohort, no control, self-report.
Enterprise knowledge access must be coupled to identity/permissions.	Microsoft Copilot architecture/security documentation. 12	Official documentation	Strong statement of product behavior; does not prove underlying customer permissions are sound.
Production AI needs task-specific, continuous evaluation rather than generic benchmarks alone.	OpenAI evaluation guidance; Google evaluation guidance. 14	Official documentation	Strong engineering convention; provider-authored and not organizational outcome research.
Agentic complexity should be justified rather than assumed.	Anthropic engineering experience across customer teams. 8	Practitioner experience	Explicit field-informed advice; vendor perspective, not comparative trial.
Advanced enterprise use increasingly combines reusable instructions, context, tools and delegation.	OpenAI Enterprise Signals 2026. 15	Company report / field evidence	Actual aggregated product telemetry; selected OpenAI enterprise customer population.
Active-user/activity metrics are insufficient proxies for value.	Microsoft analytics definitions; OpenAI Enterprise Signals. 18	Official documentation / company report	Directly establishes metric construction and provider caveats; outcome link remains local.
AI-generated work can create review problems when authors cannot defend it.	ExperiencedDevs discussion. 21	Community discussion	Anecdotal; useful for failure hypothesis, not prevalence or causal estimation.
Successful adoption is better modeled as a maintained workflow portfolio than a one-time rollout.	Synthesis of workflow-value evidence, changing tooling, continuous-eval guidance and usage/adoption gaps. 3	Interpretation	Strongly consistent with evidence; not itself experimentally tested as an organizational model.

The source classes matter. In particular, champions, office hours, playbooks, prompt libraries and communities of practice are much better supported as practitioner conventions than as independently validated causal interventions. They are reasonable things to test, but the dossier does not support claiming that they are universally necessary or sufficient. By contrast, task heterogeneity, shadow use, review cost, unequal adoption, permission dependencies and the distinction between activity and outcomes have substantially stronger empirical or direct operational evidence. 1

What cannot be established without running a real implementation

A research dossier can identify mechanisms and design a better experiment. It cannot determine the following facts for a particular company without observing that company’s work.

Which workflows actually have positive net value. Benchmark scores, vendor demos and generic time-saving estimates cannot tell you whether AI reduces end-to-end effort for your contract review, customer escalation, analyst memo, proposal generation, data investigation or software workflow. Task effects vary sharply even inside one occupation. 25 The necessary evidence is a baseline and a comparable AI-assisted cohort or time period measuring accepted output, quality, total human effort and downstream effects.

Where the company’s capability frontier is. A model being good at adjacent tasks does not establish performance on the actual local distribution of inputs, terminology, policies and exceptions. 26 This requires a representative local evaluation set.

The true verification tax. People are unreliable judges of their own time savings, and review effort can be invisible or shifted to another role. 27 Establishing it requires measuring total elapsed and active labor from trigger through accepted downstream outcome, including reviewer and exception-handler time.

Whether training transfers. Course completion, satisfaction, confidence and even immediate post-training usage do not establish durable workflow change. The stronger available field evidence still covers limited contexts. 9 A real implementation needs 30/60/90-day task-specific behavior and quality/outcome measures, ideally with comparison groups when practical.

How much shadow AI exists locally. Global surveys demonstrate the phenomenon, not the rate inside any given organization. 1 Anonymous surveys, safe disclosure mechanisms and lawful security/SaaS telemetry are needed to estimate it without incentivizing further concealment.

Whether the approved tool is competitive enough to displace personal tools. Security approval says nothing about workflow fit, latency, model quality, context window, integrations or user experience. This requires side-by-side work on representative tasks.

Which workers benefit and which lose. Existing studies consistently show heterogeneity, but disagree on some role/seniority adoption patterns and cannot substitute for a company’s own job architecture. 28 Analyze adoption and outcomes by task and role, not only company averages.

Whether embedding AI simplifies or complicates the process. Architecture diagrams cannot reveal whether employees will bypass a step, copy data manually, create exceptions, wait for approvals or invent parallel systems. Anthropic’s simplicity guidance is useful precisely because complexity costs are implementation-specific. 8 Instrument handoffs and exceptions in the live workflow.

The acceptable review rate. Neither “review everything” nor a generic 10% sample can be justified without knowing local error frequency, error detectability, consequence, reversibility and the cost of review. Production evaluation guidance can tell you how to build evidence, not the organization’s risk tolerance. 14

The real data-permission exposure. Product architecture can establish that an AI honors existing permissions; it cannot establish that those permissions are correct. 12 That requires local access analysis and adversarial retrieval testing.

The downstream externalities. Faster document generation may burden approvers; faster coding may burden reviewers; more security findings may burden remediation teams; more sales research may generate low-quality outreach. OpenAI’s security example demonstrates the mechanism, but only process-level measurement reveals where it occurs locally. 15

The correct provider strategy. There is no evidence-based universal number of AI providers. Model quality, integration depth, privacy terms, latency, cost, user preference, lock-in and switching cost matter differently by workload. The answer has to be established through an approved, evaluated portfolio rather than market benchmarks alone.

The maintenance burden. The pace of model and tooling change is observable, but the number of hours required to keep a company’s prompts, evals, retrieval, permissions, training and integrations healthy is a local operational quantity. 8 Track maintenance labor from the outset rather than treating it as free.

The causal business return. Activity telemetry is much easier to collect than incremental profit, capacity or service improvement. Even large surveys showing relationships between workflow redesign, KPI discipline and financial results remain observational. 3 Real attribution requires a credible counterfactual: randomized rollout where feasible, matched cohorts, stepped-wedge deployment, interrupted time series, or at minimum a pre-registered baseline with explicit confounders.

The most useful implementation test is therefore not “Did people adopt AI?” It is:

For this recurring unit of work, did the AI-enabled process produce a better accepted outcome at lower total cost or greater strategic value, without creating unacceptable risk or transferring hidden work elsewhere—and does that remain true after the novelty, training intervention, model version and original champion have changed?

Everything before that question is enablement. Everything after it is operations. The companies that turn scattered GenAI use into sustainable daily work are not merely the ones that get more employees prompting. The evidence points instead to organizations that convert discovered tasks into owned, permissioned, testable, observable and disposable workflows—disposable because the ability to retire an AI workflow that no longer earns its complexity is as much a sign of operational maturity as the ability to launch one. 3

Decision Evidence Dossier: From Informal Generative AI Use to Deliberate Organizational Capability

Research snapshot: October 3, 2026

The central finding is not that organizations should “adopt AI.” It is that leadership should convert an uncontrolled collection of AI experiments into an explicit portfolio of work-system decisions.

The best available causal evidence shows that generative AI can produce large improvements on some tasks, little benefit on others, and measurable harm on still others. In controlled and field studies, effects range from roughly 40% less time on professional writing tasks, to 26% more completed software-development tasks, to 15% higher customer-support productivity, to 19% longer completion times for experienced developers working in familiar codebases. AI can simultaneously increase productivity on tasks inside its capability frontier and reduce performance when people over-rely on it outside that frontier. 1

That dispersion matters more for management than the average effect.

The unit of decision should therefore usually be the task, workflow, or decision process—not “AI adoption” at company level. Broad access, targeted copilots, workflow redesign, automation, decentralized experimentation and deliberate non-adoption are all defensible choices under different conditions. The burden of proof should rise with irreversibility, risk, cost and organizational change.

A useful default stance is:

Make experimentation easy where mistakes are cheap and reversible; demand increasingly strong evidence as AI gains access to sensitive data, influences consequential decisions, changes workflows, acts autonomously, creates recurring cost, or displaces existing capacity.

This is a decision dossier, not an argument for acceleration.

Leadership decision map

Leadership is facing several different decisions that are often incorrectly collapsed into “What is our AI strategy?”

Decision leadership actually faces	Evidence needed before deciding	Reasonable outcomes
Is this activity worth supporting at all?	Baseline pain, frequency, economic relevance, current quality/cycle time	Support; tolerate experimentation; postpone; explicitly stop
What problem are we solving?	Outcome metric independent of AI usage	Faster completion, better quality, capacity, customer outcome, learning, or no compelling problem
Is GenAI the right intervention?	Comparison with no change, process simplification and deterministic automation	GenAI; conventional software/automation; workflow change; nothing
At what level should intervention occur?	Where the bottleneck actually sits	Individual assistant; team tool; workflow redesign; process automation
Which uses can remain decentralized?	Reversibility, information sensitivity, consequences of error	Unmanaged/ lightly governed experimentation; approved sandbox; central control
Should the company provide an approved general tool?	Breadth of genuinely useful low-risk tasks, security requirements, utilization economics	Enterprise tool; limited seats; reimbursement; no general license
Does a role-specific copilot make sense?	Repeatable role-specific work and stable context requirements	Deploy; prototype; retain general assistant
Should AI act or merely advise?	Error recoverability, verification ability, authorization risk	Draft only; recommend; execute with approval; bounded autonomy; no automation
Does the workflow itself need redesign?	Evidence that local time savings fail to change end-to-end performance	Keep assistant; redistribute work; remove steps; redesign process
Is a pilot enough?	Objective of pilot	Continue as local experiment; gather more evidence; productionize; terminate
Is evidence strong enough to scale?	Representative users/tasks, objective outcomes, guardrails, economics	Expand; restrict to subgroup; redesign; stop
What governance is proportionate?	Data class, decision consequence, regulation, external exposure, autonomy	Minimal rules through formal high-risk controls
Who owns the use case?	Location of business outcome and technical/control dependencies	Business owner, product/process owner, centralized platform, shared accountability
What capability must users acquire?	Observed failure modes rather than generic “AI readiness”	No training; contextual guidance; role-specific training; formal competence requirement
How will saved capacity be converted into value?	Downstream bottlenecks and management response	More throughput; shorter queues; quality improvement; capacity reduction; no realized value
When do we stop?	Pre-agreed failure conditions	Abandon; return to human process; substitute deterministic automation; re-scope

This separation is important because the empirical literature largely measures specific tasks under specific human–AI configurations, rather than an abstract organizational state called “AI adoption.” Noy and Zhang tested occupational writing; Brynjolfsson, Li and Raymond studied customer-support agents; Cui and colleagues studied professional developers; Dell’Acqua and colleagues tested consultants and later P&G product-development professionals. Their different results are evidence against treating GenAI as a homogeneous intervention. 1

Decision implication: an “AI transformation program” may be useful as an umbrella for common infrastructure and controls, but it is a poor unit for deciding whether individual uses create value.

When organizational adoption is justified

Organizational support becomes increasingly defensible when six conditions hold simultaneously:

A meaningful bottleneck exists. There is repeated work whose time, quality, capacity or service level matters economically. A cool demo is not a bottleneck.

The intervention improves an outcome that matters. AI changes quality, cycle time, throughput, rework, customer experience, learning or another endpoint—not merely prompt volume or perceived productivity. Controlled studies demonstrate that measured gains can be real, but METR’s randomized developer experiment demonstrates that users can believe themselves substantially faster while objectively becoming slower. 1

Task fit is demonstrated rather than presumed. Dell’Acqua et al.’s BCG experiment found a “jagged technological frontier”: AI improved performance on tasks inside the model’s capabilities but over-reliance could reduce performance outside them. This means superficial task similarity is insufficient evidence of reliability. 2

Verification is economically possible. A system whose outputs require more expert checking than the work it saves can have negative net value even when generation itself is fast. The importance of including verification is reinforced by experiments in which speed and quality move differently and by METR’s finding that workflow and task-selection behavior can materially complicate productivity measurement. 3

Risk can be bounded. Data exposure, consequential decisions, customer harm, security privileges and regulatory requirements are sufficiently controlled. In the EU, some governance is not discretionary: Article 4 of the AI Act requires providers and deployers to take measures supporting an appropriate level of AI literacy, while high-risk deployers face additional obligations including competent human oversight, monitoring and logging. 4

There is a credible mechanism for converting local gains into organizational value. Cutting a worker’s drafting time from 50 to 30 minutes has no automatic P&L value. It matters only when released capacity results in more valuable output, faster service, reduced backlog, avoidance of hiring, reduced external expenditure, improved quality, or some other organizational endpoint. The major causal studies establish task and worker effects; they do not establish a universal conversion from saved minutes to firm profit. 1

Conversely, organizational adoption is not yet justified merely because employees like a tool, colleagues use it, competitors announce deployments, vendor benchmarks improve, a demonstration looks impressive, or a pilot reports “hours saved.”

Doing nothing—or retaining decentralized experimentation—is rational when the potential benefit is small, the task is infrequent, workflow change costs dominate, verification is expensive, errors are difficult to detect, simpler software solves the problem better, or current evidence is too weak to justify irreversible investment.

Evidence on value and heterogeneity

The evidence base is now materially better than it was in 2023, but it remains much stronger for task productivity than for firm-level economic performance.

What has actually been demonstrated

Claimed benefit	Best evidence found	Evidence class	What can reasonably be concluded
Time savings	In Noy & Zhang’s preregistered experiment with 453 professionals, ChatGPT reduced average task time by about 40% on professional writing tasks while increasing evaluated quality. 1	Controlled experiment	Strong evidence for selected writing tasks; not a company-wide productivity estimate
Higher throughput	Randomized field experiments involving 4,867 developers across Microsoft, Accenture and another Fortune 100 company found a 26.08% increase in completed tasks among developers given an AI coding assistant. 5	Controlled field experiments	Strong evidence that coding assistants can increase output in these settings
Higher throughput	A staggered deployment covering 5,172 customer-support agents was associated with about 15% more issues resolved per hour. 3	Quasi-experimental/observational field evidence	Strong real-work evidence, though weaker causally than random assignment
Quality improvement	Noy & Zhang found approximately 18% higher evaluated writing quality alongside faster work. 1	Controlled experiment	Speed and quality need not trade off
Quality improvement—and deterioration	BCG consultants performed better on tasks inside AI’s frontier but could perform worse on tasks outside it because of over-reliance on AI. 2	Preregistered field experiment	Quality effects are task-dependent; productivity measurement without quality is unsafe
Reduced skill bottlenecks	Less experienced/lower-skill customer-support agents obtained the largest gains; experienced/high-skill agents gained less and saw small quality declines in some measures. 3	Field evidence	AI may diffuse elements of top-worker knowledge, but benefits are heterogeneous
Reduced skill bottlenecks in innovation	In P&G’s randomized experiment, individuals using AI produced ideas comparable in evaluated quality to two-person human teams and became less constrained by their own commercial versus R&D specialization. 6	Preregistered field experiment	AI can partly substitute for missing complementary expertise in some ideation work
Learning	Customer-service evidence is consistent with less-experienced workers learning useful patterns from AI suggestions, and gains were particularly strong on relatively uncommon problems. 3	Field evidence, mechanism analysis	Credible evidence of knowledge diffusion; not proof that every copilot builds durable skill
Innovation	P&G participants using GPT-4 produced higher-quality product ideas, but AI did not similarly improve people’s ability to select the best ideas. 6	Controlled field experiment	GenAI can improve generation without improving evaluation
Employee experience	P&G participants receiving AI reported more positive and fewer negative emotions during the task. 6	Controlled treatment, self-reported outcome	Evidence of improved short-run subjective task experience; not evidence of long-run engagement
Customer interaction	In the customer-support deployment, customers became more polite and less likely to ask for a manager. 3	Field observational/quasi-experimental evidence	Some downstream customer interaction benefits are plausible and observed
Cost reduction	Klarna reported substantial marketing-production savings from GenAI, including an estimated $10 million annual reduction in marketing costs; these were company-reported figures, not a controlled evaluation. 7	Company case claim reported externally	Useful hypothesis and case evidence, not causal proof transferable to another company
Firm revenue/profit	The strongest studies reviewed above generally stop at task, worker, quality, service or innovation outcomes rather than randomized firm P&L. 1	Evidence gap	Broad claims that enterprise GenAI “drives revenue” remain much less established than task-level productivity effects

The strongest counterexample to relying on perception comes from METR. Sixteen experienced open-source developers completed 246 real issues in repositories they knew well. With early-2025 AI tools, AI access made them 19% slower, even though they believed AI had accelerated them by roughly 20% and had expected even larger gains beforehand. The authors explicitly caution against generalizing this result to all developers or all AI systems. 8

METR’s follow-up is equally important for management science. By 2026 the researchers concluded that changes in developer willingness to work without AI, task selection and parallel agent usage made later causal estimates substantially less reliable. In other words, as AI becomes embedded in work, measurement itself becomes endogenous: workers may choose different tasks because AI exists. 9

Who benefits most?

There is no universal answer.

Lower-skilled or less-experienced people benefited disproportionately in the customer-support field study, and less-experienced developers obtained greater gains in the three large coding experiments. 3

But that does not imply a general “AI helps novices most” law. The BCG study found gains were contingent on whether the task sat inside the AI capability frontier, while METR found highly experienced developers working in mature codebases were slowed by the tested tools. 2

A more defensible hypothesis is:

AI produces larger gains where it possesses useful task knowledge that the worker does not already have and where its output can be cheaply incorporated and verified.

That proposition fits the customer-service, coding and P&G findings, but should still be treated as a mechanism to test rather than a universal rule. 3

Can productivity rise while quality falls?

Yes.

The customer-support evidence shows different speed and quality effects across skill groups, and the BCG study shows that AI can produce impressive-looking but wrong answers that cause humans to underperform when they trust it outside its capability frontier. 3

Therefore throughput alone is an unsafe success metric.

Does individual productivity imply company productivity?

Not automatically.

The experimental evidence principally demonstrates partial equilibrium effects: one worker or team can perform a task differently when given AI. Organizational value depends on what happens next. An employee who drafts twice as fast may simply wait longer for approval; a developer who produces more code may create more review burden; a support agent who resolves tickets faster may reduce queues—or merely create excess capacity. The causal studies demonstrate the local effects but do not prove that each local improvement survives downstream constraints. 1

Leadership should therefore distinguish:

potential capacity created → capacity actually released → capacity redeployed → economically valuable outcome.

Only the last step constitutes organizational value.

Adoption options, portfolio logic and prerequisites

There is no empirically defensible global ranking of organizational approaches.

The relevant question is which organizational commitment matches the evidence and risk of the use case.

Approach	Becomes sensible when…	Main upside	Main trade-off / failure mode	Reversibility and learning
Laissez-faire experimentation	Uses are low-consequence, data-safe and reversible; objective is discovery	Very low cost; discovers bottom-up use cases	Shadow AI, inconsistent handling of sensitive information, no systematic evidence	Very high reversibility; fast but noisy learning
Approved general enterprise assistant	Useful tasks are numerous but shallow; main need is safer access	Reduces friction and potentially shadow use	Seat cost can become detached from value; broad access may institutionalize novelty	High reversibility unless integrations proliferate
Policy plus contextual training	Employees are already using AI and need clear boundaries	Low implementation burden	Policies without useful alternatives may simply drive use underground	Highly reversible
Role-specific copilots	A family of recurring tasks shares context, tools and quality criteria	Better fit than generic chat; easier measurement	Configuration, integration and maintenance cost	Moderate reversibility
Team-level pilots	Benefits are plausible but heterogeneous or uncertain	Strong learning before irreversible investment	Volunteer teams and enthusiasts may not represent later users	Very high if designed with controls
Targeted workflow redesign	Local task gains are proven but do not improve the end-to-end process	Can turn local productivity into organizational performance	Substantial process/change-management burden	Moderate or low reversibility
Central AI enablement	Common procurement, identity, evaluation, security and model access are repeated across teams	Economies of scale in common capabilities	Central team can become a use-case approval bottleneck	Platform choices create some lock-in
Distributed champions	Use cases are local and contextual while common controls already exist	Faster contextual learning	Quality and governance can diverge between teams	High
Custom AI application	Workflow-specific context/integration creates defensible value unavailable from commodity tools	Higher task fit and control	Engineering, evaluation and maintenance burden; model/vendor churn	Lower
Agentic/automated workflow	Actions are frequent, bounded, observable and recoverable; interfaces are stable	Removes work rather than merely accelerating it	Errors acquire operational side effects; permissions and monitoring become critical	Potentially low once embedded
Delay / do nothing	Evidence is weak, verification expensive, risk high or simpler alternatives exist	Avoids premature cost and lock-in	Opportunity cost if the capability becomes genuinely useful	Maximum reversibility

The empirical basis for avoiding a one-size-fits-all answer is unusually strong: writing, support, consulting and software-development studies produce materially different effects, sometimes in opposite directions. 1

Treat adoption as a portfolio

A more useful portfolio taxonomy is not “low/medium/high AI maturity,” but seven observable classes of work:

Portfolio class	Management response
Promising, low-risk assistance	Permit and instrument; collect outcome data
Harmless experimentation	Leave decentralized unless common tooling materially reduces security/cost
Low-value novelty	Do not industrialize merely because employees use it
High-risk usage	Restrict until controls, legal basis, evaluation and ownership exist
Valuable but bottlenecked workflow	Redesign the workflow rather than buying more assistant licenses
Work where accountable human judgment is the product	Use AI only where it demonstrably strengthens that judgment; avoid substituting fluent output for responsibility
Stable deterministic process	Compare GenAI with conventional rules, scripts, workflow engines or software before choosing a probabilistic system

The critical axes are directly observable: economic value of the work; frequency; task fit; error detectability; harm if wrong; verification cost; data sensitivity; required permissions; dependency on human judgment; integration effort; and reversibility.

There is no need to convert them into an “AI readiness score.”

Which prerequisites are genuinely supported?

Some commonly cited prerequisites deserve more confidence than others.

Task-level evaluation is strongly evidence-backed. The BCG experiment demonstrates why: tasks that look similar can fall on opposite sides of a model’s capability frontier. 2

A baseline and objective outcome measurement are strongly evidence-backed. Without them, METR’s developers would have concluded almost the opposite of the measured result. 8

Representative users matter. Gains differ materially by experience and skill in both customer service and software development. 3

Integration into actual work matters operationally, although its universal causal effect has not been isolated. The strongest field evidence comes from tools embedded in actual support, development and product-innovation work rather than from abstract chatbot benchmarks. 3

AI literacy is now also a legal requirement in the EU, not merely change-management advice. Article 4 requires providers and deployers to take measures supporting an appropriate level of AI literacy, contextualized to staff knowledge, experience and use context. It does not prescribe a universal training course. 4

By contrast, the primary GenAI evidence reviewed does not establish as universal causal prerequisites:

• executive sponsorship;

• a centralized AI Center of Excellence;

• “data maturity” in the abstract;

• organization-wide prompt training;

• internal champion networks;

• protected experimentation time;

• psychological safety programs;

• AI-specific incentive schemes.

Those practices may be very useful in particular organizations, but saying “GenAI adoption fails without executive sponsorship” is materially stronger than the evidence supports. The major field studies did not randomize executive sponsorship, organizational culture or centralization. 3

Likewise, P&G participants given AI received approximately one hour of training and recommended prompting approaches, but because the study did not randomize training separately from AI access, it cannot establish that this training was necessary for the observed gains. 6

The correct distinction is therefore:

implementation plausibility ≠ experimentally established prerequisite.

Risk, governance and workforce

Governance should follow the consequence of the use, not the prestige of the technology.

A worker asking an approved model to reorganize non-sensitive meeting notes and an autonomous agent changing customer accounts may use related underlying models, but they do not warrant remotely similar controls.

A proportional governance map

Exposure	Characteristic examples	Minimum sensible controls
Low-consequence, reversible assistance	Brainstorming, rewriting non-sensitive text, personal organization	Basic acceptable-use rules, AI literacy, approved data boundaries
Internal work product	Analysis, code, internal documents, knowledge retrieval	Access controls, quality expectations, source/verification practices where appropriate, logging sufficient for incident investigation
Externally consequential output	Customer responses, contracts, published claims, production code	Named accountable owner, systematic evaluation, human review or equivalent control, incident path, change monitoring
Decisions affecting people or material rights	Some hiring, evaluation, access-to-service or other consequential decisions	Legal classification, documented oversight, testing for relevant harms, stronger logging and governance; EU high-risk rules may apply depending on use
Agentic execution	Systems able to send, modify, approve, purchase, deploy or alter records	Least privilege, bounded action space, approval gates for irreversible actions, comprehensive auditability, rollback/kill path, operational monitoring

NIST’s Generative AI Profile explicitly provides a risk-management extension to the voluntary AI Risk Management Framework for organizations deploying GenAI; it should be treated as an optional governance framework rather than law. 10

What is compliance versus optional good practice in the EU?

As of the October 3, 2026 snapshot, the EU AI Act is in force and has been amended by Regulation (EU) 2026/1744; current interpretation should therefore use the consolidated text rather than the original 2024 act alone. 11

AI literacy: Article 4 requires AI providers and deployers to take measures supporting sufficient AI literacy among staff and others operating AI systems on their behalf, taking their knowledge, experience and context into account. This requirement is broader than high-risk AI. 4

High-risk AI: where a company is a deployer of a high-risk system, Article 26 requires appropriate technical and organizational measures, competent and authorized human oversight, monitoring of system operation and, where logs are under the deployer’s control, retention for an appropriate period of at least six months unless other law provides otherwise. 12

For high-risk workplace systems, Article 26 additionally requires employers to inform workers’ representatives and affected workers before putting the system into service or use. This is a specific notification requirement; organizations should not assume that satisfying it exhausts separate labor-law duties that may apply in their jurisdiction. 12

The Act’s Annex III covers categories of potentially high-risk systems, including employment-related applications; classification nevertheless depends on the actual system and intended use rather than the fact that something uses generative AI. 13

Transparency: Article 50 transparency requirements for certain AI systems began applying on August 2, 2026, and the Commission issued implementation guidance in July/August 2026. 14

Personal data: the GDPR remains independently applicable where personal data are processed. The EDPB’s Opinion 28/2024 specifically addresses legal basis, anonymization/pseudonymization and personal-data processing in development and deployment of AI models. 15

Automated individual decision-making may additionally engage GDPR rules and established EDPB guidance; “there is a human somewhere in the process” should not be assumed to resolve those questions without examining what decision the system actually makes and the meaningfulness of human involvement. 16

Acceptable-use policy should answer operational questions

A useful policy should not merely say “use AI responsibly.” It should make it possible for an employee to determine:

What information may I provide? What systems may I use? What outputs require verification? What decisions may AI influence? What may it execute? When must I disclose AI involvement? Who remains accountable? Where do I report an incident or surprising behavior?

An organization with strong rules but no usable sanctioned alternative may still generate shadow use. Conversely, issuing an approved assistant without specifying data and decision boundaries converts a procurement decision into a de facto governance decision.

Human review is not automatically a sufficient control

The BCG evidence is particularly relevant here: humans can over-rely on plausible AI output and become worse at tasks beyond AI’s frontier. 2

“Human in the loop” therefore has value only if the reviewer:

• has the competence to detect important mistakes;

• receives enough information and time to review;

• has actual authority to override the system; and

• is not asked to validate such a volume of output that review becomes ceremonial.

For EU high-risk AI, competence, training, authority and support are explicitly part of the human-oversight obligation. 12

Pilot-to-scale and measurement framework

A pilot should answer a decision question, not demonstrate that the model can produce output.

“People used it and liked it” is enough evidence for neither workflow redesign nor a large license purchase.

What a serious pilot should test

Before starting, specify:

The counterfactual. What would happen without the AI intervention?

The unit of analysis. Task, employee, customer case, team, workflow or process.

The primary outcome. For example, customer resolution quality—not “AI usage.”

One paired productivity/quality measure. Faster output must be examined alongside errors, rework, customer outcomes or another quality endpoint. The experimental literature shows why both dimensions matter. 1

Guardrails. Security events, privacy breaches, inappropriate outputs, compliance failures, customer harm or unacceptable employee effects.

Representative tasks and users. Do not test only enthusiast volunteers doing tasks selected because AI is known to perform them well. Heterogeneous worker effects and METR’s later task-selection problem make representativeness a first-order methodological concern. 3

Verification effort. Count checking and correction, not just generation time.

Total operating cost. Licenses/inference, integration, training, evaluation, human review, support, incidents and ongoing maintenance.

Pre-agreed decision thresholds. Define the smallest benefit worth paying for and the guardrail deterioration that is unacceptable before seeing the result.

Where feasible, random assignment or phased rollout with a valid comparison group is substantially more informative than before/after surveys. The strongest organizational studies reviewed used randomization, preregistration or staggered deployment precisely to separate AI effects from ordinary differences among workers and tasks. 1

Evidence that justifies scale

A positive pilot should clear four separate gates.

Outcome gate: economically meaningful improvement appears in the intended endpoint, not merely usage or satisfaction.

Safety/quality gate: relevant quality, customer, security, privacy and compliance guardrails remain acceptable.

Economics gate: benefit survives verification, integration, model, support and maintenance costs.

Transferability gate: the result persists across representative tasks/users and is not dependent on a handful of champions or a deliberately easy task subset.

Only then should the next question be scale.

Even then, scale should generally proceed in stages with continued measurement, because models, product interfaces, usage patterns and task selection can change after initial rollout. METR’s later measurement difficulties are a concrete example of the work system adapting around AI. 9

Evidence that should trigger redesign

Redesign rather than abandonment when:

• workers become faster but the end-to-end cycle does not improve;

• gains occur primarily in one skill group, suggesting differentiated deployment;

• verification or rework absorbs the generation savings;

• AI shifts a bottleneck from creation to review;

• workers benefit on common cases but fail on unusual high-consequence cases;

• a useful assistant requires so much repetitive human transfer that workflow integration is the actual bottleneck.

The customer-support and BCG evidence supports precisely this kind of differentiated treatment: effects vary by worker experience and task capability frontier. 3

Evidence that should trigger stopping

Stop or materially narrow a use when:

net value remains negative after a reasonable learning period; quality deterioration persists; critical errors are too difficult to detect; verification cost overwhelms savings; users abandon it for real work; simpler deterministic automation performs better; required controls destroy the economics; or the organization cannot establish a lawful/acceptable operating model.

A stop decision is not an AI failure. It is portfolio discipline.

Observable measurement framework

Indicator	What it tells leadership	Main limitation / gaming risk
Task time	Local efficiency	Workers may select easier tasks; may omit verification time. METR demonstrates both perception and task-selection problems. 8
End-to-end cycle time	Whether local gains survive downstream bottlenecks	Can be affected by unrelated queue/process changes
Throughput	Capacity actually realized	Easy to inflate by lowering task complexity or quality
Task quality	Whether output remains useful/correct	Requires credible rubric, human evaluation or objective endpoint
Error rate	Reliability	Only measures detectable/defined errors
Rework	Hidden downstream cost	Rework may occur in another team and be poorly attributed
Verification effort	Cost of making probabilistic output trustworthy	Often invisible in product telemetry
Customer outcome	Whether internal gains create external value	May lag and have many confounders
Sustained active use	Whether employees continue finding the tool useful	Usage is a mediator, not an outcome; mandates can game it
Workflow penetration	Whether AI is actually part of targeted work	High penetration can mean forced adoption rather than benefit
Employee experience	Friction, perceived control, stress and usefulness	Self-reports may diverge sharply from actual productivity, as METR showed. 8
Skill/learning retention	Whether AI complements or substitutes capability	Requires testing performance without assistance over time
Incidents / near misses	Safety and governance performance	Under-reporting and rare-event statistics
Abandonment	Reveals use cases that fail in practice	Does not explain whether cause is poor tool, workflow or training
Net operating cost	Whether economics work	Hidden review, integration and support costs are easy to omit
Business endpoint	Revenue, cost, backlog, retention, conversion, quality or risk outcome the project was intended to affect	Attribution becomes harder as distance from the AI intervention increases

Do not aggregate these into a single AI score. A system that scores well on speed and badly on consequential errors should not be allowed to average those facts into a reassuring 82/100.

Leadership questions and decision reversals

Before spending materially more, redesigning roles, committing to long-term platforms or forcing broad adoption, leadership should be able to give concrete answers to the following.

Area	Question
Value	What specific business or work-system problem are we trying to change?
Value	What happens if we make no AI investment for the next twelve months?
Value	What measured outcome would constitute enough improvement to justify the full operating cost?
Alternative	Have we compared GenAI with process simplification, search, templates, rules or conventional automation?
Task fit	On which exact task classes does AI outperform the current process, and where does it fail?
Heterogeneity	Which employees benefit, which do not, and are we hiding those differences inside an average?
Quality	How would we know when AI produces a fluent but substantively wrong answer?
Verification	Who can reliably verify it, how long does that take, and is verification included in the economics?
Workflow	If a worker saves an hour, what organizational outcome actually changes?
Evidence	Is the claim based on measured behavior, self-report, a vendor benchmark or a company anecdote?
Pilot validity	Were pilot users and tasks representative of the eventual rollout population?
Governance	What information may enter the system and what actions may leave it?
Accountability	Who owns the result when the AI-assisted process fails?
Legal	Does this use process personal data, affect people materially, fall into a regulated sector, or potentially meet an AI Act high-risk category?
Autonomy	What is the maximum damage the system can cause before a human can intervene?
Economics	What does the system cost after licenses/inference, integration, checking, support, monitoring and model change?
Lock-in	What must be rebuilt if the chosen model or vendor becomes unsuitable next year?
Workforce	Are we improving work, moving work to reviewers, removing learning opportunities, or changing staffing assumptions?
Scale	What new evidence would justify the next increment of investment?
Stop	What observable result would cause us to terminate the initiative?

A leadership team unable to answer those questions does not yet have an adoption problem. It has an evidence problem.

Decision reversals

Many sensible-sounding AI recommendations become wrong when underlying conditions change.

Apparently sensible advice	It is sensible when…	Reverse the advice when…
“Give everyone an approved copilot.”	Valuable low-risk tasks are broad and recurring	Benefits are concentrated in a few roles and unused licenses dominate economics
“Let teams experiment independently.”	Mistakes are reversible and information is non-sensitive	Data exposure, external actions or consequential decisions appear
“Centralize AI adoption.”	Common security, procurement, evaluation and integration dominate	Central approval becomes the bottleneck for highly contextual low-risk uses
“Train everyone.”	There are common behaviors everyone genuinely needs, including applicable literacy obligations	Uses remain undefined and training becomes generic theater; prefer contextual just-in-time learning
“Keep a human in the loop.”	Human review can actually detect errors and reviewers have authority/capacity	Review becomes rubber-stamping, cannot scale, or experts cannot reliably verify outputs. BCG’s over-reliance result is the warning case. 2
“Scale the successful pilot.”	Tasks/users were representative and effects survive verification and cost	Enthusiasts self-selected, task mix was cherry-picked, or behavior changes after rollout. 9
“AI helps junior workers most.”	AI contains useful expertise they lack and output is verifiable, as in support and some coding contexts. 3	The model’s knowledge is unreliable for the task or junior staff lack the ability to detect failures
“AI will make experts even faster.”	The tool removes genuine expert bottlenecks	Interaction overhead exceeds assistance value; METR found exactly this for its experienced-developer sample. 8
“Automate after the copilot proves useful.”	Actions are bounded, repeatable and recoverable	A copilot is useful precisely because human judgment resolves ambiguity
“Build a custom solution.”	Proprietary workflow/context/integration creates durable value	Commodity products rapidly cover the need and maintenance becomes the dominant cost
“Standardize on one vendor.”	Integration and governance economies outweigh optionality	Model capability changes quickly or the use portfolio requires materially different providers
“Use adoption rate as the KPI.”	Only as a diagnostic for discoverability/friction	The organization begins forcing usage; adoption then measures compliance rather than value
“Scale because workers report large time savings.”	Self-report is corroborated by objective outcomes	Objective evidence disagrees; METR found workers perceived sizable acceleration while measured completion became slower. 8
“Redesign the whole workflow immediately.”	Local evidence is strong and the workflow bottleneck is understood	Task fit is still uncertain; preserve reversibility through a limited assistant/pilot first
“Wait until AI is more mature.”	Error cost and integration burden exceed current benefit	Experiments are cheap, reversible and create decision-relevant learning now

The deeper principle is that governance intensity, organizational commitment and evidence burden should rise together. The mistake is not decentralization or centralization; it is applying either indiscriminately.

Claim-to-primary-source table

The table below separates the central decision claims from weaker market narratives and traces them to primary research or official sources.

Decision-relevant claim	Primary source	Evidence type	What the source actually establishes	Important boundary
GenAI can materially accelerate professional writing while increasing quality	Noy & Zhang, Science / Stanford research record. 1	Randomized, preregistered experiment	453 professionals; average completion time fell roughly 40%, evaluated output quality rose roughly 18%	Selected occupational writing tasks, not whole jobs or firms
GenAI can raise real-world customer-support productivity	Brynjolfsson, Li & Raymond, QJE, “Generative AI at Work.” 3	Staggered field deployment / quasi-experimental evidence	5,172 agents; roughly 15% more issues resolved per hour on average	One organizational/work setting
GenAI’s effects differ substantially by worker skill/experience	Brynjolfsson, Li & Raymond. 3	Field evidence	Larger gains for less-experienced/lower-skill agents; small gains and some quality decline among top workers	Does not prove novices always gain most
Coding assistants can increase completed development work	Cui et al., Management Science, 2026. 5	Three randomized field experiments	Combined n=4,867; treatment increased completed tasks by 26.08%	Specific assistant/work environments
AI can also slow experienced developers	METR randomized productivity study. 8	Randomized field experiment	16 experienced open-source developers, 246 real issues; early-2025 AI made completion roughly 19% slower	Narrow sample and tool era; authors explicitly warn against broad generalization
Self-reported productivity can disagree sharply with actual productivity	METR. 8	Experimental measurement	Developers believed AI made them about 20% faster despite measured slowdown	Particularly strong warning against survey-only ROI
AI has a “jagged” task capability frontier	Dell’Acqua et al., Organization Science, 2026. 2	Preregistered randomized field experiment at BCG	AI improved performance on some realistic consulting tasks but could cause over-reliance and worse performance on tasks outside its frontier	Frontier changes as models improve
A productivity intervention can simultaneously create quality risk	Dell’Acqua et al.; Brynjolfsson et al. 2	Controlled/field evidence	Speed/output and correctness/quality need not move together	Requires task-specific quality definitions
GenAI can partly substitute for complementary expertise during ideation	Dell’Acqua et al., P&G “Cybernetic Teammate.” 6	Preregistered randomized field experiment, 791 professionals	Individual+AI performance could match human teams; AI reduced functional specialization differences	Product ideation, not universal team substitution
Better generation does not imply better judgment	P&G experiment. 6	Randomized field evidence	AI improved idea quality but did not comparably improve selection of the best ideas	Particularly relevant to decision-support claims
AI can improve short-run subjective work experience	P&G experiment. 6	Experimental treatment with self-reported affect	Positive emotions increased and negative emotions declined in AI conditions	Does not establish retention, engagement or long-run well-being
One-hour AI training was part of a successful intervention, but its causal necessity is unknown	P&G study design. 6	Experimental protocol	AI participants received training/prompt guidance	Training itself was not independently randomized
Later AI-productivity measurement can be distorted by worker/task selection	METR 2026 update. 9	Methodological update from continuing field study	Willingness to work without AI, task selection and parallel agent use complicated causal estimation	Highlights a measurement issue rather than estimating universal impact
EU providers/deployers have an AI-literacy obligation	EU AI Act Article 4, official AI Act Service Desk. 4	Binding regulation / official text	Organizations must take measures supporting sufficient contextual AI literacy	“Sufficient” is context-dependent; not a mandate for one standardized course
EU high-risk AI deployers have stronger operational obligations	AI Act Article 26. 12	Binding regulation	Includes competent human oversight, monitoring, certain logging and workplace information obligations	Applies when the system/use is classified as high-risk
High-risk workplace AI requires workers and representatives to be informed before use	AI Act Article 26(7). 12	Binding regulation	Explicit workplace information obligation for high-risk AI	Other national/EU labor duties may be separate
AI-related personal-data processing remains subject to GDPR analysis	EDPB Opinion 28/2024. 15	Official supervisory-board opinion	Addresses legal basis, anonymity and processing of personal data in AI development/deployment	Exact obligations depend on processing roles and facts
NIST’s GenAI Profile is a risk-management framework, not a legal mandate	NIST AI RMF Generative AI Profile. 10	Official voluntary framework	Provides GenAI-specific risk-management guidance	Voluntary unless incorporated into contracts/rules elsewhere

Overall evidentiary judgment

The evidence now supports a proposition narrower—and more useful—than “GenAI makes companies productive”:

Generative AI can generate substantial causal improvements in particular cognitive tasks and work settings. Those effects are highly heterogeneous across tasks, workers and human–AI configurations; objectively negative effects are possible; and evidence that these task gains systematically become firm-level revenue, profit or broad organizational performance remains substantially weaker. 1

Accordingly, the strongest leadership posture is neither “AI-first” nor “AI-resistant.” It is evidence-first and portfolio-based:

decentralize discovery where consequences are low; centralize common controls and infrastructure where economies or risks justify it; test outcomes rather than enthusiasm; redesign workflows only after locating the real bottleneck; increase governance as consequences and autonomy increase; and treat stopping a weak use case as successful capital allocation rather than failed transformation.