# 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](https://pubsonline.informs.org/doi/10.1287/mnsc.2025.02523) 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](https://academic.oup.com/qje/article/140/2/889/7990658) 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](https://www.census.gov/library/stories/2026/05/ai-use-businesses.html) 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](https://www.fedsmallbusiness.org/reports/survey/2026/2026-report-on-employer-firms) 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](https://pubsonline.informs.org/doi/10.1287/mnsc.2025.02523) ### 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](https://www.theaustralian.com.au/business/technology/aussies-bring-their-own-ai-to-work-according-to-microsoft-linkedin/news-story/1b8537633e3a93be3fbcfabb5acfb539) **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](https://academic.oup.com/qje/article/140/2/889/7990658) **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](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai) **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](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838) 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](https://www.census.gov/library/stories/2026/05/ai-use-businesses.html) ## 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](https://pubsonline.informs.org/doi/10.1287/mnsc.2025.02523) 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](https://www.stlouisfed.org/on-the-economy/2026/apr/why-does-ai-adoption-differ-so-much-countries) 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](https://pubsonline.informs.org/doi/10.1287/mnsc.2025.02523) ### 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](https://www.census.gov/library/stories/2026/05/ai-use-businesses.html) 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](https://www.census.gov/library/stories/2026/05/ai-use-businesses.html) 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](https://www.census.gov/library/stories/2026/05/ai-use-businesses.html) ### 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](https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Use_of_artificial_intelligence_in_enterprises) 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](https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Use_of_artificial_intelligence_in_enterprises) 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](https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Use_of_artificial_intelligence_in_enterprises) ### 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](https://www.fedsmallbusiness.org/reports/survey/2026/2026-report-on-employer-firms) 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](https://www.fedsmallbusiness.org/reports/survey/2026/2026-report-on-employer-firms) 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](https://www.fedsmallbusiness.org/reports/survey/2026/2026-report-on-employer-firms) ## 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](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838) “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](https://academic.oup.com/qje/article/140/2/889/7990658) 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](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/) 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](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838) 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](https://www.stlouisfed.org/on-the-economy/2026/apr/why-does-ai-adoption-differ-so-much-countries) 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](https://www.stlouisfed.org/on-the-economy/2026/apr/why-does-ai-adoption-differ-so-much-countries) “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](https://www.theaustralian.com.au/business/technology/aussies-bring-their-own-ai-to-work-according-to-microsoft-linkedin/news-story/1b8537633e3a93be3fbcfabb5acfb539) 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](https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Use_of_artificial_intelligence_in_enterprises) 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](https://academic.oup.com/qje/article/140/2/889/7990658) 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](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838) 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](https://www.census.gov/library/stories/2026/05/ai-use-businesses.html) 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](https://www.stlouisfed.org/on-the-economy/2026/apr/why-does-ai-adoption-differ-so-much-countries) 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](https://pubsonline.informs.org/doi/10.1287/mnsc.2025.02523) **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](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838) **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](https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Use_of_artificial_intelligence_in_enterprises) **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](https://www.fedsmallbusiness.org/reports/survey/2026/2026-report-on-employer-firms) ### 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](https://academic.oup.com/qje/article/140/2/889/7990658) ## 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](https://academic.oup.com/qje/article/140/2/889/7990658) 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](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838) 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](https://pubsonline.informs.org/doi/10.1287/orsc.2025.20702) ### 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](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838) 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](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/) 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](https://academic.oup.com/qje/article/140/2/889/7990658) ### 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](https://pubsonline.informs.org/doi/10.1287/mnsc.2025.02523) 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](https://www.nber.org/papers/w33777) 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](https://www.chicagobooth.edu/research/center-for-applied-artificial-intelligence/research/our-faculty-research/2026/still-waters-rapid-currents) ### 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](https://www.fedsmallbusiness.org/reports/survey/2026/2026-report-on-employer-firms) 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](https://arxiv.org/abs/2507.09089) 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](https://www.fedsmallbusiness.org/reports/survey/2026/2026-report-on-employer-firms) 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](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/) ### 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](https://www.chicagobooth.edu/research/center-for-applied-artificial-intelligence/research/our-faculty-research/2026/still-waters-rapid-currents)
**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](https://arxiv.org/abs/2507.09089)
**Employee workarounds / shadow AI** **Supported descriptively.** Microsoft/LinkedIn provider research documented extensive use of personally sourced AI among knowledge workers. [5](https://www.theaustralian.com.au/business/technology/aussies-bring-their-own-ai-to-work-according-to-microsoft-linkedin/news-story/1b8537633e3a93be3fbcfabb5acfb539)
**Fragmented individual practices** **Supported indirectly.** High employee use can coexist with low formal integration; small firms report much experimentation and partial integration. [4](https://www.fedsmallbusiness.org/reports/survey/2026/2026-report-on-employer-firms)
**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](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/)
**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](https://academic.oup.com/qje/article/140/2/889/7990658)
**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](https://www.chicagobooth.edu/research/center-for-applied-artificial-intelligence/research/our-faculty-research/2026/still-waters-rapid-currents)
**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](https://www.fedsmallbusiness.org/reports/survey/2026/2026-report-on-employer-firms)
### 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](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838) **Gross benefit is consumed by verification.** Generation becomes cheaper but checking, correcting and integrating outputs does not; the bottleneck simply moves. [7](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838) **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](https://www.fedsmallbusiness.org/reports/survey/2026/2026-report-on-employer-firms) **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](https://www.stlouisfed.org/on-the-economy/2026/apr/why-does-ai-adoption-differ-so-much-countries) **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](https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Use_of_artificial_intelligence_in_enterprises) **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](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/) **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](https://www.census.gov/library/stories/2026/05/ai-use-businesses.html) 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](https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Use_of_artificial_intelligence_in_enterprises) 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](https://academic.oup.com/qje/article/140/2/889/7990658) 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](https://academic.oup.com/qje/article/140/2/889/7990658) 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](https://www.fedsmallbusiness.org/reports/survey/2026/2026-report-on-employer-firms) 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](https://www.stlouisfed.org/on-the-economy/2026/apr/why-does-ai-adoption-differ-so-much-countries) 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](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838) 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](https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Use_of_artificial_intelligence_in_enterprises) 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](https://www.census.gov/library/stories/2026/05/ai-use-businesses.html) 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](https://pubsonline.informs.org/doi/10.1287/mnsc.2025.02523) 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](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](https://www.census.gov/library/stories/2026/05/ai-use-businesses.html) 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](https://www.census.gov/library/stories/2026/05/ai-use-businesses.html) 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](https://www.fedsmallbusiness.org/reports/survey/2026/2026-report-on-employer-firms) 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](https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Use_of_artificial_intelligence_in_enterprises) 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](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](https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Use_of_artificial_intelligence_in_enterprises) 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](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](https://academic.oup.com/qje/article/140/2/889/7990658) Staggered workplace deployment / field empirical study 5,172 customer-support agents Feb. 4, 2025 2026-10-03 [https://doi.org/10.1093/qje/qjae044](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](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838) Preregistered randomized field experiment 758 BCG consultants Published online Mar. 11, 2026 2026-10-03 [https://doi.org/10.1287/orsc.2025.21838](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](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838) Randomized experiment BCG consultants Mar. 11, 2026 2026-10-03 [https://doi.org/10.1287/orsc.2025.21838](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](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/) 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://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/) ; [https://doi.org/10.48550/arXiv.2507.09089](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](https://www.stlouisfed.org/on-the-economy/2026/apr/why-does-ai-adoption-differ-so-much-countries) 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](https://www.stlouisfed.org/on-the-economy/2026/apr/why-does-ai-adoption-differ-so-much-countries) 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](https://pubsonline.informs.org/doi/10.1287/orsc.2025.20702) Organizational field experiment Procter & Gamble professionals / product-development tasks 2026 publication 2026-10-03 [https://doi.org/10.1287/orsc.2025.20702](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](https://www.nber.org/papers/w33777) Linked worker/employer surveys + administrative records Denmark, AI-exposed occupations/workplaces Revised 2026 2026-10-03 [https://www.nber.org/papers/w33777](https://www.nber.org/papers/w33777) ; [https://doi.org/10.3386/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](https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Use_of_artificial_intelligence_in_enterprises) Official statistics + regulation EU enterprises 2025–2026 2026-10-03 [https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai](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](https://www.fedsmallbusiness.org/reports/survey/2026/2026-report-on-employer-firms) 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](https://www.fedsmallbusiness.org/reports/survey/2026/2026-report-on-employer-firms) 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](https://pubsonline.informs.org/doi/10.1287/mnsc.2025.02523) **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](https://www.census.gov/library/stories/2026/05/ai-use-businesses.html) **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](https://pubsonline.informs.org/doi/10.1287/orsc.2025.20702) **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](https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Use_of_artificial_intelligence_in_enterprises) **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](https://www.census.gov/library/stories/2026/05/ai-use-businesses.html) **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](https://academic.oup.com/qje/article/140/2/889/7990658) **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](https://academic.oup.com/qje/article/140/2/889/7990658) **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](https://academic.oup.com/qje/article/140/2/889/7990658) **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](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838) ### 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](https://learn.chatgpt.com/docs/computer-use) **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](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/) **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](https://academic.oup.com/qje/article/140/2/889/7990658) **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](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai) **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](https://www.chicagobooth.edu/research/center-for-applied-artificial-intelligence/research/our-faculty-research/2026/still-waters-rapid-currents) **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](https://pubsonline.informs.org/doi/10.1287/mnsc.2025.02523) 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](https://www.fedsmallbusiness.org/reports/survey/2026/2026-report-on-employer-firms)