# What Empirical Research Actually Establishes About Generative AI and Data-Adjacent Knowledge Work

**Research snapshot: October 3, 2026**

The evidence supports a narrower—and more interesting—conclusion than either “AI is a productivity revolution” or “AI is eliminating knowledge work.”

Generative AI can produce **large causal improvements on some bounded knowledge-work tasks**, including writing, customer support, coding, consulting, translation, and some analytical work. But the effect is **task-contingent, expertise-contingent, model-contingent, and workflow-contingent**. In credible studies, AI ranges from making work substantially faster and better to making experienced workers measurably slower or causing them to make more wrong decisions. [1](https://www.science.org/doi/10.1126/science.adh2586)

The strongest recurring heterogeneity result is not “AI helps everyone equally.” In several occupations, initially weaker or less-experienced workers obtain larger **relative** productivity gains. That pattern appears in customer support, consulting, professional translation, and several software-development studies. But it is not a universal law, it does not imply experts are unimportant, and it certainly does not imply junior jobs become unnecessary. [2](https://academic.oup.com/qje/article/140/2/889/7990658)

Learning is even less one-directional. A sustained customer-support deployment provides credible evidence that workers learned from AI and retained some benefit during later AI outages. A randomized educational study, by contrast, found that unrestricted GPT assistance improved performance while available yet harmed subsequent unaided performance. The important empirical distinction is therefore **assisted performance versus acquired human capability**; most workplace studies measure only the former. [2](https://academic.oup.com/qje/article/140/2/889/7990658)

Most claims about **data-team size, disappearance of junior roles, analysts becoming strategic advisers, data engineering becoming architecture-only, analyst/analytics-engineer convergence, or AI-native teams requiring dramatically fewer people remain inadequately tested**. Almost all of the causal productivity literature stops at the task or individual worker. Direct experiments at the team level are rare, and direct studies of organizational headcount, career ladders, or end-to-end data-team output are rarer still. The P&G team experiment is unusually informative precisely because it shows why extrapolation is dangerous: an individual with AI matched a two-person human team on one product-innovation exercise, but AI-equipped teams did not dramatically outperform human teams on average, and the experiment says nothing about whether an organization should employ fewer people. [3](https://pubsonline.informs.org/doi/10.1287/orsc.2025.20702)

The strongest labor-market evidence is also not one-sided. Online freelance markets show displacement in exposed work; Danish administrative data show widespread adoption but essentially no detectable effect on earnings or recorded hours through the observation period; and 2026 U.S. payroll data show a large and widening *descriptive* early-career employment shortfall in AI-exposed occupations, concentrated in hiring, while the authors explicitly decline to interpret it as a causal AI estimate. [4](https://pubsonline.informs.org/doi/abs/10.1287/orsc.2023.18441?journalCode=orsc)

My bottom-line assessment is therefore:

> **The evidence is strong that GenAI changes the production function of particular knowledge-work tasks. It is moderate that it changes individual workflows and skill gradients. It is emerging but contradictory on learning and task reallocation. It is weak on persistent team redesign, career ladders, occupational convergence, organizational output, and headcount.**

## Evidence boundaries and the direct data-team record

### Level-of-analysis map

The most important methodological error in this literature is to move silently between levels of analysis. They need to be treated as different outcomes.

<table header-row="true">
	<tr>
		<td>Level</td>
		<td>What would count as evidence?</td>
		<td>What current GenAI evidence actually establishes</td>
	</tr>
	<tr>
		<td>**Task**</td>
		<td>Time to complete a specified task; completion rate; quality/error score</td>
		<td>**Strongest evidence base.** Numerous randomized experiments show positive, zero, and negative effects depending on task and setting. [1](https://www.science.org/doi/10.1126/science.adh2586)</td>
	</tr>
	<tr>
		<td>**Worker**</td>
		<td>Total output over meaningful periods; quality-adjusted throughput; learning/retention</td>
		<td>**Moderate evidence.** Customer support and corporate coding experiments are strongest; persistence is much less studied. [2](https://academic.oup.com/qje/article/140/2/889/7990658)</td>
	</tr>
	<tr>
		<td>**Team**</td>
		<td>Team output, coordination, division of labor, combined quality</td>
		<td>**Sparse.** P&G's experiment is a major exception; it is a one-day innovation exercise rather than a longitudinal operating team. [3](https://pubsonline.informs.org/doi/10.1287/orsc.2025.20702)</td>
	</tr>
	<tr>
		<td>**Occupation**</td>
		<td>Changes in occupational tasks, employment or wages</td>
		<td>**Emerging, mostly observational.** Effects differ materially by platform, country, age and whether AI appears substitutive or complementary. [4](https://pubsonline.informs.org/doi/abs/10.1287/orsc.2023.18441?journalCode=orsc)</td>
	</tr>
	<tr>
		<td>**Organization**</td>
		<td>Revenue, cost, service volume, quality, headcount, span of control, restructuring</td>
		<td>**Very weak for causal GenAI effects.** Tool rollouts are studied much more often than organization-level output or structure. Danish evidence finds major adoption without detectable earnings/hour changes; customer-support evidence measures operational worker output, not firm profits or optimal staffing. [2](https://academic.oup.com/qje/article/140/2/889/7990658)</td>
	</tr>
	<tr>
		<td>**Labor market**</td>
		<td>Employment, hiring, separations, wages across firms/occupations</td>
		<td>**Mixed and causally difficult.** Freelance displacement, Danish null effects, and U.S. early-career divergence coexist. [4](https://pubsonline.informs.org/doi/abs/10.1287/orsc.2023.18441?journalCode=orsc)</td>
	</tr>

</table>

Thus the inference

**“AI makes task X 30% faster → worker is 30% more productive → team needs 30% fewer workers”**

is not supported. The first arrow requires evidence about workers' task portfolios, bottlenecks, demand, idle time, quality-control burden, and reallocation of saved time. The second requires evidence about complementarities, coordination, workload growth, demand response and organization design. Current studies seldom observe all of those. Indeed, the Stanford labor-market paper explicitly notes that higher occupational productivity can raise *or* reduce employment depending, among other things, on demand elasticity. [5](https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf)

### Direct data-team evidence versus transferred evidence

The direct empirical literature on actual data work is substantially thinner than the literature on software development, writing and customer service.

The strongest recent direct item I found is Ali Merali's preregistered 2025 experiment involving **more than 500 consultants, data analysts and managers using 13 LLMs**. Across models, later model progress was associated with shorter task completion time—approximately 8% per year of model progress—with larger improvements on non-agentic analytical tasks than on workflows requiring models to operate tools. Because the result is a preprint, because occupations are pooled in the headline results, and because this is a controlled task environment rather than longitudinal data-team production, it should be treated as promising direct evidence, not an estimate of data-team productivity. [6](https://arxiv.org/abs/2512.21316)

Small HCI studies provide useful process evidence rather than productivity estimates. *Data Formulator 2*, for example, studied eight participants with varying data-science expertise performing visualization/transformation work; participants developed explicit strategies for verification, correction and iteration around AI-generated transformations. That is evidence that AI changes analytical workflow and creates verification work; it is not evidence that eight-person analytics teams can be reduced to five. [7](https://arxiv.org/abs/2408.16119)

A related participatory study, Drosos et al.'s *“It's like a rubber duck that talks back”: Understanding Generative AI-Assisted Data Analysis Workflows*, directly examines AI-assisted data-analysis practices, but is qualitative/HCI evidence rather than a controlled productivity field experiment. Its existence is itself illustrative: the direct data-analysis literature has so far been stronger on *how people interact with AI* than on long-run quality-adjusted organizational output. [8](https://www.microsoft.com/en-us/research/wp-content/uploads/2025/01/lee_2025_ai_critical_thinking_survey.pdf)

For **SQL specifically**, I did not find a mature body of randomized field evidence measuring professional SQL analysts' end-to-end productivity, error rates, downstream corrections and later unaided competence. There are many Text-to-SQL benchmarks and autonomous-agent demonstrations, but system benchmark accuracy is not equivalent to analyst productivity. A 2025 system paper reporting a large speed comparison between an autonomous database-research system and a professional analyst, for example, is a system evaluation rather than a randomized human productivity study. [9](https://arxiv.org/abs/2512.14622)

Likewise, the empirical base for professional **spreadsheet analysis, BI dashboard development, analytics engineering and data engineering as complete jobs** remains thin. Claims about these roles therefore rely substantially on transfers from software engineering, consulting, financial analysis and general knowledge work.

A useful transferability map is:

<table header-row="true">
	<tr>
		<td>Data-team activity</td>
		<td>Best empirical analogue</td>
		<td>Transfer confidence</td>
	</tr>
	<tr>
		<td>Python/R coding, transformation code, dbt-style development</td>
		<td>Software-development experiments</td>
		<td>**Moderate–high for coding subtasks**, lower for architecture, operations and domain modeling. [10](https://doi.org/10.48550/ARXIV.2302.06590)</td>
	</tr>
	<tr>
		<td>SQL generation</td>
		<td>LLM coding + emerging data-analysis HCI</td>
		<td>**Moderate for query drafting; low for end-to-end analyst performance**. [7](https://arxiv.org/abs/2408.16119)</td>
	</tr>
	<tr>
		<td>Exploratory analysis</td>
		<td>Merali; Data Formulator; financial-analysis studies</td>
		<td>**Moderate but immature.** [6](https://arxiv.org/abs/2512.21316)</td>
	</tr>
	<tr>
		<td>Writing findings and summaries</td>
		<td>Professional-writing experiments</td>
		<td>**High for bounded writing tasks**, lower for decision impact. [1](https://www.science.org/doi/10.1126/science.adh2586)</td>
	</tr>
	<tr>
		<td>Information retrieval/synthesis</td>
		<td>Consulting and financial-analysis studies</td>
		<td>**Moderate.** More information does not necessarily produce better judgment. [11](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838)</td>
	</tr>
	<tr>
		<td>Business decision support</td>
		<td>BCG, P&G, AI-advice experiments</td>
		<td>**Moderate for mechanisms; low for data-team organization design.** [12](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838)</td>
	</tr>
	<tr>
		<td>Forecasting</td>
		<td>Financial-analyst evidence</td>
		<td>**Emerging and mixed.** [13](https://papers.ssrn.com/sol3/Delivery.cfm/5821142.pdf?abstractid=5821142&mirid=1)</td>
	</tr>
	<tr>
		<td>Software/data engineering</td>
		<td>Corporate coding RCTs, CodeFuse, METR</td>
		<td>**Moderate**, with unusually strong evidence of task-context heterogeneity. [14](https://www.microsoft.com/en-us/research/publication/the-effects-of-generative-ai-on-high-skilled-work-evidence-from-three-field-experiments-with-software-developers/)</td>
	</tr>

</table>

### Analytical-work evidence

The emerging financial-analyst literature illustrates why “better information retrieval” should not be equated with “better decisions.” A 2025–2026 SSRN study exploiting FactSet's GenAI integration finds AI-associated analyst reports became richer—using more distinct information sources, broader topical coverage and more analytical methods—and more timely. Yet forecast accuracy worsened under higher information-processing demands; the authors interpret this as evidence that information acquisition can cease to be the bottleneck while human synthesis becomes one. This is adjacent rather than direct evidence for corporate data analysts, and it remains a preprint, but the mechanism is highly relevant. [13](https://papers.ssrn.com/sol3/Delivery.cfm/5821142.pdf?abstractid=5821142&mirid=1)

That mechanism is consistent with the broader evidence: GenAI can lower the marginal cost of generating queries, alternatives, summaries or analyses **without lowering the marginal cost of deciding which result is correct or consequential**. BCG's experiment and P&G's idea-selection result provide causal examples of that distinction. [12](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838)

## Productivity, quality, expertise, learning and review

### Productivity evidence

The productivity evidence is real. It is also much more conditional than headline averages suggest.

Noy and Zhang's preregistered experiment with 453 college-educated professionals found that ChatGPT cut completion time on occupation-specific writing assignments by roughly **40%** while raising evaluator-rated quality by roughly **18%**. Lower initial performers gained more, compressing the performance distribution. These were short, well-defined writing assignments, however—not months of whole-job production. [1](https://www.science.org/doi/10.1126/science.adh2586)

Dell'Acqua et al.'s BCG experiment is perhaps the clearest demonstration of a **jagged technological frontier**. Among 758 consultants, GPT-4 users working on tasks within the model's capability frontier completed roughly **12% more tasks**, completed them about **25% faster**, and generated outputs rated more than **30% higher in quality**. But on a deliberately difficult business-strategy problem outside the frontier, AI users were about **19 percentage points less likely to reach the correct answer**. The same technology that augmented one set of professional tasks degraded another. [11](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838)

Brynjolfsson, Li and Raymond provide the strongest sustained workplace evidence among the canonical studies. In a staggered deployment covering **5,172 customer-support agents and roughly three million chats**, AI assistance increased issues resolved per hour by around **15% overall**. Less-skilled and less-experienced agents gained around 30%, whereas the highest-skilled workers saw much smaller productivity gains and, on some quality measures, slight declines. [2](https://academic.oup.com/qje/article/140/2/889/7990658)

Coding produces some of the largest—and most contradictory—effects. Peng et al.'s controlled GitHub Copilot experiment found recruited developers completed a fixed JavaScript HTTP-server task **55.8% faster** with Copilot. [15](https://doi.org/10.48550/ARXIV.2302.06590) Corporate field RCTs at Microsoft, Accenture and a Fortune 100 company, pooling **4,867 developers**, found a **26.08% increase in completed tasks**, with higher adoption and larger gains among less-experienced developers. [14](https://www.microsoft.com/en-us/research/publication/the-effects-of-generative-ai-on-high-skilled-work-evidence-from-three-field-experiments-with-software-developers/) A 12-week CodeFuse quasi-experiment involving 1,219 programmers found sharply higher code output and task completion, concentrated among junior programmers. [16](https://www.sciencedirect.com/science/article/pii/S1572308926000458)

Then METR randomized 246 real issues from mature open-source repositories among **16 highly experienced developers working on codebases they knew extremely well**. AI availability—primarily Cursor Pro with Claude 3.5/3.7 Sonnet—made completion **19% slower**. Developers had predicted a 24% speedup and, even after participating, believed AI had made them approximately 20% faster. Final PR quality was similar across conditions. The tasks averaged about two hours and included real bug fixes, features and refactors. [17](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/)

This is not a contradiction to be resolved by choosing one's favorite paper. The studies are examining materially different production functions:

- fixed greenfield coding versus mature repositories;

- unfamiliar versus highly familiar code;

- juniors versus experts;

- code completion versus agent/chat workflows;

- prototype completion versus code satisfying repository-specific style, testing and documentation requirements;

- early-2023 versus early-2025 models and tooling. [18](https://doi.org/10.48550/ARXIV.2302.06590)

METR's 2026 follow-up is especially instructive about persistence and measurement. Later raw data suggested AI may have become beneficial for the same class of developers, but the researchers explicitly concluded the estimate was unreliable because workers who expected high AI value increasingly refused to participate in AI-disallowed conditions, developers withheld tasks they did not want to perform without AI, agentic workflows enabled parallel work that made time measurement difficult, and AI itself altered which tasks developers chose to undertake. Among 57 developers and more than 800 tasks, the experimental design had effectively become endogenous to the technology. [19](https://metr.org/blog/2026-02-24-uplift-update/)

That is an important research result in its own right: sufficiently integrated AI may change not merely **task speed**, but **task selection and concurrency**, making old productivity metrics increasingly incomplete. [19](https://metr.org/blog/2026-02-24-uplift-update/)

### Quality evidence

There is no defensible general statement that AI makes work “the same quality but faster.”

Three regimes appear empirically:

**Faster and better.** Noy–Zhang's writing study, BCG's within-frontier tasks and Merali's translation experiment all found improvements in speed and evaluated quality. In Merali's experiment with 300 professional translators and 1,800 tasks, scaling to more capable models improved both completion speed and graders' scores. [1](https://www.science.org/doi/10.1126/science.adh2586)

**Faster but quality-neutral or ambiguously measured.** Some coding studies primarily use completion or output metrics rather than long-horizon maintenance cost. The corporate developer RCTs find substantially more completed tasks, but that does not establish proportional increases in quality-adjusted software value. [14](https://www.microsoft.com/en-us/research/publication/the-effects-of-generative-ai-on-high-skilled-work-evidence-from-three-field-experiments-with-software-developers/)

**Faster/richer but worse on the ultimate decision.** BCG's outside-frontier task and the FactSet financial-analysis study are the clearest examples. More fluent output or more available information can coexist with worse correctness or forecasting. [11](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838)

A recurring weakness is that many task experiments end when the submitted artifact is graded. They therefore do **not** measure subsequent peer review, debugging, stakeholder clarification, data-quality investigation, maintenance or downstream correction. METR is unusually relevant because the completion-time measure includes actual developer interaction with AI and the final work had to satisfy repository expectations; it found that review/waiting/interaction costs could dominate generation gains in its particular expert setting. [17](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/)

For data teams, that omission matters disproportionately. A generated query that executes is not necessarily semantically correct; an analysis can be statistically valid yet answer the wrong business question; a transformation can pass tests yet encode a mistaken business definition. Direct empirical estimates of these **downstream verification costs for professional analytics pipelines remain scarce**. Small data-analysis HCI studies do observe verification and correction becoming explicit parts of AI-assisted workflow. [7](https://arxiv.org/abs/2408.16119)

### Heterogeneity by expertise

The claim **“lower performers benefit more”** is one of the better-supported cross-study regularities, but only if stated conditionally.

Noy and Zhang found larger gains for initially weaker writers. [1](https://www.science.org/doi/10.1126/science.adh2586) BCG found larger percentage gains for its lower-performing consultant half, though higher performers also benefited inside the frontier. [11](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838) Customer-support gains were dramatically larger for novice/lower-skill agents. [2](https://academic.oup.com/qje/article/140/2/889/7990658) Merali's professional translation experiment reports gains around four times larger for lower-skilled workers. [20](https://arxiv.org/abs/2409.02391) The pooled corporate coding RCTs also found larger productivity improvements among less-experienced developers. [14](https://www.microsoft.com/en-us/research/publication/the-effects-of-generative-ai-on-high-skilled-work-evidence-from-three-field-experiments-with-software-developers/) CodeFuse's significant gains were concentrated among entry-level programmers. [16](https://www.sciencedirect.com/science/article/pii/S1572308926000458)

This is meaningful evidence for **skill compression on some assisted tasks**. It is not evidence that expertise loses value.

First, the comparisons are predominantly **relative gains**, not final absolute performance. An expert improving by 10% may still produce substantially more valuable work than a novice improving by 30%. Second, many AI systems appear most useful where they can supply patterns, language or knowledge that the worker previously lacked; marginal value can therefore naturally fall with prior mastery. Third, expertise becomes particularly important when the model exits its capability frontier. BCG's consultants were not protected against overreliance, but the experiment demonstrates that the problem shifts from production to correct boundary recognition. [11](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838)

The expert-developer METR result adds another mechanism: when humans possess deep repository-specific context that the model lacks, interacting with AI can impose a **context-transfer and review tax**. That is almost the mirror image of customer support, where AI encoded practices from high-performing colleagues and effectively supplied novices with organization-specific tacit knowledge. [17](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/)

Hence:

**“Experts gain less from AI” — conditionally supported.**

**“Experts do not benefit from AI” — contradicted by several studies.**

**“Experts use AI better” — inadequately tested as a general claim.**

**“AI erases expertise differences” — contradicted as a general claim; it compresses differences on some tasks but not reliably outside the model's frontier.** [21](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838)

### Junior-worker evidence

For less-experienced workers, three different outcomes have to be separated: assisted productivity, learning, and employment.

On **assisted productivity**, evidence is fairly strong that juniors sometimes benefit disproportionately. Customer support, corporate coding and CodeFuse all show this pattern. [2](https://academic.oup.com/qje/article/140/2/889/7990658)

On **learning**, customer support is unusually encouraging. AI-exposed new agents reached performance levels comparable with substantially more experienced untreated agents much sooner. During occasional system outages, previously exposed workers still handled chats faster, and the residual advantage tended to increase with accumulated AI exposure. This provides rare workplace evidence that at least part of the improvement represented acquired capability rather than pure real-time substitution. [2](https://academic.oup.com/qje/article/140/2/889/7990658)

On **employment**, the evidence is much weaker causally. The August 2026 Stanford payroll study finds that employment among 22–25-year-olds in more AI-exposed occupations stood about **19% below the level it would have reached had it kept pace with less-exposed peers**, with the divergence operating primarily through lower hiring rather than elevated separations. But the authors explicitly describe these as early **descriptive indicators rather than causal estimates**; they note pre-existing divergent trends, attenuation with education controls, and differences between the ADP sample and national benchmarks. [5](https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf)

Moreover, Danish evidence creates an important counterweight. Humlum and Vestergaard link AI adoption surveys to administrative employment records and do not find the exposed-occupation changes translating into detectable worker earnings or hours; their later discussion specifically cautions against attributing observed early-career declines mechanically to workplace GenAI adoption. [22](https://bfi.uchicago.edu/wp-content/uploads/2025/04/BFI_WP_2025-56-1.pdf)

So the current evidence supports:

> **AI can disproportionately improve juniors' performance on some tasks. It does not establish that junior jobs are redundant.**

Indeed, if AI raises novice productivity while simultaneously substituting for some entry-level tasks, the net effect on junior employment depends on demand expansion, job redesign and the value of apprenticeship—none of which can be inferred from task RCTs.

### Expert-worker evidence

Experts do not exhibit one universal behavior.

Highly familiar open-source experts were slowed in METR. [17](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/) Top customer-support workers experienced little gain, with some quality/customer-satisfaction measures deteriorating slightly. [2](https://academic.oup.com/qje/article/140/2/889/7990658) Higher-performing BCG consultants nevertheless improved substantially on tasks inside GPT-4's frontier. [11](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838)

The hypothesis that experts' future role is primarily “reviewing AI” has intuitive support but surprisingly little direct longitudinal evidence. METR demonstrates that review and AI interaction can itself become costly. P&G demonstrates that AI-assisted participants can become worse at identifying which of their generated ideas is actually best. BCG demonstrates that skilled consultants can fail to override a plausible but misleading model recommendation. None establishes that senior staff become an organization-wide review bottleneck. [17](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/)

Direct GenAI evidence that senior data professionals systematically detect AI errors better than juniors remains notably scarce.

### Learning evidence versus deskilling

The central empirical lesson is that **performance with AI and learning from AI are different dependent variables**.

The customer-support field deployment provides evidence for positive workplace learning: novices apparently internalized some high-performer practices and retained productivity benefits when assistance temporarily disappeared. [2](https://academic.oup.com/qje/article/140/2/889/7990658)

Bastani et al.'s randomized high-school mathematics study demonstrates the opposite possibility. Students using an unrestricted GPT interface performed better while the tool was available, but subsequently performed **17% worse than controls when AI access was removed**. A specially designed GPT tutor with guardrails preserved much of the assisted benefit without the same post-removal penalty. This is education rather than workplace evidence, so transfer to analysts must be conservative; nevertheless, it directly falsifies the assumption that higher assisted performance necessarily means more human learning. [23](https://www.pnas.org/doi/10.1073/pnas.2422633122)

The CHI 2025 survey of **319 knowledge workers** found self-reported changes in critical-thinking effort when people used GenAI, with reliance patterns related to confidence in AI. But it is a survey about reported cognitive effort, not a longitudinal measurement of skill loss, and therefore cannot establish atrophy. [8](https://www.microsoft.com/en-us/research/wp-content/uploads/2025/01/lee_2025_ai_critical_thinking_survey.pdf)

Accordingly, **“GenAI deskills workers” is too strong**. The evidence supports several weaker mechanisms:

AI can reduce occasions for independent practice; AI can reduce cognitive effort in some workflows; unguarded assistance can impair later unaided performance in an educational experiment; and reliance can create errors. Conversely, an embedded workplace assistant can transmit expert practices and accelerate learning. [23](https://www.pnas.org/doi/10.1073/pnas.2422633122)

Actual **skill atrophy** implies deterioration over sustained time. Most current experiments last minutes, hours, days or weeks. We do not yet have strong multi-year evidence showing experienced analysts, data engineers or software engineers losing independently measured core skills because of GenAI.

### Automation bias, overreliance and quality assurance

Overreliance is better established than long-run deskilling.

BCG's outside-frontier treatment is a particularly relevant professional example: access to GPT-4 made consultants materially less likely to reach the correct strategic recommendation because many incorporated misleading AI analysis. [11](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838)

Klingbeil, Grützner and Schreck's incentivized behavioral experiment randomized **319 participants** to AI advice, expert advice or control conditions. Participants followed AI advice even when it conflicted with contextual information and their own prior assessment; AI-advised participants' own payoffs were on average 22% below control in the relevant comparison. Because the task was an abstract trust game rather than professional data review, transfer is mechanistic rather than occupational—but it cleanly establishes that an “AI” label can induce costly overreliance. [24](https://doi.org/10.1016/J.CHB.2024.108352)

P&G provides a subtler form of review failure. AI improved idea generation, yet AI-assisted individuals and teams were worse than unaided teams at identifying their own best idea: unaided teams selected their highest-rated idea roughly half the time versus around 37% in the AI conditions. Thus increased production of candidate outputs can make **evaluation**, rather than generation, the bottleneck. [3](https://pubsonline.informs.org/doi/10.1287/orsc.2025.20702)

For data teams, the practical analogue is obvious but should remain an inference: a model may cheaply generate ten plausible SQL queries, statistical explanations or transformation strategies, while the human still needs enough domain and technical knowledge to recognize the one that encodes the correct business semantics. The direct data-analysis studies' observed verification/correction behavior is consistent with that mechanism. [7](https://arxiv.org/abs/2408.16119)

Evidence on moderators is uneven:

<table header-row="true">
	<tr>
		<td>Moderator</td>
		<td>What can be claimed</td>
	</tr>
	<tr>
		<td>**Expertise**</td>
		<td>Expertise does not eliminate overreliance; BCG professionals still followed wrong AI. Whether experts systematically detect errors better than juniors in GenAI-generated data work is inadequately tested. [11](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838)</td>
	</tr>
	<tr>
		<td>**Task difficulty / model frontier**</td>
		<td>**Strong evidence.** AI can flip from positive to negative when task requirements exceed model capability. [11](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838)</td>
	</tr>
	<tr>
		<td>**Confidence in AI**</td>
		<td>Associated with reduced self-reported critical-thinking effort in knowledge-worker research, but causal performance implications remain less established. [8](https://www.microsoft.com/en-us/research/wp-content/uploads/2025/01/lee_2025_ai_critical_thinking_survey.pdf)</td>
	</tr>
	<tr>
		<td>**Explanations**</td>
		<td>Earlier human-AI evidence suggests explanations alone do not reliably calibrate reliance; cognitive forcing can outperform passive explanation. The GenAI-specific workplace evidence remains limited. The CHI review explicitly identifies this distinction. [8](https://www.microsoft.com/en-us/research/wp-content/uploads/2025/01/lee_2025_ai_critical_thinking_survey.pdf)</td>
	</tr>
	<tr>
		<td>**Uncertainty displays**</td>
		<td>Plausibly important, but little direct evidence yet establishes optimal uncertainty presentation for GenAI-assisted professional data review.</td>
	</tr>
	<tr>
		<td>**High historical AI accuracy**</td>
		<td>A theoretically serious vigilance problem, but longitudinal professional evidence directly measuring declining vigilance as AI accuracy rises is still sparse.</td>
	</tr>

</table>

The implication is not “humans must check everything manually.” That would erase much of the productivity gain. The unresolved research problem is **how to allocate human review selectively according to error cost, model uncertainty, task novelty and worker expertise**.

## Coding, task composition, teams and organizational consequences

### Coding and engineering evidence

Software engineering is the richest empirical analogue for data engineering and analytics engineering, but transfer should be made at the **task** level.

The evidence establishes that code assistants can accelerate fixed, self-contained implementation; can increase completed-task counts during corporate rollouts; can disproportionately help less-experienced engineers; and can substantially increase measured code output. [10](https://doi.org/10.48550/ARXIV.2302.06590)

It equally establishes that AI may slow highly experienced engineers on mature, context-heavy repositories, particularly where satisfying tacit codebase conventions and reviewing generated suggestions are material parts of the task. [17](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/)

That maps plausibly onto data engineering as follows:

**High-transfer tasks:** boilerplate transformations, tests, API usage, syntax, query drafting, documentation, unfamiliar library use and local debugging.

**Lower-transfer claims:** data architecture, semantic model design, ownership boundaries, migration sequencing, distributed-system trade-offs, incident accountability, stakeholder negotiation and deciding what data should exist in the first place.

The latter have barely been tested in randomized GenAI studies. Therefore the industry claim **“data engineering is shifting from coding to architecture”** is a reasonable hypothesis about task mix, not an empirical conclusion.

The METR follow-up adds another emerging effect: agentic coding can induce **parallelism**. Developers may delegate work to an agent and perform a second task while it runs, making simple “elapsed time on task” measures inadequate. AI can also alter which tasks developers choose to attempt. That means future engineering productivity measurement needs to capture throughput across portfolios, not merely stopwatch time for predefined tickets. [19](https://metr.org/blog/2026-02-24-uplift-update/)

### Task-composition evidence

There is now meaningful evidence that AI changes what workers spend time doing, not merely how quickly they do a fixed task.

Humlum and Vestergaard's Danish surveys and administrative linkage find widespread employer initiatives, enterprise tools and training, while workers report time savings and the creation of **new AI-related tasks**, including oversight and integration work. Most users report reallocating saved time into other work rather than simply reducing hours. Yet these reported workflow changes had not translated into detectable aggregate earnings or recorded-hour effects over the study horizon. [22](https://bfi.uchicago.edu/wp-content/uploads/2025/04/BFI_WP_2025-56-1.pdf)

Customer-support AI changed communication as well as speed: lower-skill workers' language became more similar to that of high-performing workers, consistent with diffusion of organizational practices through the model. [2](https://academic.oup.com/qje/article/140/2/889/7990658)

P&G found that AI changed functional content. Commercial specialists produced more technically balanced ideas and R&D specialists more commercially balanced ones, reducing differences between functional groups. [3](https://pubsonline.informs.org/doi/10.1287/orsc.2025.20702)

FactSet-enabled financial analysts drew on broader information and more analytical approaches. [13](https://papers.ssrn.com/sol3/Delivery.cfm/5821142.pdf?abstractid=5821142&mirid=1) METR's later developers reported changing which issues they attempted and working concurrently while agents operated. [19](https://metr.org/blog/2026-02-24-uplift-update/)

Taken together, there is credible emerging evidence for **task reallocation and boundary blurring**. There is much less evidence that this has stabilized into new job descriptions or occupations.

### Team-level evidence

P&G's *Cybernetic Teammate* experiment is the major direct team-level study in the evidence base. The preregistered experiment involved 826 employees, with 791 in the reported analysis, spanning R&D and commercial functions. Participants were assigned to work individually or in cross-functional two-person teams, with or without AI, on a product-innovation task. [3](https://pubsonline.informs.org/doi/10.1287/orsc.2025.20702)

Relative to individuals without AI, human teams improved solution quality by about **0.24 standard deviations**, individuals with AI by about **0.37 SD**, and teams with AI by about **0.39 SD**. Thus an individual with AI performed approximately as well on average as a two-person team without AI on this specific task. [3](https://pubsonline.informs.org/doi/10.1287/orsc.2025.20702)

But two details matter enormously.

First, **teams with AI did not significantly outperform human-only teams by anything resembling a multiplicative amount on average quality**, although they had a greater probability of producing top-decile ideas. Second, the task was a one-day virtual innovation exercise, not the sustained operation of an established team with production systems, incidents, mentoring, coordination costs, backlog dynamics and organizational accountability. [3](https://pubsonline.informs.org/doi/10.1287/orsc.2025.20702)

Therefore:

> “On one product-innovation task, an AI-equipped individual achieved output quality comparable to a human dyad” is directly supported.

> “AI lets companies replace two-person teams with one person” is not.

I found no similarly strong randomized evidence showing that professional **data teams** using GenAI require fewer analysts, analytics engineers or data engineers while maintaining long-term output, quality, resilience and development of future senior staff.

### Organizational adoption and outcomes

There is a large gap between **access to a tool** and **organization-level productivity**.

The customer-support study is unusually strong because AI was integrated into real production work and outcomes were observed over time. It establishes higher individual operational throughput and some quality/learning effects. It still does not directly estimate the optimal number of employees, firm profitability or long-run organization design. [2](https://academic.oup.com/qje/article/140/2/889/7990658)

The three software-company RCTs improve external validity by randomizing coding-assistant access in ordinary business operations. Yet the headline outcome is completed tasks per developer, not business revenue, product adoption, total engineering cost or future maintenance burden. [14](https://www.microsoft.com/en-us/research/publication/the-effects-of-generative-ai-on-high-skilled-work-evidence-from-three-field-experiments-with-software-developers/)

Humlum and Vestergaard's 25,000-worker, roughly 7,000-workplace Danish evidence provides the reverse perspective: employer encouragement, enterprise solutions and training dramatically increase adoption, but administrative earnings and hours show essentially null effects over the observation window; the estimates rule out large short-run changes. [22](https://bfi.uchicago.edu/wp-content/uploads/2025/04/BFI_WP_2025-56-1.pdf)

This suggests an important distinction:

**tool adoption → usage** is empirically well established;

**usage → task gains** is established for selected tasks;

**task gains → firm-level output or lower headcount** remains mostly unmeasured.

### Job design, substitution and demand expansion

GenAI can simultaneously substitute for tasks and complement workers. Current evidence contains examples of both.

Hui, Reshef and Zhou find reduced employment and earnings among freelancers in occupations most affected by ChatGPT and image-generation systems. The marketplace allows demand for discrete contracted tasks to adjust rapidly, making substitution unusually visible. Interestingly, high prior freelancer quality did not reliably protect workers; the paper finds suggestive evidence that top freelancers could also be disproportionately affected. [4](https://pubsonline.informs.org/doi/abs/10.1287/orsc.2023.18441?journalCode=orsc)

By contrast, customer support shows augmentation: the AI system increased human agents' throughput rather than eliminating the observed worker role during the study. [2](https://academic.oup.com/qje/article/140/2/889/7990658)

Denmark shows another outcome: adoption and modest reported time savings accompanied task reallocation but not detectable changes in total hours or earnings. [22](https://bfi.uchicago.edu/wp-content/uploads/2025/04/BFI_WP_2025-56-1.pdf)

The 2026 Stanford payroll paper separates occupations according to real-world AI usage characterized as more substitutive versus complementary. Young-worker employment declines are concentrated in occupations classified as having more automative usage; complementary occupations show flatter or increasing employment, particularly among experienced workers. But the authors stress that their estimates are descriptive rather than a causal effect of AI. [5](https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf)

This is the right way to treat “demand expansion.” A productivity improvement can reduce labor required per unit while simultaneously lowering price, improving quality, enabling new outputs and increasing demand. The net employment effect is an equilibrium outcome, not a mathematical corollary of task speed. The Stanford paper explicitly states this ambiguity. [5](https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf)

No current study establishes that a 30% improvement in analytics-task productivity leads to anything like a 30% reduction in analyst employment.

## Careers, occupational boundaries and the labor market

### Apprenticeship and career development

The apprenticeship question may be strategically more important than the immediate productivity question, but the evidence is weak.

Customer support demonstrates one possible positive mechanism: AI can act as a scalable carrier of high-performer practices, allowing novices to climb the performance curve faster. [2](https://academic.oup.com/qje/article/140/2/889/7990658)

The education literature demonstrates a negative mechanism: when AI substitutes for the cognitive practice required to develop a skill, assisted performance can rise while unaided competence deteriorates. [23](https://www.pnas.org/doi/10.1073/pnas.2422633122)

Software-development productivity studies largely measure output **while assistance is present**. They do not follow a cohort of junior engineers for several years to determine whether AI-assisted juniors become equally capable senior engineers, acquire architectural judgment faster, or fail to develop skills previously learned through debugging and routine implementation. [10](https://doi.org/10.48550/ARXIV.2302.06590)

The labor-market evidence raises the stakes without resolving the mechanism. The Stanford data show reduced hiring among young workers in exposed occupations; they cannot tell us whether employers are removing apprenticeship tasks because AI performs them, whether other macro factors drive the pattern, or what happens to the future supply of senior expertise. [5](https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf)

Therefore the claim **“AI is breaking the junior-to-senior pipeline” is plausible but inadequately tested**.

### Role convergence

There is empirical evidence for **task-level convergence**, not occupational merger.

At P&G, commercial and technical specialists using AI produced proposals with more similar mixtures of business and technical content. Non-core specialists with AI could sometimes reach performance comparable to groups containing a core-domain participant. [3](https://pubsonline.informs.org/doi/10.1287/orsc.2025.20702)

In customer support, novice workers' language shifted toward patterns used by top performers. [2](https://academic.oup.com/qje/article/140/2/889/7990658)

These are examples of AI reducing functional differences in the *content of a task*. They do not show role titles disappearing, compensation structures converging, reporting lines merging or firms combining analyst and analytics-engineer occupations.

I found no strong empirical evidence that **analyst and analytics-engineer occupations have actually merged because of GenAI**. Nor is there strong evidence that data engineering has empirically become an “architecture rather than coding” occupation.

Thus:

**skill overlap — supported in selected settings;**

**task redistribution — emerging evidence;**

**role-title convergence — little evidence;**

**actual occupational merger — unsupported.**

### Labor-market and hiring evidence

The three most informative labor-market studies point in different directions because they examine different markets.

**Online freelancing:** Hui, Reshef and Zhou use the launch of ChatGPT and image generators to identify short-run changes on a large online labor platform. More-exposed freelance occupations experience reductions in employment and earnings. Online marketplaces are particularly responsive because jobs are granular and contracts are easy to reallocate, so transfer to salaried corporate data teams is limited. [4](https://pubsonline.informs.org/doi/abs/10.1287/orsc.2023.18441?journalCode=orsc)

**Denmark:** Humlum and Vestergaard survey roughly 25,000 workers across 11 exposed occupations and link responses to administrative records. Through the observed period they find rapid adoption, employer investment, reported time savings and new tasks, yet **no detectable average effect on earnings or recorded hours**, with confidence intervals excluding large effects. [22](https://bfi.uchicago.edu/wp-content/uploads/2025/04/BFI_WP_2025-56-1.pdf)

**United States:** Brynjolfsson, Chandar and Chen use ADP payroll records covering millions of workers through June 2026. They find no economy-wide mass displacement, but a large and widening relative employment shortfall for 22–25-year-olds in exposed occupations—19% on their simple kept-pace comparison—concentrated in reduced hiring rather than separations. They explicitly label the findings descriptive and document potential confounders and pre-trends. [5](https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf)

It would be bad causal inference to cite the Stanford number as proof that “AI eliminated 19% of entry-level jobs.” The paper itself says not to do that. [5](https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf)

It would be equally wrong to use Denmark's null average effect as proof that AI has no labor-market impact anywhere. Freelance markets provide direct evidence of adjustment where task contracting is unusually flexible. [4](https://pubsonline.informs.org/doi/abs/10.1287/orsc.2023.18441?journalCode=orsc)

### AI skills

There is strong evidence that **ability and willingness to use AI affects realized productivity**, but much weaker evidence that “prompt engineering” constitutes a durable, separately priced occupational skill.

Corporate coding experiments find adoption heterogeneity; less-experienced developers both used the assistant more and gained more. [14](https://www.microsoft.com/en-us/research/publication/the-effects-of-generative-ai-on-high-skilled-work-evidence-from-three-field-experiments-with-software-developers/) Denmark shows that employer encouragement, enterprise access and training strongly increase chatbot adoption and reduce demographic adoption gaps. Yet even workers reporting sizable AI benefits do not show clearly higher subsequent earnings over the available horizon. [22](https://bfi.uchicago.edu/wp-content/uploads/2025/04/BFI_WP_2025-56-1.pdf)

METR's later experiment also highlights selection: by 2026, many developers valued AI enough to refuse experimental conditions requiring them to work without it. That is evidence of revealed preference for AI-assisted workflows, though not a causal wage premium for “AI skill.” [19](https://metr.org/blog/2026-02-24-uplift-update/)

As of the research cutoff, I would distinguish:

**AI tool literacy:** clearly becoming relevant to task performance and adoption.

**ability to select appropriate tasks/models and verify output:** plausibly a durable complement, with meaningful supporting evidence from failure cases.

**prompt syntax as a standalone skill:** no strong evidence that it is a stable long-run human-capital category.

**labor-market wage premium caused by GenAI skill:** not yet cleanly established for data occupations.

### Business, communication and judgment skills

The popular claim that “technical skill matters less; business understanding matters more” gets ahead of the evidence.

BCG does show that knowing when an AI's apparently competent answer is inappropriate is valuable. [11](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838) P&G shows that GenAI can help technical and commercial professionals cross functional boundaries, while human evaluation remains difficult. [3](https://pubsonline.informs.org/doi/10.1287/orsc.2025.20702) FactSet's analyst evidence suggests information acquisition becomes cheaper while synthesis and judgment can become binding constraints. [13](https://papers.ssrn.com/sol3/Delivery.cfm/5821142.pdf?abstractid=5821142&mirid=1)

Those results support a **complementarity hypothesis** for problem framing, contextual judgment and verification.

They do not show that technical skill becomes less valuable in wages, hiring or causal performance. Indeed, reliable review of generated SQL, transformations, metrics or forecasts itself requires substantial technical and domain understanding.

A more defensible statement is:

> As generation and retrieval become cheaper, some workflows shift their bottleneck toward specification, contextual interpretation and verification.

That is different from claiming:

> Business skills now matter more than technical skills.

The latter requires relative returns data that the current literature does not provide.

## Claim audit and contradictory evidence

### Common industry claims

<table header-row="true">
	<tr>
		<td>Claim</td>
		<td>Classification</td>
		<td>Evidence verdict</td>
	</tr>
	<tr>
		<td>**AI will shrink data teams.**</td>
		<td>**Plausible but inadequately tested**</td>
		<td>No causal study directly shows sustained GenAI adoption reducing internal data-team headcount while holding output/quality constant. Freelance substitution does not establish corporate-team restructuring. [4](https://pubsonline.informs.org/doi/abs/10.1287/orsc.2023.18441?journalCode=orsc)</td>
	</tr>
	<tr>
		<td>**Analysts will become strategic advisers.**</td>
		<td>**Plausible but inadequately tested**</td>
		<td>Generation/retrieval become cheaper and judgment can become a bottleneck, but longitudinal job-design evidence for analysts is absent. [13](https://papers.ssrn.com/sol3/Delivery.cfm/5821142.pdf?abstractid=5821142&mirid=1)</td>
	</tr>
	<tr>
		<td>**SQL skill will stop mattering.**</td>
		<td>**Unsupported**</td>
		<td>Direct professional SQL evidence is thin, and verification of semantic correctness remains a real human task. Data-analysis HCI studies show active verification/correction rather than elimination of understanding. [7](https://arxiv.org/abs/2408.16119)</td>
	</tr>
	<tr>
		<td>**Business understanding becomes more important than technical skill.**</td>
		<td>**Plausible but inadequately tested**</td>
		<td>Context/judgment clearly matter, but no convincing evidence estimates the relative labor-market returns to business versus technical skill after GenAI. [11](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838)</td>
	</tr>
	<tr>
		<td>**Juniors benefit most from AI.**</td>
		<td>**Conditionally supported**</td>
		<td>Larger relative gains recur in support, consulting, translation and coding, but not as a universal occupational law. [2](https://academic.oup.com/qje/article/140/2/889/7990658)</td>
	</tr>
	<tr>
		<td>**Junior roles will disappear.**</td>
		<td>**Unsupported as a general claim / labor-market warning signs exist**</td>
		<td>U.S. exposed occupations show a sizable descriptive young-worker hiring shortfall, but causality is unresolved; Danish adoption data do not show comparable causal labor effects. [5](https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf)</td>
	</tr>
	<tr>
		<td>**Seniors become review bottlenecks.**</td>
		<td>**Plausible but inadequately tested**</td>
		<td>Review costs can erase AI speed gains and evaluation can become difficult; no team-level evidence shows senior review queues as an organizational bottleneck. [17](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/)</td>
	</tr>
	<tr>
		<td>**AI deskills workers.**</td>
		<td>**Mixed / inadequately tested longitudinally**</td>
		<td>Educational harm and reduced self-reported cognitive effort coexist with workplace evidence of accelerated learning. [23](https://www.pnas.org/doi/10.1073/pnas.2422633122)</td>
	</tr>
	<tr>
		<td>**AI accelerates learning.**</td>
		<td>**Conditionally supported**</td>
		<td>Strong customer-support evidence, but not universal; unguarded educational assistance can harm unaided learning. [2](https://academic.oup.com/qje/article/140/2/889/7990658)</td>
	</tr>
	<tr>
		<td>**Analyst and analytics-engineer roles will converge.**</td>
		<td>**Plausible but inadequately tested**</td>
		<td>Some cross-functional/task-boundary convergence exists; occupational merger does not. [3](https://pubsonline.informs.org/doi/10.1287/orsc.2025.20702)</td>
	</tr>
	<tr>
		<td>**Data engineering shifts from coding to architecture.**</td>
		<td>**Plausible but inadequately tested**</td>
		<td>Coding assistance is demonstrably productive in several contexts; no longitudinal occupational study establishes a durable architecture-heavy task mix. [14](https://www.microsoft.com/en-us/research/publication/the-effects-of-generative-ai-on-high-skilled-work-evidence-from-three-field-experiments-with-software-developers/)</td>
	</tr>
	<tr>
		<td>**Teams using AI need fewer people.**</td>
		<td>**Unsupported**</td>
		<td>P&G shows AI-equipped individuals can match human dyads on one innovation task; that is not an employment or staffing experiment. [3](https://pubsonline.informs.org/doi/10.1287/orsc.2025.20702)</td>
	</tr>
	<tr>
		<td>**AI-native teams dramatically outperform traditional teams.**</td>
		<td>**Unsupported**</td>
		<td>P&G team+AI had interesting top-end gains but not dramatic average superiority over human teams. Direct longitudinal team evidence remains scarce. [3](https://pubsonline.informs.org/doi/10.1287/orsc.2025.20702)</td>
	</tr>

</table>

### Why apparently contradictory studies disagree

**Coding is the clearest contradiction.** GitHub Copilot reduced time by 55.8% on a bounded JavaScript task, corporate field experiments raised completed tasks by roughly 26%, CodeFuse strongly raised output, while METR's experienced developers became 19% slower. [10](https://doi.org/10.48550/ARXIV.2302.06590)

The likely reconciliation is not “one paper is wrong.” Task familiarity, repository maturity, participant expertise, model generation, quality standards, AI interface and outcome measurement differ radically. METR itself explicitly warns against generalizing its result to most developers. [17](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/)

**Skill compression versus expert value** is similarly conditional. AI often lifts low performers more, but experts still improve on some tasks; on deeply familiar tasks, experts may have more context than the model and therefore incur greater interaction cost. [21](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838)

**Learning versus deskilling** depends on whether AI exposes workers to useful expert patterns or substitutes for the practice that produces knowledge. Customer support looks like the former; unguarded math assistance demonstrates the latter. [2](https://academic.oup.com/qje/article/140/2/889/7990658)

**Quality versus quantity** changes with the evaluation metric. AI can create more ideas, reports or code while making selection, synthesis or later verification harder. That explains why P&G can find better idea generation alongside poorer idea selection, and why FactSet users can create richer reports alongside worse forecasts under cognitive load. [3](https://pubsonline.informs.org/doi/10.1287/orsc.2025.20702)

**Labor-market findings** differ because online marketplaces, salaried Danish employment and U.S. payroll panels have different adjustment margins. Freelance contracts can disappear almost immediately; salaried firms can redirect workers to additional tasks; economy-wide payroll patterns are affected by monetary policy, remote work, industry cycles, education and cohort composition. [4](https://pubsonline.informs.org/doi/abs/10.1287/orsc.2023.18441?journalCode=orsc)

**Model generation matters.** Studies using 2022–2023 ChatGPT/Copilot, early-2023 GPT-4, early-2025 Claude/Cursor and late-2025 agentic tools are experiments on materially different technologies. METR's inability to cleanly reproduce its original design by 2026 because developers had integrated newer agents into their workflows is unusually direct evidence of this moving-target problem. [19](https://metr.org/blog/2026-02-24-uplift-update/)

## Material-study register

The table records the principal studies carrying the conclusions above. “Transferability” refers specifically to professional data, analytics and adjacent engineering teams—not to knowledge work in general.

### Core task and worker studies

<table header-row="true">
	<tr>
		<td>Study; status; direct URL</td>
		<td>Design, population, N</td>
		<td>Task / AI / comparator</td>
		<td>Outcome and duration</td>
		<td>Main result</td>
		<td>Key limitations</td>
		<td>Data-team transferability</td>
	</tr>
	<tr>
		<td>**Noy & Zhang, 2023, Science**, peer reviewed. DOI `https://doi.org/10.1126/science.adh2586` [1](https://www.science.org/doi/10.1126/science.adh2586)</td>
		<td>Preregistered randomized online experiment; college-educated professionals; **N=453**</td>
		<td>Occupation-specific professional writing; ChatGPT vs no AI</td>
		<td>Completion time and independent quality ratings; one-off tasks</td>
		<td>~40% faster; ~18% higher quality; weaker initial performers gain more</td>
		<td>Short tasks; no workflow integration, downstream correction or long-run learning</td>
		<td>**Moderate** for analyst writing/synthesis; low for whole job</td>
	</tr>
	<tr>
		<td>**Dell'Acqua et al., “Jagged Technological Frontier,” Organization Science, 2026**, peer reviewed. DOI `https://doi.org/10.1287/orsc.2025.21838` [11](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838)</td>
		<td>Preregistered field/lab RCT at BCG; **N=758 consultants**</td>
		<td>18 consulting tasks plus outside-frontier strategy problem; GPT-4 vs no AI, with an additional prompt-instruction condition</td>
		<td>Speed, completion count, quality/correctness; experimental session</td>
		<td>Inside frontier: ~25% faster, ~12% more tasks, &gt;30% quality gain; outside frontier: ~19 ppt lower correctness</td>
		<td>Short experimental work; early GPT-4; consulting tasks</td>
		<td>**High** for analysis/problem-solving mechanisms; medium for data jobs</td>
	</tr>
	<tr>
		<td>**Brynjolfsson, Li & Raymond, QJE, 2025**, peer reviewed. DOI `https://doi.org/10.1093/qje/qjae044` [2](https://academic.oup.com/qje/article/140/2/889/7990658)</td>
		<td>Staggered field deployment; **5,172 customer-support agents**, ~3m chats</td>
		<td>Real support conversations; organization-specific GenAI assistant vs pre-access/no access</td>
		<td>Resolutions/hour, quality, language, retention, learning; months</td>
		<td>~15% throughput overall; ~30% for low-skill/novice workers; faster ramp and some persistent gains during outages</td>
		<td>Single company/occupation; not data work</td>
		<td>**Medium**; excellent evidence on expertise and learning mechanisms</td>
	</tr>
	<tr>
		<td>**Merali, 2025, “Scaling Laws … Consulting, Data Analyst, and Management Tasks”**, arXiv preprint. `https://arxiv.org/abs/2512.21316` [6](https://arxiv.org/abs/2512.21316)</td>
		<td>Preregistered experiment; **&gt;500 consultants, data analysts and managers**</td>
		<td>Professional tasks using one of **13 LLMs**</td>
		<td>Task time; controlled sessions</td>
		<td>Each year of model progress associated with ~8% lower task time; gains larger on non-agentic analytical tasks than tool-using agentic work</td>
		<td>Preprint; pooled headline estimates; limited longitudinal validity</td>
		<td>**High/direct**, but still controlled task evidence</td>
	</tr>
	<tr>
		<td>**Merali, 2024, LLM-assisted translation**, arXiv preprint. `https://arxiv.org/abs/2409.02391` [20](https://arxiv.org/abs/2409.02391)</td>
		<td>Preregistered experiment; **300 professional translators; 1,800 tasks**</td>
		<td>Translation with 13 different LLMs/control</td>
		<td>Speed, grades, earnings/min</td>
		<td>10× model compute → 12.3% speed gain, +0.18 SD grades, +16.1% earnings/min; gains ~4× greater for lower-skilled</td>
		<td>Translation-specific; short tasks</td>
		<td>**Medium-low occupationally; high for heterogeneity mechanism**</td>
	</tr>

</table>

### Coding and engineering studies

<table header-row="true">
	<tr>
		<td>Study; status; direct URL</td>
		<td>Design, population, N</td>
		<td>Task / AI / comparator</td>
		<td>Outcome and duration</td>
		<td>Main result</td>
		<td>Key limitations</td>
		<td>Data-team transferability</td>
	</tr>
	<tr>
		<td>**Peng et al., 2023, GitHub Copilot**, arXiv preprint. DOI `https://doi.org/10.48550/arXiv.2302.06590` [15](https://doi.org/10.48550/ARXIV.2302.06590)</td>
		<td>Controlled randomized experiment; recruited software developers; **95 recruited**</td>
		<td>Build JavaScript HTTP server; GitHub Copilot vs no Copilot</td>
		<td>Completion time; one fixed exercise</td>
		<td>Treatment finished **55.8% faster**</td>
		<td>Narrow greenfield task; early Copilot; limited maintenance/quality horizon</td>
		<td>**Moderate–high** for coding subtasks; low for whole data-engineering job</td>
	</tr>
	<tr>
		<td>**Cui et al., 2025, “Effects of Generative AI on High-Skilled Work”**, Microsoft/SSRN preprint. `https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4945566` [14](https://www.microsoft.com/en-us/research/publication/the-effects-of-generative-ai-on-high-skilled-work-evidence-from-three-field-experiments-with-software-developers/)</td>
		<td>Three ordinary-business RCTs at Microsoft, Accenture, Fortune 100; **N=4,867 developers**</td>
		<td>AI code-completion assistant vs no access</td>
		<td>Completed tasks over company field-experiment periods</td>
		<td>Pooled **+26.08%** completed tasks; larger adoption/gains among less-experienced developers</td>
		<td>Individual studies noisy; completion counts not complete quality-adjusted business value</td>
		<td>**High** for coding portions of analytics/data engineering</td>
	</tr>
	<tr>
		<td>**Gambacorta et al., Journal of Financial Stability, 2026**, peer reviewed. DOI `https://doi.org/10.1016/j.jfs.2026.101543` [16](https://www.sciencedirect.com/science/article/pii/S1572308926000458)</td>
		<td>Quasi-experiment at Ant Group; random sample **N=1,219 programmers**, matched treated/control groups</td>
		<td>CodeFuse coding assistant</td>
		<td>Code output, completed tasks, commits; **12 weeks**</td>
		<td>Code output ~+55%; task completion substantially higher; significant gains concentrated among juniors</td>
		<td>One company/tool; output metrics imperfect proxies for long-run software value</td>
		<td>**High** for engineering subtasks; medium for data engineering as a role</td>
	</tr>
	<tr>
		<td>**Becker et al./METR, 2025**, RCT preprint. `https://arxiv.org/abs/2507.09089` [25](https://arxiv.org/abs/2507.09089)</td>
		<td>Within-developer randomized experiment; **16 experienced OSS developers, 246 tasks**, average ~5 years on repository</td>
		<td>Real issues in mature repos; primarily Cursor + Claude 3.5/3.7 vs no AI</td>
		<td>Implementation time and PR quality; Feb–Jun 2025; tasks average ~2h</td>
		<td>AI **increased completion time by 19%**; similar final quality; developers incorrectly perceived speedup</td>
		<td>Small/selected expert sample; mature OSS repositories; specific tool generation</td>
		<td>**Very high** for warning against extrapolation; especially relevant to expert codebase work</td>
	</tr>
	<tr>
		<td>**METR 2026 follow-up**, research update, not peer reviewed. `https://metr.org/blog/2026-02-24-uplift-update/` [19](https://metr.org/blog/2026-02-24-uplift-update/)</td>
		<td>Continued task randomization; **57 developers, 143 repos, 800+ tasks**</td>
		<td>Later-2025 agentic tools; AI allowed/disallowed</td>
		<td>Completion time; Aug 2025 onward</td>
		<td>Raw estimates move toward speedup, but authors deem signal unreliable due severe participant/task selection and concurrency measurement issues</td>
		<td>Selection directly compromises causal estimate</td>
		<td>**High methodological relevance**: AI changes task choice and time accounting</td>
	</tr>

</table>

### Team, learning, review and analytical studies

<table header-row="true">
	<tr>
		<td>Study; status; direct URL</td>
		<td>Design, population, N</td>
		<td>Task / AI / comparator</td>
		<td>Outcome and duration</td>
		<td>Main result</td>
		<td>Key limitations</td>
		<td>Data-team transferability</td>
	</tr>
	<tr>
		<td>**Dell'Acqua et al., “Cybernetic Teammate,” Organization Science, 2026**, peer reviewed. DOI `https://doi.org/10.1287/orsc.2025.20702` [3](https://pubsonline.informs.org/doi/10.1287/orsc.2025.20702)</td>
		<td>Preregistered 2×2 field experiment; P&G R&D/commercial employees; **826 recruited, 791 analyzed**</td>
		<td>Individual/team × AI/no AI product innovation</td>
		<td>Idea quality, top-decile output, functional content, selection; **one day**</td>
		<td>Individual+AI ≈ human team quality; team+AI +0.39 SD vs unaided individual; functional boundaries blur; AI conditions worse at identifying own best idea</td>
		<td>Flash teams, single firm, one innovation task, no persistent staffing outcome</td>
		<td>**Medium-high** for team mechanisms; low for headcount inference</td>
	</tr>
	<tr>
		<td>**Bastani et al., PNAS, 2025**, peer reviewed. DOI `https://doi.org/10.1073/pnas.2422633122` [23](https://www.pnas.org/doi/10.1073/pnas.2422633122)</td>
		<td>Randomized field experiment; nearly 1,000 high-school mathematics students; analytic N varies</td>
		<td>GPT Base, guarded GPT Tutor, control</td>
		<td>Practice performance then unaided exam</td>
		<td>Unrestricted GPT improves assisted performance but unaided exam performance **17% below control**; tutor guardrails avoid much of harm</td>
		<td>Education, not workplace; different motivation/task structure</td>
		<td>**Low occupational transfer**, **high conceptual relevance** to assisted performance vs learning</td>
	</tr>
	<tr>
		<td>**Klingbeil, Grützner & Schreck, Computers in Human Behavior, 2024/25**, peer reviewed. DOI `https://doi.org/10.1016/j.chb.2024.108352` [24](https://doi.org/10.1016/J.CHB.2024.108352)</td>
		<td>Randomized incentivized behavioral experiment; **N=319**</td>
		<td>Trust-game decisions; AI vs expert advice vs control</td>
		<td>Reliance, contextual errors, monetary payoffs; four rounds</td>
		<td>Participants overrelied on AI even against contextual information; AI-treatment payoffs ~22% below control in relevant comparison</td>
		<td>Abstract decision task; not professional analytics</td>
		<td>**Medium for automation-bias mechanism**</td>
	</tr>
	<tr>
		<td>**Lee et al., CHI 2025**, peer-reviewed conference. DOI `https://doi.org/10.1145/3706598.3713778` [8](https://www.microsoft.com/en-us/research/wp-content/uploads/2025/01/lee_2025_ai_critical_thinking_survey.pdf)</td>
		<td>Survey/critical-incident study; **319 knowledge workers**</td>
		<td>Real recalled GenAI work tasks</td>
		<td>Self-reported critical thinking and cognitive effort</td>
		<td>GenAI changes where/when workers report applying critical thinking; confidence in AI is associated with reduced effort</td>
		<td>Self-report, selection, no objective skill-loss measure</td>
		<td>**Medium for workflow hypotheses; low for deskilling causality**</td>
	</tr>
	<tr>
		<td>**Wang et al., Data Formulator 2**, HCI/system study, 2024 preprint. [7](https://arxiv.org/abs/2408.16119)</td>
		<td>Exploratory user study; **N=8**, including professional data-science experience</td>
		<td>AI-assisted data transformations and visualization</td>
		<td>Workflow, verification and iteration behavior</td>
		<td>Users develop active verification/error-correction strategies</td>
		<td>Tiny N; no clean no-AI productivity counterfactual</td>
		<td>**High/direct process relevance; low causal strength**</td>
	</tr>
	<tr>
		<td>**Xue, Zhang & Zhu, “Generative AI for Analysts,” 2025/2026**, SSRN preprint, revised Sept. 2026. DOI `https://doi.org/10.2139/ssrn.5821142` [13](https://papers.ssrn.com/sol3/Delivery.cfm/5821142.pdf?abstractid=5821142&mirid=1)</td>
		<td>Quasi-experimental financial-analyst panel around FactSet GenAI integration; large report panel, N not stated in accessible abstract</td>
		<td>Professional financial-analysis/report production</td>
		<td>Report richness/timeliness, forecasting</td>
		<td>More sources, topics and analytical methods; timelier output; forecast accuracy deteriorates under higher information-processing load</td>
		<td>Preprint; financial analysts; adoption/identification assumptions</td>
		<td>**Medium-high** for analytical work and cognitive bottlenecks</td>
	</tr>

</table>

### Labor-market and organization-level studies

<table header-row="true">
	<tr>
		<td>Study; status; direct URL</td>
		<td>Design, population, N</td>
		<td>Exposure/comparator</td>
		<td>Outcome and duration</td>
		<td>Main result</td>
		<td>Key limitations</td>
		<td>Data-team transferability</td>
	</tr>
	<tr>
		<td>**Hui, Reshef & Zhou, Organization Science, 2024**, peer reviewed. DOI `https://doi.org/10.1287/orsc.2023.18441` [4](https://pubsonline.informs.org/doi/abs/10.1287/orsc.2023.18441?journalCode=orsc)</td>
		<td>Difference-in-differences/event-study style online labor-market analysis; large freelancer panel</td>
		<td>Occupations differentially affected by ChatGPT, DALL-E, Midjourney</td>
		<td>Contracts/employment and earnings after releases</td>
		<td>Highly affected freelance occupations experience reduced employment and earnings; high prior quality does not clearly protect workers</td>
		<td>Online gig market; concurrent shocks; exposure ≠ individual AI adoption</td>
		<td>**Medium-low** for salaried teams; important substitution evidence</td>
	</tr>
	<tr>
		<td>**Humlum & Vestergaard, NBER/RFBerlin, 2025–2026**, working paper. `https://www.nber.org/papers/w33777` [26](https://www.nber.org/papers/w33777)</td>
		<td>Two adoption surveys linked to administrative records; ~**25,000 workers, 7,000 workplaces, 11 exposed occupations**</td>
		<td>Actual chatbot adoption/employer initiatives</td>
		<td>Adoption, time savings, tasks, earnings, hours; through roughly two years post-ChatGPT</td>
		<td>Rapid adoption/new tasks/modest reported savings; **no detectable average earnings or recorded-hour impact**</td>
		<td>Denmark; non-random individual adoption despite quasi-experimental analyses; early horizon</td>
		<td>**Medium-high** for organization/adoption boundary</td>
	</tr>
	<tr>
		<td>**Brynjolfsson, Chandar & Chen, “Canaries in the Coal Mine?”, Aug. 2026**, working paper. Direct PDF: `https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf` [5](https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf)</td>
		<td>High-frequency ADP payroll panel; **millions of U.S. workers** through June 2026</td>
		<td>Occupation-level AI exposure and substitutive/complementary usage</td>
		<td>Employment, hiring, separations, compensation</td>
		<td>No economy-wide displacement; 22–25-year-old employment in exposed occupations 19% below kept-pace comparator; driven mainly by hiring; authors explicitly call evidence descriptive</td>
		<td>Noncausal exposure design, pre-trends, education sensitivity, ADP representativeness</td>
		<td>**Medium** for early-career risk; not specific to data teams</td>
	</tr>

</table>

## What the evidence does not support, and the research frontier

### What current evidence does not support

As of October 3, 2026, the evidence reviewed here does **not** legitimately support the following strong statements.

It does not support a general conversion from task speed into headcount. A 25% faster consultant or coder task is not evidence for a 20% smaller organization. The strongest organization-level evidence shows workers reallocate time, adoption creates new tasks and aggregate hours may remain unchanged. [22](https://bfi.uchicago.edu/wp-content/uploads/2025/04/BFI_WP_2025-56-1.pdf)

It does not support saying that AI generally improves professional quality. Quality rises strongly inside some capability frontiers and falls outside them. [11](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838)

It does not support saying that experts no longer matter because AI compresses performance distributions. Experts can have smaller relative gains while retaining substantial absolute advantage; their contextual knowledge can also determine whether AI is useful at all. [2](https://academic.oup.com/qje/article/140/2/889/7990658)

It does not support saying that experts inevitably benefit less. BCG's higher performers still gained materially, and the relevant interaction is task × expertise × model capability rather than expertise alone. [11](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838)

It does not support saying that AI necessarily teaches juniors. Customer support supplies positive evidence, while controlled learning experiments show that the opposite can happen when assistance replaces cognitive effort. [2](https://academic.oup.com/qje/article/140/2/889/7990658)

It does not support a strong workplace **deskilling** claim. We have evidence of cognitive offloading and one important educational learning penalty, but not multi-year evidence of professional skill atrophy among data workers. [8](https://www.microsoft.com/en-us/research/wp-content/uploads/2025/01/lee_2025_ai_critical_thinking_survey.pdf)

It does not support saying that junior roles are disappearing *because of AI*. There is a serious descriptive hiring signal in U.S. exposed occupations, but the best current paper explicitly stops short of causal attribution, while Danish evidence produces much smaller aggregate labor-market effects. [5](https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf)

It does not support “SQL is becoming irrelevant.” The direct empirical literature needed to establish that proposition simply does not exist at sufficient quality.

It does not support “technical skill matters less than business skill.” Existing research shows that context and judgment remain important; it does not estimate that technical human capital has fallen in relative causal value.

It does not support saying analysts and analytics engineers are empirically merging into a common role. Cross-functional output can converge without occupations converging. [3](https://pubsonline.informs.org/doi/10.1287/orsc.2025.20702)

It does not support “data engineers are becoming architects rather than coders.” Coding assistance is well established; occupational task substitution is not.

It does not support saying senior employees are becoming review bottlenecks. Review burden is demonstrably real in some workflows, but the organizational queueing claim has not been measured. [17](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/)

It does not support saying AI-native teams dramatically outperform traditional teams. The rare randomized team experiment provides a much subtler result. [3](https://pubsonline.informs.org/doi/10.1287/orsc.2025.20702)

And it does not support assuming that self-reported productivity is a good approximation to causal productivity. METR is a direct warning: developers predicted a large speedup, objectively became slower, and still believed afterwards that AI had sped them up. [17](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/)

### Most important longitudinal research gaps

The decisive missing experiment for data teams is not another one-hour benchmark. It is a **six-to-twenty-four-month randomized or phased deployment** in real analytics and engineering organizations that measures the complete production system.

Such work needs to observe the same people both with and without assistance at intervals; measure query/code/report quality plus downstream defects; track stakeholder value rather than artifact production alone; capture time spent prompting, reviewing and correcting; identify task creation and task disappearance; and test independent competence periodically with AI removed.

The highest-priority unresolved questions are these.

**Whole-worker productivity.** When analysts save thirty minutes writing SQL, where does the thirty minutes actually go? Humlum and Vestergaard provide early evidence of reallocation, but not detailed data-team production functions. [27](https://static1.squarespace.com/static/5d35e72fcff15f0001b48fc2/t/6878c9ba2c66f90282b47e0b/1752746431549/chatbots_july25.pdf)

**Quality-adjusted end-to-end output.** Does faster generation reduce cycle time from business question to trusted production decision after validation, review, dashboard changes, stakeholder clarification and remediation?

**Learning trajectories.** Do AI-assisted junior analysts reach senior-level independent competence faster, slower or differently? Customer support and educational experiments show opposite mechanisms are both possible. [2](https://academic.oup.com/qje/article/140/2/889/7990658)

**Loss of practice.** Which skills decay when rarely exercised, at what rate, and does improved higher-level problem solving compensate? Current knowledge-worker surveys cannot answer this. [8](https://www.microsoft.com/en-us/research/wp-content/uploads/2025/01/lee_2025_ai_critical_thinking_survey.pdf)

**Review economics.** As generated output volume rises, does reviewer workload rise linearly, sublinearly or become selectively automatable? P&G and METR show why this matters without resolving it. [3](https://pubsonline.informs.org/doi/10.1287/orsc.2025.20702)

**Team topology.** Does AI reduce the value of specialist handoffs, increase span of control, favor smaller cross-functional teams or merely expand the amount of work teams undertake? P&G provides an initial mechanism but not longitudinal organizational evidence. [3](https://pubsonline.informs.org/doi/10.1287/orsc.2025.20702)

**Apprenticeship.** If firms automate routine tasks traditionally assigned to juniors, what replaces their learning function? The U.S. hiring pattern makes this urgent, but no study yet links GenAI-driven removal of junior tasks causally to later senior-skill supply. [5](https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf)

**Demand response.** When analysis or engineering becomes cheaper, how much more analysis or software do organizations demand? Without this elasticity, productivity cannot be translated into employment effects. [5](https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf)

**Model-generation effects.** The treatment called “AI” is moving too rapidly for a coefficient estimated using GPT-3.5, early GPT-4 or 2023 Copilot to be treated as a stable technology parameter. METR's 2026 measurement difficulties are an unusually vivid example. [19](https://metr.org/blog/2026-02-24-uplift-update/)

**Data-role-specific exposure.** We need separate studies of BI analysts, product analysts, data scientists, analytics engineers, data engineers and ML engineers rather than generic “knowledge worker” or “software developer” categories.

### Research frontier

The frontier is moving from **“Does access to an LLM make a fixed task faster?”** toward harder questions.

One frontier is **endogenous task choice**: workers start selecting work because AI makes previously expensive tasks cheap. METR's late-2025 experiment already encountered this experimentally. [19](https://metr.org/blog/2026-02-24-uplift-update/)

Another is **concurrent agentic work**: a person may supervise several machine processes, so human active time, wall-clock time and produced value diverge. Conventional productivity experiments are not designed for this. [19](https://metr.org/blog/2026-02-24-uplift-update/)

A third is **verification as a scarce complement**. BCG, P&G, FactSet and human-overreliance experiments independently point toward a regime in which producing candidate work becomes cheaper faster than evaluating it. [12](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838)

A fourth is **AI-mediated knowledge diffusion**. Customer support shows that a model can encode practices associated with high performers and spread them to novices, potentially transforming organizational learning. Whether the same mechanism works for tacit data-modeling judgment, architecture or causal analysis is unknown. [2](https://academic.oup.com/qje/article/140/2/889/7990658)

A fifth is **task-versus-role convergence**. P&G provides genuine evidence that AI can narrow differences between specialists' outputs. The unanswered question is whether that ultimately creates broader generalist jobs, changes team interfaces, or simply lets specialists collaborate more effectively. [3](https://pubsonline.informs.org/doi/10.1287/orsc.2025.20702)

A sixth is the **entry-level paradox**. The same technology can make novices much more productive while potentially reducing employer demand for some novice tasks. Customer support and coding provide the former evidence; U.S. payroll data raise concern about the latter. Whether demand expansion, faster learning or new responsibilities offset task substitution is now one of the central empirical questions. [2](https://academic.oup.com/qje/article/140/2/889/7990658)

The most defensible synthesis as of the snapshot date is therefore not that GenAI is “automating knowledge workers,” nor that it is merely “augmenting them.” Both verbs are too coarse.

**GenAI is selectively automating, augmenting and creating tasks inside jobs. It changes the location of bottlenecks and sometimes the distribution of performance across workers. We have good causal evidence for many of those task-level effects. We have much less evidence about the stable jobs, teams and organizational structures that result.**

For data leaders specifically, the evidence currently justifies treating AI as a potentially large **task-production technology** and an increasingly important **workflow-design variable**. It does **not yet justify treating smaller teams, eliminated junior layers, merged analyst/engineering roles, or architecture-only data engineers as empirical facts.** Those are hypotheses awaiting the level of evidence that the task-productivity question has finally begun to receive.