Productivity

Productivity describes the relationship between output and the inputs used to produce it. Labor productivity is commonly measured as output per hour worked, while multifactor productivity relates output to a bundle of inputs such as labor and capital.^[1][2] In the context of artificial intelligence, the term can refer to at least three different outcomes: completing a task faster or better, increasing an organization's output relative to its costs, or increasing productivity across an economy. Evidence at one level does not establish an effect at another.

Modern generative AI tools can draft and transform text, retrieve and synthesize information, assist with code, transcribe meetings, and carry out sequences of tool calls. Many use a large language model as a core component. Controlled and field studies have found substantial benefits for some well-specified tasks, including professional writing and customer support.^[3][4] Other studies have found smaller effects, no detected change in measured output, or slower completion in settings where the worker already knew a complex codebase well.^[5][8][10] The size and even the direction of an estimate therefore depend on the task, worker, system, outcome measure, and study period.

This article covers AI-related productivity as an empirical and measurement question. It does not rank products, repeat vendor return-on-investment claims, or treat projected economic value as observed productivity.

What productivity means

Productivity is a ratio, not simply an activity count. A worker may send more messages without producing more useful work, and a team may finish a draft faster while creating additional review or correction costs. The denominator also matters. An intervention that reduces staff time but adds model fees, integration work, supervision, or security controls may still be worthwhile, but its full productivity effect cannot be inferred from time saved alone.

The main levels of analysis are:

Scroll sideways for more →

The US Bureau of Labor Statistics defines major-sector labor productivity as sectoral output divided by hours worked. It defines multifactor productivity as output divided by a combined input measure.^[1] The OECD likewise distinguishes single-factor measures from multifactor measures and explains that a productivity statistic depends on how output and inputs are defined.^[2] These definitions are useful guardrails for interpreting AI studies. A faster task is evidence about task efficiency. It becomes evidence about firm or economy-wide productivity only when output, quality, relevant inputs, and aggregation are also addressed.

How AI can change a workflow

AI productivity software is better understood by the part of a workflow it changes than by a single product category. A system may reduce the cost of producing a first draft, help a worker find internal information, recommend a next action, or perform an action through connected tools. These functions can coexist in one application.

Scroll sideways for more →

Provider documentation establishes what a product is designed to do, not how much productivity it creates in a particular workplace. Product capabilities also change more quickly than peer-reviewed evidence. For those reasons, durable evaluation should describe the tested product version, enabled features, connected data, user population, and date.

Some systems use retrieval-augmented generation to ground an answer in selected documents, and users may apply prompt engineering to specify the task and desired format. Neither technique guarantees correctness. They change the information available to the system and the way the request is expressed, while the final output can still require verification.

Evidence from controlled and field studies

The strongest evidence uses a credible comparison group and measures an outcome that matters for the assigned task. Even then, each estimate is local to its study design. The table below reports the population, intervention, measured result, and an important boundary rather than presenting the figures as a single universal productivity range.

Scroll sideways for more →

The writing and customer-support results show that AI can improve both speed and evaluated quality when the task and system are well matched.^[3][4] They also show why average effects can hide distributional differences. In the customer-support study, the assistant appears to have helped less-experienced workers apply practices present in the organization's data, while highly experienced workers had little room for the same type of gain.^[4] This pattern is important but is not universal: effects can depend on whether expertise helps a user direct and check the system or instead makes the system's suggestions redundant.

The consulting experiment provides a direct warning against defining productivity only as rapid completion. Workers did better on tasks within the model's tested capability range but worse on the deliberately selected outside-frontier task.^[5] A plausible answer accepted without adequate checking can increase apparent throughput while reducing correct output. For high-consequence work, accuracy and downstream correction are part of the outcome, not optional adjustments.

Software-development evidence illustrates the importance of context. A fixed greenfield coding task yielded a large speedup,^[6] and pooled field experiments found more completed tasks on company systems.^[7] The METR trial instead studied experienced maintainers doing real work in repositories they had used for years and found a slowdown with early-2025 tools.^[8] In February 2026, METR reported that its later study could not provide a reliable causal estimate because developers and tasks expected to benefit most were increasingly selected out, the pay rate changed, and concurrent agents made time measurement unreliable. Its raw later estimates were therefore described as weak evidence rather than a definitive update.^[9]

These studies are not contradictions that can be resolved by averaging their headline percentages. They estimate different interventions for different workers and outcomes. Together they support a narrower conclusion: AI can produce material task-level gains, but performance is heterogeneous and empirical results should remain attached to their study conditions.

Why estimates differ

Several design choices can change a productivity estimate:

• Task fit. A model may be reliable on drafting, classification, or familiar transformations and unreliable on a nearby task requiring missing context or a subtle decision. The "jagged frontier" describes this uneven boundary.^[5]

• Worker experience. Novices may benefit from examples and suggestions that experienced workers already know. In other settings, domain expertise may be necessary to supply context, reject errors, and integrate output.

• Baseline and comparison. A comparison against unaided work differs from a comparison against existing search, templates, automation, or specialist software.

• Tool configuration. Model version, context window, connected data, permissions, latency, interface, and approval rules can all affect the result.

• Adoption and compliance. Giving access is not the same as sustained use. Treatment effects based on access, effects among users, and self-selected observational comparisons answer different questions.

• Outcome definition. Time, activity counts, accepted cases, code changes, sales, quality scores, errors, and value added are not substitutes. A faster draft is not necessarily a better final deliverable.

• Time horizon. Short experiments capture immediate assistance and learning costs. Longer deployments may capture adaptation, new workflows, process redesign, and changes in work allocation.

• Spillovers. An individual may save time while shifting review, coordination, or error-correction work to colleagues. Team-level measurement is needed to detect that transfer.

The OECD's 2025 review of experimental evidence similarly concludes that results depend on task and user characteristics and that external validity and longer-run effects remain limited. It emphasizes the continuing role of human expertise and oversight.^[12] An International Labour Organization review published in 2026 describes the emerging gains as real but uneven and often unverified. It also notes that worker-reported time savings have not yet consistently appeared as measured output, earnings, or employment effects.^[13]

From task gains to organizational productivity

An organization realizes a productivity benefit only when a useful task effect survives the surrounding workflow. Deployment can require data preparation, identity and access controls, procurement, model evaluation, user training, integration, monitoring, and exception handling. These inputs should be counted. So should review time, rework, incidents, and the cost of tasks that the system initiates but does not complete correctly.

The six-month multi-firm experiment by Dillon and colleagues is informative because it moved beyond a single benchmark task. It found less email time among users without detecting a corresponding change in the number or composition of activities visible in the applications studied.^[10] That outcome may still benefit workers, especially when time outside regular hours falls, but it is not equivalent to an increase in measured organizational output.

Evidence from Denmark also separates work reorganization from aggregate labor-market outcomes. Humlum and Vestergaard linked adoption surveys in highly exposed occupations to administrative records. Their March 2026 revision reports widespread new tasks involving content generation, AI oversight, and AI integration, alongside statistically precise null effects on earnings and recorded hours at worker and workplace levels. Their estimates rule out effects larger than 2 percent on those outcomes two years after ChatGPT's launch.^[11] The finding applies to the measured Danish setting and horizon; it does not prove that future or different deployments have no effect.

This delay is consistent with the productivity J-curve hypothesis for general-purpose technologies. Brynjolfsson, Rock, and Syverson argue that complementary intangible investment, including redesigned processes and organizational capabilities, may be costly before its benefits appear in conventional productivity statistics.^[17] Applied to AI, the hypothesis suggests a reason task demonstrations and aggregate statistics can diverge. It is a framework for interpretation, not evidence that every AI investment will eventually pay off.

Practical organizational evaluation should therefore begin with a defined workflow rather than a broad mandate to "use AI." Teams can map the current process, identify the decision or transformation being assisted, specify who remains accountable, and test whether the intervention changes a valued outcome after all material inputs are counted.

Economy-wide productivity

Economy-wide estimates are more uncertain than task-level experiments because they require assumptions about exposure, adoption, task savings, prices, investment, new tasks, substitution between inputs, and spillovers. They should be labeled as models or scenarios.

Daron Acemoglu uses a task-based model and existing estimates of AI exposure and task-level cost savings. His 2024 analysis derives an increase of no more than 0.66 percent in total factor productivity over ten years under its baseline assumptions and argues that even this may be high when future affected tasks are harder for AI systems to learn.^[14] This is a model-based calculation, not an observed ten-year result.

An IMF working paper applies related methods to 31 European countries and includes alternative assumptions about adoption, exposure, regulation, and task complementarity. Its preferred scenario produces about a 1.1 percent cumulative productivity gain over the medium term for Europe, with meaningful variation across countries and scenarios.^[15] That figure is also a conditional model result rather than a measurement of realized productivity.

An OECD analysis asks whether generative AI has characteristics of a general-purpose technology: broad applicability, continuing improvement, and the capacity to generate complementary innovation. It concludes that generative AI shows potential along these dimensions while stressing uncertainty and the importance of complementary investment, skills, competition, and diffusion.^[16] Calling a technology general purpose does not specify the timing or magnitude of a productivity effect.

The responsible comparison is therefore not between a 40 percent writing-task time reduction and a roughly 1 percent macroeconomic scenario. The figures refer to different units, denominators, time horizons, and causal mechanisms. Task effects enter a macro model only for the share of work that is exposed, technically feasible, adopted, and translated into valuable output after costs and equilibrium responses.

Measuring an AI productivity intervention

A useful evaluation protocol can be stated before deployment:

1. Define the unit. Specify whether the target is a document, case, software task, worker-week, team process, business unit, or firm.

2. Record the baseline. Measure the existing workflow, tools, completion time, quality, error rate, queue, and cost before adding AI.

3. Create a comparison. When practical, randomize access or rollout timing. If randomization is not practical, document the comparison group and likely sources of selection bias.

4. Pin the intervention. Record the model and product version, enabled features, prompts or instructions, connected sources, permissions, and evaluation dates.

5. Measure adoption separately. Track access, actual use, frequency, abandonment, and use of substitute tools. Report whether an estimate is based on assignment, access, or use.

6. Choose an outcome with value. Measure accepted or correct output, not only activity. Pair speed or throughput with quality, defects, customer outcomes, or another domain-appropriate measure.

7. Count relevant inputs. Include staff time, subscriptions or inference, integration, training, supervision, security, review, rework, and incident response.

8. Look for displacement. Test whether work moved to reviewers, colleagues, customers, or a later stage of the process.

9. Measure distribution. Report results by role or prior experience when the sample supports it. An average can conceal gains for one group and losses for another.

10. Report uncertainty and duration. Include sample size, confidence intervals or standard errors, attrition, missing data, and study dates. Repeat measurement after users and workflows have adapted.

This protocol distinguishes three useful quantities. Gross task savings capture the direct change in time or performance on the assisted task. Net workflow savings subtract added review, correction, coordination, and operating costs. Organizational productivity relates the valued output of the whole process to all relevant inputs. Reporting all three where possible prevents a local speedup from being misrepresented as a firm-wide result.

Self-reports can reveal user experience and help identify mechanisms, but they are weak substitutes for timed or administrative outcomes. In the METR study, developers believed AI had sped them up even though the randomized comparison found longer completion times.^[8] Telemetry also has limits: it can show time in an application without observing the quality or value of the work. A credible evaluation uses the measurement method that matches the claim.

Risks and tradeoffs

Productivity claims should include failure costs. NIST's Generative AI Profile defines confabulation as confidently presented erroneous or false content and treats it as one of several risks that organizations should measure and manage. The profile recommends mapping use contexts, defining human oversight, evaluating systems in conditions relevant to deployment, and monitoring behavior after deployment.^[18] The AI Wiki article on hallucination covers the related terminology and technical problem.

Important tradeoffs include:

• Verification and automation bias. Fluent output can reduce the chance that a user checks an error. High-consequence decisions need evidence, review responsibility, and an escalation path.

• Privacy and access. Productivity tools may process email, documents, meetings, source code, or internal records. Access should follow the underlying user's authorization, and the deployment should specify retention, logging, and data-handling rules.

• Security and agent actions. A tool that can send messages, modify code, or change records creates risks beyond incorrect text. Least privilege, sandboxing, confirmation for consequential actions, and auditable logs limit the impact of a failure.

• Work intensity and job quality. Time savings may become reduced working time, additional output, or higher workload expectations. The ILO review identifies coordination, autonomy, inequality, entry-level opportunities, and job quality as outcomes that should be measured alongside throughput.^[13]

• Unequal effects. Some experiments find larger benefits for less-experienced workers,^[3][4][7] but the pattern depends on the task and organization. It should not be assumed for every occupation or used to infer long-run wages or employment.

The evidence supports neither a blanket claim that AI always raises productivity nor a blanket claim that it cannot do so. The most defensible conclusion is conditional: a defined AI intervention can raise productivity when it improves a valued outcome more than it increases the complete set of inputs and risks. Establishing that condition requires measurement at the level of the claim.

See also

• Erik Brynjolfsson
• Productivity ChatGPT Plugins
• Productivity Custom GPTs

References

1. ^US Bureau of Labor Statistics. "Major Sector Productivity and Costs: Calculation." Updated January 29, 2026. bls.gov/...calculation
2. ^OECD. *Measuring Productivity: OECD Manual*. 2001. oecd.org/...9789264194519-en.pdf
3. ^Noy, Shakked, and Whitney Zhang. "Experimental evidence on the productivity effects of generative artificial intelligence." *Science* 381, no. 6654 (2023): 187-192. doi.org/...science.adh2586
4. ^Brynjolfsson, Erik, Danielle Li, and Lindsey R. Raymond. "Generative AI at Work." *Quarterly Journal of Economics* 140, no. 2 (2025): 889-942. academic.oup.com/...7990658
5. ^Dell'Acqua, Fabrizio, et al. "Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality." *Organization Science* (2026). doi.org/...orsc.2025.21838
6. ^Peng, Sida, Eirini Kalliamvakou, Peter Cihon, and Mert Demirer. "The Impact of AI on Developer Productivity: Evidence from GitHub Copilot." arXiv:2302.06590 (2023). arxiv.org/...2302.06590
7. ^Cui, Zheyuan Kevin, et al. "The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers." *Management Science* (2026). doi.org/...mnsc.2025.00535
8. ^Becker, Joel, Nate Rush, Beth Barnes, and David Rein. "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity." METR, July 2025. metr.org/...AI_Experienced_OS_Devs_Study-paper.pdf
9. ^METR. "We are Changing our Developer Productivity Experiment Design." February 24, 2026. metr.org/...2026-02-24-uplift-update
10. ^Dillon, Eleanor Wiske, Sonia Jaffe, Nicole Immorlica, and Christopher T. Stanton. "Shifting Work Patterns with Generative AI." arXiv:2504.11436, revised November 2025. arxiv.org/...2504.11436
11. ^Humlum, Anders, and Emilie Vestergaard. "Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI." NBER Working Paper 33777, revised March 2026. nber.org/...w33777
12. ^Calvino, Flavio, Lucia Reggi, and Filippo Squicciarini. "The Effects of Generative AI on Productivity, Innovation and Entrepreneurship." OECD Artificial Intelligence Papers, no. 39, 2025. doi.org/...b21df222-en
13. ^International Labour Organization. "The impact of GenAI on jobs, productivity and work organization: a review of the empirical evidence." June 1, 2026. ilo.org/...-and-work-organization-review-empirical
14. ^Acemoglu, Daron. "The Simple Macroeconomics of AI." *Economic Policy* 40, no. 121 (2025): 13-58. doi.org/...eiae042
15. ^Misch, Florian, Ben Park, Carlo Pizzinelli, and Galen Sher. "Artificial Intelligence and Productivity in Europe." IMF Working Paper 2025/067. 2025. imf.org/...AI-and-Productivity-in-Europe-565924
16. ^Filippucci, Francesco, Peter Gal, and Matthias Schief. "Is Generative AI a General-Purpose Technology? Implications for Productivity and Policy." OECD Artificial Intelligence Papers, no. 40, 2025. doi.org/...704e2d12-en
17. ^Brynjolfsson, Erik, Daniel Rock, and Chad Syverson. "The Productivity J-Curve: How Intangibles Complement General Purpose Technologies." *American Economic Journal: Macroeconomics* 13, no. 1 (2021): 333-372. doi.org/...mac.20180386
18. ^National Institute of Standards and Technology. *Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile*. NIST AI 600-1. 2024. doi.org/...NIST.AI.600-1
19. ^Microsoft Support. "What you can do with Microsoft Copilot." Accessed July 28, 2026. support.microsoft.com/...do-with-microsoft-copilot
20. ^Google Workspace Learning Center. "Collaborate with Gemini in Google Docs." Accessed July 28, 2026. support.google.com/...15123226
21. ^Notion. "Notion AI." Accessed July 28, 2026. notion.com/...notion-ai
22. ^GitHub Docs. "GitHub Copilot documentation." Accessed July 28, 2026. docs.github.com/...copilot
23. ^Cursor. "Agent overview." Accessed July 28, 2026. docs.cursor.com/agent
24. ^Otter.ai. "AI Meeting Agent." Accessed July 28, 2026. get.otter.ai/ai-meeting-agent