Why AI Productivity Gains Do Not Always Reach Firm Level

Why can task-level AI gains fail to show up at firm level?

A task can get faster without raising firm productivity. The tool may help with only a slice of work; review, coordination, process constraints and measurement stand between time saved and valuable additional output. Evidence is mixed: task studies and some firm studies report gains, but neither automatically generalizes. The question is where a gain goes from task to firm.

Trace that movement as a value-conversion chain: a tool produces a local result; people use it repeatedly widely; the full workflow improves after review and handoffs; released capacity is applied to valuable output; and a firm measure records that output. Each step is a condition, not a promise. A tool may be capable without frequent use. One step may speed up while the next queue stays unchanged. A process gain may not appear fully in the metric.

The Federal Reserve’s “The AI Buildout and the Economy” explains that task experiments measure outcomes such as task speed, while adjustment costs or bottlenecks can erode gains before they reach firm or aggregate productivity. The International Labour Organization’s “The Aggregation Paradox of AI” describes mixed firm evidence and uneven adoption; its brief says gains are concentrated in larger, digitally advanced enterprises, while many firms report little measurable effect beyond pilots. This frames a conversion problem, not proof that gains are imaginary or universally absent.

Treat a faster task as a reason to test a recurring workflow through final review. Check use, net effort, handoffs, quality and what happened to released time before making a firm-wide claim. Productivity evidence alone cannot establish whether jobs will be added or removed.

Sources: The AI Buildout and the Economy: Publicly Available Data to Assess AI’s Impact; The Aggregation Paradox of AI: Why do micro-economic productivity gains from AI disappear at scale

What does an individual productivity gain actually measure?

An individual productivity gain measures a specified result for a specified unit and period. It does not by itself show that the firm produced more value per labor hour. In “Generative AI at Work,” a staggered rollout among 5,172 customer-support agents was associated with an average 15% increase in issues resolved per hour. This is a worker-level result in one support operation, not a 15% increase in company output or a prediction for every workflow. The study reports substantial variation across workers in its studied setting. Firm labor productivity has a wider boundary: output relative to labor input across the firm’s activities. The Federal Reserve note “The AI Buildout and the Economy: Publicly Available Data to Assess AI’s Impact” distinguishes task outcomes such as speed from firm or sector measures that aggregate many activities. It notes that adjustment costs or bottlenecks can erode an upstream gain before broader output changes. Total factor productivity is broader again, relating output to multiple inputs rather than labor alone; it is not another name for issues resolved per hour. The task percentage therefore cannot simply be multiplied across company hours. A wider estimate would require knowing the task’s share of labor, how often the tool is used, whether quality holds, how the task contributes to final output, and what other inputs or costs change. Without those links, arithmetic conceals assumptions. A field experiment shows why a worker outcome and work-system change are separate questions. “Shifting Work Patterns with Generative AI” randomized access among 7,137 knowledge workers in 66 firms for six months. In the second half, the 80% of treated workers who used the tool spent two fewer hours on email weekly and reduced work outside regular hours. Researchers detected no change in task quantity or composition from individual-level access during the experiment. The email-time result did not establish task redesign in that period. To compare a task headline with a firm claim, check the unit, counted output, denominator and observation window. Ask whether the result is per task, per worker or across the firm, and whether it counts time saved or accepted output. This keeps a local gain within the boundary the study measured.

Sources: Generative AI at Work; Shifting Work Patterns with Generative AI; The AI Buildout and the Economy: Publicly Available Data to Assess AI’s Impact

Why does adoption intensity matter more than a yes-or-no label?

Firm adoption is not a binary bridge from no effect to full-scale productivity. A tool may be available or reported as adopted while reaching only a narrow set of workers, tasks, or hours. The relevant question is how often it is used in eligible work, and how much of the firm’s labor that work represents. A gain on one task can have little effect on an average that also includes many untouched tasks.

The nationally representative U.S. surveys reported in “The Rapid Adoption of Generative AI” make the distinction concrete. As of late 2024, 27% of employed respondents said they had used generative AI for work at least once in the prior week, while 10% said they used it every workday. The study estimated that 1% to 7% of all work hours were assisted, and respondents reported time savings equivalent to 1.4% of total work hours. These are different measures: occasional use is not daily use, assisted hours are not a count of all output improved, and reported time saved is not audited firm productivity. They describe U.S. respondents at that time, not a current 2026 rate or every employer.

Firm-level patterns also caution against treating an adoption checkbox as a uniform dose. The OECD working paper “A Portrait of AI Adopters Across Countries” applied harmonized statistical methods to official firm surveys in 11 countries. It found AI use more prevalent in ICT and professional services and more widespread among large firms; AI users tended to be more productive, especially the largest firms. The report also found that ICT skills, high-speed digital infrastructure, and use of other digital technologies were related to AI use and appeared relevant to adopters’ productivity advantage. This is an association in cross-country firm data, not proof that adoption alone caused higher productivity, and its 2023 evidence predates the generative-AI wave.

A 2026 Federal Reserve note reports that Census survey adoption rates have a positive association with firm size. It warns that headline uptake figures do not show usage intensity, which surveys suggest remains shallow even where reported adoption is broad. These sources suggest a conversion problem: licenses, policy statements, or reported use tell little about the share of recurring work actually assisted. Differences in size and digital capability may shape which firms put tools into routine processes; the evidence does not show every large adopter uses tools deeply or gains the same result.

For a local diagnosis, define eligible tasks, then track actual assisted frequency or share over a representative period. Separate occasional experiments from routine use; purchased seats are not completed work. This does not establish a firm productivity effect. It identifies how much work could plausibly carry a task-level gain into the broader measure.

Sources: The Rapid Adoption of Generative AI; A Portrait of AI Adopters Across Countries; The AI Buildout and the Economy: Publicly Available Data to Assess AI’s Impact

What has to change around a faster task for a firm to gain?

A faster task becomes a firm gain only when the rest of the work can use its result. Information must be usable, people able to apply and check the output, and responsibility for the next step clear. If a draft arrives sooner but waits in the same queue, the local saving may stop there. These conditions can help a gain travel through a workflow, but do not guarantee productivity. The OECD’s 2024 report, The Impact of Artificial Intelligence on Productivity, Distribution and Growth, describes AI as a bundle of complementary inputs, including software, skills, data and computing capacity. It treats AI as a possible general-purpose technology whose broader effects depend on diffusion and changes in work organization. A team may need accessible records, connected software, training, and rules for verifying results or handling exceptions. Buying access does not itself supply those complements. The report outlines mechanisms; it does not establish that every firm needs the same investment or that it will pay off. A Portrait of AI Adopters Across Countries, an OECD working paper published in 2023, analyzes official firm surveys from 11 countries using harmonized statistical code. It reports that AI use was more common in ICT and professional services and larger firms; AI users tended to be more productive, especially the largest firms. ICT skills, high-speed digital infrastructure and other digital technologies were significantly related to AI use and appeared to play a critical role in users’ productivity advantages. This is an association, not proof that complements caused the advantage. The study predates the generative-AI wave, so it helps explain why adoption may travel with broader capabilities, but is not a current estimate of generative-AI returns. A randomized field experiment shows why access is not the same as organizational redesign. Shifting Work Patterns with Generative AI studied 7,137 knowledge workers across 66 firms. Workers were randomly selected to receive an AI tool integrated into work applications. In the second half of the six-month experiment, the 80% of treated workers who used it spent two fewer hours on email each week and reduced work outside regular hours. Yet researchers detected no change in task quantity or composition from individual-level provision. That does not mean the tool had no effect: the reported time saving was an individual outcome. Nor does six months show redesign cannot happen later. It does show that local savings did not establish a changed task mix in this study. For a team, ask what changed after the assisted step. Did handoffs, queue rules, data access, training, decision authority or process ownership change enough for the next stage to act sooner or better? If not, identify the blocking step before scaling a claim from one task to the firm. OECD evidence makes complementary capabilities a plausible explanation; the experiment cautions against assuming they follow automatically from individual access. A task gain is evidence about that task. Whether it becomes firm productivity depends on the surrounding process and measuring the result.

Sources: The Impact of Artificial Intelligence on Productivity, Distribution and Growth; A Portrait of AI Adopters Across Countries; Shifting Work Patterns with Generative AI

How can review, correction and coordination consume a speed gain?

A faster first draft or decision is not yet a net productivity gain. Compare the process from task start through an accepted result, including checking, correcting and handing work onward. If those steps absorb the minutes saved upstream, total workflow time may not fall. If review is shortened at the expense of accuracy or accountability, the apparent gain has a quality cost. The Federal Reserve note, *The AI Buildout and the Economy: Publicly Available Data to Assess AI’s Impact*, identifies adjustment costs and bottlenecks elsewhere in production as reasons a task-level improvement need not translate proportionally into firm output. It does not claim that every workplace has a specific review burden. The implication is narrower: measure the route to usable output, because a capability on one step does not establish that the full process became cheaper or faster. Verification matters when errors are costly, source material is incomplete, or someone must explain and stand behind the result. Those conditions vary by task, so a speed figure alone cannot settle the tradeoff. Consider an illustrative document workflow: a summary arrives sooner, but a reviewer checks source records, repairs omissions and asks a specialist to confirm an ambiguity. If the summary is sound, review may stay brief and the workflow can improve. If more drafts arrive than reviewers can clear, the queue moves downstream. This example explains a mechanism; it is not a reported case or measured result from the cited studies. Coordination adds time when people must resolve who owns a correction, which version is authoritative, or whether a decision can proceed. Count those minutes and delays rather than treating them as invisible overhead. A stronger outcome measure captures accepted work, not just a draft. In *Generative AI at Work*, the staggered deployment study of 5,172 customer-support agents reported an average 15% increase in issues resolved per hour. That measure combines speed with successful resolution, making it more informative than drafting time alone. The study concerns one company and a specific support workflow; it does not show that review costs disappeared elsewhere or estimate firm-wide productivity. It illustrates why outcome definition matters: an unaccepted answer would not count as resolved. For a local check, record first-pass time, review and correction time, rework, handoff delay, and the share accepted without material repair. Compare similar work over a defined period and include task-appropriate quality criteria. If accepted output rises while total effort and error burden remain within bounds, that is evidence of a workflow gain in that setting. If only the first step gets faster, or correction and coordination erase the difference, the task has changed but net productivity has not been shown.

Sources: The AI Buildout and the Economy: Publicly Available Data to Assess AI’s Impact; Generative AI at Work; The Impact of Artificial Intelligence on Productivity, Distribution and Growth

Why might saved time fail to become more firm output?

Saved time is a change in capacity, not yet a change in the firm's recorded output. Whether it becomes more output depends on what follows and how the organization uses released time. If another step limits throughput, making an earlier step faster can shorten its queue without increasing completed cases. If demand is fixed, the firm may have no reason to produce more units. Released hours might instead go to quality checks, exceptions, reducing overtime, or postponed work, uses that can matter even when output per hour barely changes. Consider an illustrative workflow: an analyst uses an AI tool to prepare a first-pass review faster, but it still waits for a data handoff and a fixed weekly approval meeting. The faster first pass does not by itself make approvals happen sooner or increase approved decisions. This is an illustrative bottleneck, not an observed case. The Federal Reserve note, “The AI Buildout and the Economy: Publicly Available Data to Assess AI’s Impact,” explains that task experiments can measure speed on a task while adjustment costs or bottlenecks elsewhere erode gains before they reach firm output. Its descriptive sector analysis cannot establish every firm's constraints. The time can still have value. A team might use capacity for additional work if demand and downstream steps allow it, improve quality, take on deferred work, reduce extra hours, or leave slack. The ILO brief “The Aggregation Paradox of AI: Why do micro-economic productivity gains from AI disappear at scale” identifies diffusion and complementary reorganization as explanations for why micro-level gains may not appear uniformly in broader measures. Neither source establishes how a particular firm will use saved time. For a local test, separate time on the assisted task; elapsed time through review and handoffs; accepted output or quality; and the downstream measure the team cares about. Ask what constraint the faster task could relax. If none is observable, local speed may be real while a firm-output gain remains unshown. If a constraint changes, look for a corresponding change in volume, quality, waiting time, or another defined outcome. More completed cases and fewer errors are different results; neither should stand in for the other. This is a practical reasoning approach, not a validated instrument. Test the whole recurring workflow before scaling a claim. Continue when the net result improves on the relevant measure without an unacceptable quality or accountability cost; redesign a handoff if work accumulates there; or stop if review absorbs the saving. Do not infer headcount reductions or an individual's displacement from saved-time arithmetic. Productivity concerns output and inputs under a chosen measure; employment effects require separate evidence about labor demand and the firm's response.

Sources: The AI Buildout and the Economy: Publicly Available Data to Assess AI’s Impact; The Aggregation Paradox of AI: Why do micro-economic productivity gains from AI disappear at scale

What do current firm-level studies actually disagree about?

Current firm-level evidence does not support a universal claim that AI produces no productivity gains. The European Investment Bank’s EIB Working Paper 2026/02 reports a positive estimate for its matched sample, while the Federal Reserve’s 2026 note finds no clear productivity divergence across U.S. industries with different AI exposure. These findings can coexist because they concern different populations, measures and designs. The EIB estimates an association for firms classified by AI adoption; the Federal Reserve tracks sector trends and asks whether more exposed industries are pulling away. Neither directly measures how much one worker’s task improvement raises output at their firm.

The EIB paper uses matched EIBIS–ORBIS data for more than 12,000 non-financial firms in the European Union and United States. Its summary reports a 4% labor-productivity increase associated with AI adoption, concentrated in medium and large firms. The authors instrument EU firms’ adoption using U.S. peer rates, aiming to address the possibility that more productive firms are simply more likely to adopt. The paper attributes the estimate to capital deepening rather than job losses and identifies complementary investment in software, data and workforce training. This is counterevidence to a blanket “no firm gains” conclusion. It remains a working-paper estimate dependent on its sample, measure and instrumental-variable assumptions. The summary describes AI generally, not a generative-AI-specific return for every adopter. A productivity estimate also does not establish worker demand at a particular firm.

The Federal Reserve note takes a broader, descriptive route. It compares labor-productivity trends in U.S. industries grouped by AI exposure, measuring real value added relative to employment and average weekly hours. More exposed industries appear to have higher labor-productivity growth, but trends across groups have been relatively consistent, suggesting micro-level gains have not yet added up to a clear aggregate signal. This is not a firm-by-firm estimate or causal test of adoption. The authors warn that exposed industries may have had different productivity trends before AI, so a later divergence would not automatically prove causation. Their conclusion is time-bounded: no large aggregate signal as of 2026 does not rule out future effects, especially while reported use remains shallow.

The International Labour Organization’s May 2026 research brief offers a synthesis, not a new causal estimate. It describes firm evidence as mixed, adoption as uneven, and gains as concentrated in larger, digitally advanced enterprises; many firms report little measurable impact beyond pilots. This helps explain how positive estimates can coexist with weak broad indicators: gains may be concentrated, adoption may remain shallow, and studies count different outcomes. The brief cannot establish the size of a general firm-level effect on its own.

The fair verdict is conditional. Positive firm-level findings, including the EIB estimate, do not show that every organization captures task-study gains. The Federal Reserve comparison is evidence against claiming a clear broad productivity break in its data, not proof that firm gains are absent. For a worker or team, test whether a local task improvement survives review and handoffs and reaches a valued output measure. Keep the conclusion at the measured level: firm productivity is separate from labor demand, and neither study gives an individual displacement probability.

Sources: EIB Working Paper 2026/02; The AI Buildout and the Economy: Publicly Available Data to Assess AI’s Impact; The Aggregation Paradox of AI: Why do micro-economic productivity gains from AI disappear at scale

Why can measured productivity lag, miss or misattribute a gain?

A productivity measure can miss a real change for a time, record it imperfectly, or show it in a different place from where the AI-assisted work occurred. That makes a weak aggregate signal worth examining, not dismissing. The Federal Reserve note, “The AI Buildout and the Economy: Publicly Available Data to Assess AI’s Impact,” reports that labor-productivity trends across its high-, medium- and low-AI-exposure sector groups remained relatively consistent through its 2026 analysis. The authors call this suggestive of micro-level gains not adding up in aggregate, while cautioning that exposure groups may already have had different productivity trends. Timing is one possible gap. A firm may need to pay for and build complementary inputs, such as data, software, training and process changes, before output reflects the new capability. The OECD report “The Impact of Artificial Intelligence on Productivity, Distribution and Growth” explains the productivity J-curve hypothesis: early intangible investment can be difficult to capture in national accounts, while later benefits may become visible as complementary assets are used. This is a framework for interpreting a possible lag, not evidence that AI gains must eventually appear. The Federal Reserve note likewise says measured gains from general-purpose technologies have historically lagged investment, but explicitly leaves future effects uncertain. A delay is one explanation among several, not a promise about the destination. The unit and definition of productivity also matter. Labor productivity relates output to labor input; total factor productivity tries to account for multiple inputs, including capital. An AI-assisted process may raise measured output per hour partly because a firm adds software or computing capital, while a broader productivity calculation may classify the change differently. The Federal Reserve note says AI effects can be misattributed between capital deepening and total factor productivity, and that complementary intangible assets are often incompletely captured in data. These categories answer different questions, so a number without its denominator and accounting method is difficult to compare with a task-level result. Prices and sector boundaries add further complications. For services, output is often inferred from revenue; if prices fall, revenue-based measures can understate changes in the amount or quality of service delivered. The Fed note also explains that AI’s contribution may be embodied in intermediate inputs and therefore appear in another sector rather than as final output in the firm using it. Neither issue proves that a measured result is wrong; each limits what the statistic represents. For a productivity claim, ask what output was counted, over what period, and whether the measure reflects quality, prices, capital and intangible inputs. Treat a null as evidence about that measure and window, not as meaningless noise or proof that gains are inevitable later.

Sources: The AI Buildout and the Economy: Publicly Available Data to Assess AI’s Impact; The Impact of Artificial Intelligence on Productivity, Distribution and Growth

How should you interpret a productivity claim at work?

A productivity claim applies to the unit, outcome, population and period measured. A faster task is evidence about that task . A survey of use is evidence about reported use. A firm estimate concerns its sampled firms and chosen productivity measure. None, by itself, says whether a particular worker’s role will shrink, expand or stay much the same. The studies in this debate answer different questions. The Quarterly Journal of Economics study reports that access to an assistant raised customer-support agents’ issues resolved per hour by 15% on average at one company. That is a measured worker-level outcome in one workflow; it does not estimate the company’s total output after every cost and input. The Management Science study uses U.S. survey responses to estimate workplace use and assisted hours, alongside reported time savings. It describes reported use in late 2024, not audited firm productivity. The American Economic Review: Insights experiment randomized access among knowledge workers in 66 firms. It found less email time among users, with no detected task quantity or composition change over six months. That finding is limited to this intervention and period. At the firm level, The EIB’s January 2026 working paper reports a positive labor-productivity estimate for a matched sample of more than 12,000 non-financial EU and U.S. firms, using an instrumental-variable design. It is a working-paper result for that sample and method, and its AI measure is not limited to generative AI. By contrast, the Federal Reserve note “The AI Buildout and the Economy: Publicly Available Data to Assess AI’s Impact” finds no clear productivity divergence among sector groups in its descriptive analysis as of 2026; it cautions that underlying sector trends may differ. Their different units, designs, measures and periods allow these results to coexist; neither warrants a universal forecast. For a workplace decision, use a level-matched check. Choose one recurring task and record its usual completion time and accepted output as a baseline. Track how often the tool is actually used, time through final review, quality or rework, handoff delays, and what observable work received the released capacity. This is a reasoned measurement approach, not a validated test. A local improvement supports a decision about that workflow; it cannot alone establish company-wide productivity, labor demand or anyone’s displacement risk.

Sources: Generative AI at Work; The Rapid Adoption of Generative AI; Shifting Work Patterns with Generative AI; EIB Working Paper 2026/02; The AI Buildout and the Economy: Publicly Available Data to Assess AI’s Impact

What is a proportionate next step?

A proportionate next step is a small comparison of one recurring task, carried through to accepted work. First define the outcome that matters and record a baseline from the usual process: elapsed time, review effort, quality and any waiting between handoffs. Then try the AI-assisted version on a bounded sample, recording whether it was used each time and how much checking or correction the result required. Compare like with like, including the downstream wait rather than stopping the clock when a first draft appears. The useful result might be more accepted output in the same time, the same output with less effort, or a quality improvement that matters to the work. If the apparent saving shifts effort to reviewers or weakens accountability, redesign the workflow or stop; speed at the first step is not enough. The Federal Reserve note, “The AI Buildout and the Economy: Publicly Available Data to Assess AI's Impact,” explains why adjustment costs and bottlenecks can interrupt the path from a task result to firm productivity. The Quarterly Journal of Economics study “Generative AI at Work” measured issues resolved per hour in a particular customer-support deployment, illustrating why the outcome and setting must be named. A local comparison can guide a decision about that workflow, but it cannot establish a firm-wide productivity effect or predict redundancy. If you want to map which parts of your own work face change pressure, the free task checker can help organize that question; its signals are task-level, not a validated probability of displacement.

Sources: The AI Buildout and the Economy: Publicly Available Data to Assess AI’s Impact; Generative AI at Work

Questions readers ask