Productivity Gains Across Sectors

Dr. Rachel Kim used to lose four hours a day to her keyboard. After the patients — the diagnosing, the treating, the reassuring — came the documentation: the notes, the codes, the translation of messy human illness into the bureaucratic dialect insurers demand. She would type until the evening emptied out. It was exhausting, and it was the job.

Then her hospital installed an AI scribe. The system listens to the visit, pulls out what matters, and drafts the note in real time. Dr. Kim now spends about forty-five minutes on paperwork instead of four hours. She gets home before her children are asleep. By any sensible measure, she has become far more productive.

Now hold that story against a different one. In the summer of 2025, researchers at METR ran a careful experiment with sixteen seasoned open-source software developers — the kind who maintain projects with millions of lines of code. Each was given real tasks from their own repositories, some to do with AI assistance, some without. Before starting, the developers predicted the AI would make them 24 percent faster. Afterward, they reported that it had made them roughly 20 percent faster. The stopwatch disagreed. With AI, they were 19 percent slower (METR, 2025). They felt accelerated while they were being slowed down, and they never noticed the difference.

Both stories are true, and the gap between them is the subject of this chapter. AI is producing real, sometimes dramatic gains for individual workers and firms. It is also producing confident illusions, mismeasured savings, and benefits that vanish the moment you zoom out to the level of a whole economy. Understanding where the gains are real, where they are imagined, who captures them, and why they have not yet shown up in the national statistics is one of the central economic puzzles of the moment.

What the Gains Actually Are

Start with the measured reality at the worker level, because that is where the evidence is strongest. When researchers put people in controlled settings and measure their output with and without AI, the tools usually help. In the landmark study of customer-support agents by Erik Brynjolfsson, Danielle Li, and Lindsey Raymond, access to a generative AI assistant raised the number of issues resolved per hour by about 15 percent on average — and by as much as 34 percent for the least experienced agents (Brynjolfsson et al., Generative AI at Work, NBER). Studies of professional writing find that ChatGPT cuts the time to complete tasks and compresses the quality gap between weak and strong writers by roughly 40 percent. GitHub Copilot has let programmers finish narrow coding tasks around 56 percent faster in controlled trials.

Zoom out one level, from the task to the firm, and the numbers shrink fast. The U.S. Federal Reserve estimates that generative AI currently saves the average worker about 5.4 percent of total work hours — a little over two hours in a forty-hour week (St. Louis Fed, 2025). A 2026 survey of corporate executives by the Federal Reserve Bank of Atlanta found firms reporting an average productivity gain from AI of about 1.8 percent in 2025, with large firms expecting output per worker to be around 3 percent higher in 2026 than it otherwise would have been (Atlanta Fed, 2026). And when economists tried to verify those self-reports against actual revenue data, the implied gain shrank again — to a fraction of a percent.

That shrinkage is the whole story in miniature. Controlled studies show task-level gains of 15 to 55 percent. Firm surveys report a few percent. Hard revenue data shows a fraction of a percent. National statistics show almost nothing attributable to AI at all. The gain does not disappear all at once — it leaks away at every step between the individual keystroke and the aggregate ledger.

Where the Gains Land — and Where They Don't

The leakage is not uniform. AI's benefits are pooling in some corners of the economy and barely dampening others, and the pattern is consistent enough to be predictive. The sectors capturing the most are the ones whose work is made of words, code, and numbers — information that current AI can read and generate. The sectors capturing the least are the ones made of physical acts: lifting, driving, building, serving.

The evidence lines up cleanly. In the Atlanta Fed's revenue-based analysis, high-skill services and finance showed implied 2025 labor-productivity gains of roughly 0.8 percent, while low-skill services, manufacturing, and construction clustered around 0.4 percent — half as much. Information-services workers spend the largest share of their hours using generative AI; workers in leisure, accommodation, and physical trades spend the least. The tools are best at exactly the tasks that dominate high-wage desk work, which means the first wave of AI productivity is flowing toward sectors and workers that were already highly productive and well paid.

Sector AI exposure Reported/implied productivity effect Leading applications
Information services & software Highest gen-AI use Largest task-level gains; ~0.8% implied firm gain Coding assistants, content, analysis
Finance $20B+ invested ~0.8% implied gain; large loss-prevention value Fraud detection, trading, personalization
Healthcare Fast-growing adoption Up to ~40% in specific diagnostic tasks Imaging triage, documentation, billing
Manufacturing ~77% of firms piloting ~0.4% implied; ~23% downtime reduction Predictive maintenance, quality control
Retail, accommodation, trades Lowest gen-AI use ~0.4% or less Inventory, basic customer service

The applications that generate the most consistently documented gains share a signature: they take a repetitive, pattern-heavy, information-rich task and let a machine do the first pass. In radiology, an AI platform integrated into chest-CT interpretation cut reading time by 22 percent, and AI-assisted breast-cancer screening has reduced false positives by 5.7 percent and false negatives by 9.4 percent while slashing the second-reader workload by 88 percent (structured reviews in PMC, 2025). In manufacturing, predictive-maintenance systems that watch sensor data for the signature of a failing bearing have cut unplanned downtime by around 23 percent. In finance, fraud-detection engines scan millions of transactions a second and stop billions of dollars in losses a year. In software, coding assistants draft boilerplate and suggest completions. Diagnostics, predictive maintenance, fraud detection, and code generation are the four workhorses — narrow, measurable, and repeatable.

The Novice-or-Expert Puzzle

Even inside a single AI-adopting firm, the gains fall unevenly across the workforce — and the research on who benefits most points in two opposite directions at once.

One body of evidence says AI is a great leveller. The support-agent study found the biggest gains among the least experienced workers; the writing studies found AI compressing the quality distribution by lifting weaker performers. The intuition is that AI behaves like a capable junior colleague who has read everything: it raises the floor for people who lack expertise and adds little for those who already have it. If that is the dominant effect, AI could narrow skill-based inequality within occupations.

The other body of evidence says the reverse. A separate strand of research finds AI boosting the performance of highly skilled workers substantially while doing little for novices, on the theory that expert judgment is what turns a raw AI draft into something valuable. And then there is the METR result from the opening — where the most experienced developers were slowed down, not sped up, because integrating AI suggestions into a large, unfamiliar-to-the-model codebase cost them more time in review and correction than the suggestions saved.

These findings are not simply contradicting each other by mistake. They are measuring genuinely different situations: different tasks, different tools, different levels of task complexity, different amounts of context the model can and cannot see. AI seems to help most when the task is well-scoped and the worker is below the model's competence, and to help least — or hurt — when the task is sprawling and the worker already exceeds it. The lesson is not that one camp is wrong. It is that a single number labelled "AI productivity" hides a landscape too varied to be summarized by one figure. The METR team underlined this themselves: after finding selection issues in their first study, they revised their estimate to a milder 4 percent slowdown with a confidence interval running from a 15 percent slowdown to a 9 percent speedup (METR, 2026). The honest reading of the coding evidence is not "AI slows experts down." It is "we are much less sure than the marketing implies."

Solow's Ghost

Now the paradox itself. Add up all these firm-level gains, real and reported, and you would expect the national productivity statistics to be climbing. For the most part, they are not — or at least, not in a way anyone can confidently pin on AI. In early 2026 Goldman Sachs economists reported that they still could find "no meaningful relationship between AI and productivity at the economy-wide level," even as the management teams they surveyed claimed a median 30 percent gain on specific tasks (Goldman Sachs, 2026). A survey of thousands of chief executives that same season found most admitting AI had made no measurable difference to their firms' productivity or employment — a result striking enough that economists began invoking a ghost from four decades earlier (Fortune, 2026).

The ghost is Robert Solow, who quipped in 1987 that "you can see the computer age everywhere but in the productivity statistics." His paradox described the strange two-decade stretch when computers filled every office but the productivity numbers stayed flat. We are living through its sequel. Call it the AI productivity paradox: a technology that visibly changes how millions of people work, without yet changing what the economy measurably produces.

There is a complication worth naming. U.S. nonfarm productivity did jump in late 2025 — up 4.9 percent in the third quarter, with unit labor costs falling for two straight quarters (Bureau of Labor Statistics, 2025). But almost no economist attributes that spike to AI; it tracks the business cycle and post-pandemic normalization more than any technological wave. The absence of a clear AI signal in the macro data is exactly what needs explaining.

Four Suspects

Researchers have narrowed the explanations to four, and they are not mutually exclusive.

The first is mismeasurement. GDP was built to count physical output, and much of what AI improves is not extra output but better, faster, less painful work. When Dr. Kim reclaims three hours of her evening without seeing more patients, national accounts register nothing. When a diagnosis is more accurate but not more numerous, when a fraud is prevented rather than a product produced, when a document is less stressful to write — value is created that the standard yardstick was never designed to catch. An economy that has shifted toward information and services is being measured with a ruler built for factories.

The second is redistribution. Perhaps the gains are real and correctly measured, but they are flowing to capital rather than labor, so they show up as profits rather than as broad productivity or wages. There is unsettling support for this. AI investment concentrates in a handful of "superstar" firms with the data and compute to exploit it; the labor share of income has continued its long decline; and between-firm wage dispersion — the gap between what leading and lagging firms pay — remains the dominant driver of rising U.S. earnings inequality and shows no sign of reversing (IMF, AI Adoption and Inequality, 2025). Gains that accrue to shareholders do not register in the statistics that track how ordinary workers are doing.

The third is overstated firm-level results — the possibility that the impressive numbers are partly an artifact of cherry-picked case studies, survivorship bias, and vendors and executives with reasons to report success. The METR developers who felt 20 percent faster while running 19 percent slower are a warning here: self-reported productivity gains, which underlie many of the rosier surveys, can be systematically wrong. When 93 percent of developers use AI coding tools but organizational productivity has barely moved, the reported task-level gains may simply not be adding up to anything at the level that counts.

The fourth, and the one most economists lean on, is implementation lag. New general-purpose technologies have always demanded years of complementary investment — new infrastructure, retrained workers, redesigned workflows, updated rules — before their benefits appear in aggregate data. The hospital has an AI scribe, but the insurer still demands documentation in formats designed for paper. The factory has predictive-maintenance software, but not enough technicians trained to act on its alerts. The technology has arrived; the surrounding machinery of the economy has not caught up.

graph TD
    A[Task-level AI gain<br/>15–55%] --> B[Firm-level gain<br/>~2%]
    B --> C[Revenue-verified gain<br/>fraction of 1%]
    C --> D[Macro productivity<br/>no clear AI signal]
    A -.mismeasurement.-> D
    A -.redistribution to capital.-> D
    A -.overstated results.-> D
    A -.implementation lag.-> D

The four suspects are hard to separate because they can all be operating at once, and the honest position is that we cannot yet say which dominates. The implementation-lag story is the most reassuring and the best supported by history, but confidence in it is a bet about the future, not a reading of the present. If lag is the answer, the gains are coming. If redistribution is the answer, they are here — just not for most people. If mismeasurement is the answer, they are here but invisible. Distinguishing these matters enormously, and the current evidence does not clinch it.

Why Words Beat Widgets

The lopsided sectoral pattern has a cause worth spelling out, because it shapes everything downstream. Today's AI is, at bottom, a system for processing symbols — text, code, images, structured data. It excels where the work product is itself a symbol and where the raw material can be fed to it cheaply. A legal brief, a marketing email, a block of code, a radiology image, and a loan application are all, in this sense, the same kind of thing: information in, information out.

Physical service and trade sectors are different in kind. A restaurant meal, a haircut, a repaired roof, and a delivered package require a body acting on the physical world in real time. AI can help schedule the plumber, optimize the delivery route, and manage the restaurant's inventory — but it cannot yet fix the pipe, cut the hair, or carry the box. The binding constraint in those sectors is physical labor, and software does not relax it. Until robotics catches up to language models — a gap that remains wide — the sectors that employ the most people and most need a productivity lift are structurally on the outside of this wave. Whether they stay there is one of the genuine unknowns: it is possible that AI's current form is simply the wrong tool for physical work, not an early version of the right one.

The Risks Hiding Inside the Gains

Efficiency is not free of hazard, and each sector's signature application carries a second-order risk that the headline numbers ignore.

In finance, the same algorithmic trading that improves liquidity and shaves transaction costs also synchronizes behavior across the market. When many firms run similar models trained on similar data, they can stampede in the same direction at the same microsecond, turning a small shock into a flash crash. Efficiency at the level of a single trade can become fragility at the level of the whole system.

In healthcare, the danger is subtler and arguably more troubling: skill decay. When endoscopists who had grown used to AI assistance went back to working without it, their adenoma detection rate — a core measure of how many precancerous growths they catch — fell from 28.4 percent to 22.4 percent (widely reported from European colonoscopy data in 2025). The AI had quietly eroded the very expertise it was meant to augment. And automation bias cuts the other way too: in a controlled study, deliberately wrong AI prompts raised false-positive recalls by up to 12 percent even among experienced breast-imaging radiologists. A tool that makes the average case faster can make the unusual case more dangerous, precisely because the human stops looking as hard.

The third risk is structural: automating part of a process can create a new bottleneck rather than removing one. If AI helps developers write twice as much code but code still has to be reviewed by a human, the review queue becomes the choke point — one reason the pull-request review burden climbs even as code output rises. Speeding up one stage of a pipeline while leaving the next untouched does not speed up the pipeline; it just moves the congestion.

Who Gets the Gains

Suppose the optimists are right and the aggregate gains eventually arrive. The distributional question remains, and on current trajectory the answer is uncomfortable. If AI's benefits stay concentrated in high-productivity sectors and among high-skill workers and capital owners, the technology will widen the gaps it touches rather than close them. Information-sector firms pull further ahead of physical-service firms. Superstar companies with the compute and data pull further ahead of everyone else. Shareholders capture returns that never reach wages. The IMF's modelling suggests that when firms can choose how much to adopt, the effects on wealth inequality are especially pronounced, because the capital that owns the AI compounds its advantage while the labor it displaces does not.

This is where measurement and morality intersect. If AI raises output while flowing overwhelmingly to capital, standard statistics may eventually record it as "productivity growth" and call it progress. But a productivity gain that lifts profits while leaving most workers' pay and prospects untouched is progress only in a narrow, accounting sense. Whether it counts as genuine economic advance depends on criteria the productivity number does not contain — whether living standards rise broadly, whether the gains are shared, whether the people whose work was automated are better or worse off. The distribution question is not a footnote to the productivity story. It is the story.

Making the Gains Spread

For AI's benefits to travel from early adopters to the wider economy, the technology is the least of what is required. The binding inputs are complementary: digital infrastructure so that data can actually flow; workforce skills so that people can use the tools well and know when not to trust them; organizational redesign so that workflows are rebuilt around the new capability rather than bolted onto the old one; and regulatory adaptation so that rules written for a paper world stop forcing digital work back into analog formats. The manufacturing J-curve — the well-documented pattern in which firms' productivity dips after adopting AI before it recovers — is precisely the signature of these complementary investments being made. The dip is the cost of rebuilding the system; the recovery, for firms with the capital and patience to reach it, is the payoff. Firms that install the software without doing the surrounding work stay stuck in the trough.

What History Suggests About Timing

The most useful frame for the paradox is historical, because we have watched this movie before. Electricity was commercially available from the 1880s, but the productivity surge it enabled did not show up in the data until the 1920s. The reason was not the dynamo; it was the factory. Manufacturers had to stop replacing their single central steam engine with a single central electric motor and instead redesign the entire plant around many small motors, each driving its own machine — a reorganization that took a generation. Computers tell the same story: they spread through offices in the 1970s and 1980s, and the productivity payoff became clearly visible only in the late 1990s. Solow's paradox resolved itself; the gains were real, just delayed by the slow work of institutional adaptation.

By this analogy, 2026 looks like the mid-1980s of computing or the 1900s of electrification: the technology is real, adoption is climbing, the complementary rebuild is underway, and the aggregate payoff is still ahead. Goldman Sachs, having found no economy-wide effect yet, nonetheless forecasts that AI will begin lifting productivity and GDP around 2027 and continue through the 2030s, adding perhaps 0.4 percentage points to annual growth later in the decade (Goldman Sachs, 2026). That is meaningful but not miraculous — an acceleration, not a discontinuity. The plausible range for when aggregate gains materialize clusters in the late 2020s to mid-2030s, with wide error bars in both directions.

The historical parallel is a source of hope, not proof. It is possible that AI diffuses faster than electricity did, because it rides on infrastructure that already exists and requires no new wires in the ground. It is equally possible that it diffuses more slowly or more narrowly, delivering its gains to capital and a slice of high-skill workers while the physical-service majority waits for a robotics revolution that may be decades away. The precedents tell us the lag is normal. They do not guarantee the destination.

What the Numbers Should Learn to See

There is a final point that sits underneath all the others. Part of the reason AI's gains are hard to find is that we are looking with the wrong instruments. Productivity, as conventionally measured, is output per hour — and it was designed for an economy that made things you could count and stack. It is poorly suited to an economy that increasingly produces quality, safety, and relief.

If AI gives a physician back three hours a day, that is a real gain in human welfare even though no additional patient was seen. If it prevents a catastrophic equipment failure, that shows up in the statistics only as the absence of a disaster. If it stops a fraudulent transaction, value is preserved though nothing was produced. If it lowers the cognitive burden of complex work, the benefit is felt but not recorded. A measurement framework that captures none of these is not neutral; it is systematically blind to exactly the kinds of value AI is best at creating. Updating productivity measurement to register quality improvements, reduced cognitive load, and harm prevented is not an accounting nicety. It is a precondition for knowing whether this technology is delivering on its promise at all — and for making sure that, when the gains do arrive, we measure them well enough to argue about who should get them.

Summary

AI produces real and sometimes large productivity gains at the level of individual tasks and workers — 15 to 55 percent in controlled studies — but those gains shrink at every step toward the aggregate, and have not yet appeared as a clear signal in national productivity statistics. This is the AI productivity paradox, a direct echo of Robert Solow's 1987 observation about computers. The gains are heavily concentrated in information-intensive sectors — software, finance, and specific healthcare diagnostics — whose work is made of symbols that current AI can process, and largely absent from physical service and trade sectors whose work requires a body acting in the world. The best-documented applications are diagnostics, predictive maintenance, fraud detection, and coding assistance: narrow, repetitive, information-rich tasks where a machine can take a reliable first pass.

Four explanations compete for the macro-level gap: mismeasurement of quality and welfare gains in a service economy, redistribution of gains from labor to capital, overstated firm-level results, and implementation lag. Most economists favor lag, drawing on the electrification and computing precedents where aggregate gains arrived decades after the technology, but the evidence does not yet decisively rank the four, and the choice among them determines whether the gains are coming, already captured by a few, or merely invisible. The distributional stakes are high: on current trajectory, benefits concentrate among high-skill workers, superstar firms, and capital owners, which would widen inequality rather than narrow it. Whether this counts as genuine progress depends on criteria — broad living standards, shared gains — that the productivity number does not contain, and capturing AI's true impact will require measurement frameworks that can finally see quality, reduced burden, and harm prevented.

Key Takeaways

  1. AI's productivity gains are real but leak away with scale: 15–55 percent at the task level in controlled studies, a few percent in firm surveys, a fraction of a percent in verified revenue data, and no clear signal in national statistics — a pattern economists call the AI productivity paradox, after Solow's 1987 quip about computers.
  2. Self-reported gains are unreliable. In METR's 2025 trial, experienced developers felt AI made them 20 percent faster while it actually made them 19 percent slower — a warning that inflates many optimistic surveys. A 2026 revision softened the slowdown to about 4 percent, underscoring how uncertain the coding evidence is.
  3. The gains concentrate where work is made of symbols — software, finance, diagnostics — and barely touch physical service and trade sectors, because today's AI processes information but cannot yet act on the physical world. This favors sectors and workers that were already high-productivity and well-paid.
  4. Whether AI helps novices or experts more depends on the task and context; the contradictory findings reflect genuine variation, not error, and a single "AI productivity" number hides that landscape.
  5. Four explanations for the macro gap — mismeasurement, redistribution to capital, overstated results, and implementation lag — are not mutually exclusive; lag is best supported by history, but which dominates is genuinely unresolved and determines whether the gains are coming, captured, or invisible.
  6. Efficiency carries second-order risks: synchronized algorithmic trading can amplify market crashes, AI diagnostics can erode clinicians' skills (colonoscopy detection fell from 28.4 to 22.4 percent after AI was removed) and induce automation bias, and automating part of a process can shift bottlenecks rather than remove them.
  7. Historical precedent — electrification's 40-year lag, computing's 20-year lag — suggests aggregate AI gains may arrive in the late 2020s to mid-2030s (Goldman Sachs projects roughly +0.4 points to annual growth from around 2027), but the precedents promise a lag, not a destination, and the distribution of the eventual gains is a political choice, not a technical certainty.

Sources

Last updated: 2026-07-07

V2 (in progress) Previous: V1