What workplace AI studies actually show

"AI makes workers 40% faster." "AI slows developers down." Both headlines come from careful studies. Neither tells you what will happen when your team starts using an assistant on Monday.

The gap lies in what the studies measured. One timed professionals writing short documents in a single online session. Another counted chats resolved per hour across months of real customer support. A third had experienced programmers fix real issues in their own codebases and found they got slower. Each result holds for its own setting, and very few hold beyond it.

This review covers the studies that most often anchor claims about workplace AI productivity, plus the main 2026 additions. It sets out what each measured, what it found, and where its limits sit. Our companion piece on AI and jobs deals with employment, hiring and wages. This one stays at the level of tasks: what happens to a piece of work when a person does it with AI.

Our assessment: the evidence clearly supports one claim. On well-specified tasks that the tool handles well, assistants usually make people faster, and less experienced workers often gain the most. It does not support a single productivity number for "AI at work". The same studies show losses on tasks outside the tool's competence. They also show a slowdown among experts in complex codebases, and much smaller effects once you look at a whole working week.

Why task-level studies need their own reading

A productivity study answers a narrow question: when people do this task with this tool, what changes in this measure? The measure might be minutes per task, tasks per week, a grader's quality score or the share of correct answers. Each one can move in a different direction.

That narrowness is the strength of the best studies, not a defect. The trouble starts when a result is lifted out of its setting. A time saving on a 30-minute writing exercise does not tell you about a full working week. A gain in a support centre does not tell you about a law firm. And a study of 2023 tools tells you little about 2026 tools in either direction.

Four details decide how far a result can travel:

• The task. Was it real work from the participants' own jobs, a realistic task designed by researchers, or a standard exercise?
• The people. Who took part, how experienced were they, and how were they recruited?
• The tool and the date. Which model or product, and when? Effects measured on early tools may not hold for later ones.
• The outcome. Speed, volume, quality, correctness, or several of these at once?

The same discipline applies to AI benchmarks. Our guide to what AI benchmarks actually measure makes the parallel point for model scores.

The studies, one by one

Sources: Peng et al.,[1] Noy and Zhang,[2] Dell'Acqua et al.,[3] Brynjolfsson, Li and Raymond,[4] Cui et al.,[5] METR,[6] and Dillon et al.[7] Tool details are as reported by each paper.

Customer support: real work, one firm

Brynjolfsson, Li and Raymond studied the staggered rollout of a GPT-3-based assistant that suggested replies to chat agents. The sample covered 3 million chats by 5,172 agents at a Fortune 500 firm selling business-process software. Most agents (89%) worked outside the United States, mainly in the Philippines.[4]

Access raised issues resolved per hour by 15% on average. In the preferred specification that is 0.30 more resolutions per hour, or 15.2%. Less skilled and less experienced agents improved by 30%. Agents with two months of tenure and AI access performed as well as agents without it who had more than six months. The most experienced and highest-skilled agents gained little speed and saw small declines in quality.[4]

The gain came mostly through speed. Chats got shorter and agents handled more of them at once. In the main results table, the estimated changes in resolution rate and in customer satisfaction scores were small and not statistically significant.[4] Customers wrote more positively and asked for a manager less often, and attrition fell, mainly among newer agents.

The authors draw the limit themselves:

"We emphasize that these findings capture medium-run effects in a single firm."

— Brynjolfsson, Li and Raymond, Generative AI at Work, Quarterly Journal of Economics (2025).[4]

The design is also a staggered rollout rather than random assignment, so the estimate relies on comparing agents who got access at different times.

Consulting: the "jagged frontier"

Dell'Acqua and colleagues ran a preregistered experiment with 758 Boston Consulting Group consultants, about 7% of the firm's individual-contributor consulting workforce. Participants were randomly assigned to no AI, GPT-4, or GPT-4 plus a short prompt-engineering overview. The paper, first circulated in 2023, was published in Organization Science in 2026.[3]

On 18 product-development tasks designed to sit inside GPT-4's capabilities, consultants with AI completed 12.2% more tasks and worked 25.1% faster on average. Human graders rated their work more than 30% higher in quality. Consultants in the bottom half of a baseline assessment gained most, but the top half also gained.[3]

The second task was built to sit outside the frontier. It was a business case where the spreadsheet pointed one way and the interview notes, read carefully, pointed another. Without AI, about 84.5% of consultants reached the correct recommendation. With AI, 70.6% did in the GPT-only group and 60% in the group given the prompting overview: an average drop of 19 percentage points.[3]

The authors name their limits. There was a single outside-frontier task. Much of the inside-frontier grading was subjective. Participants were early in their careers. And the setting rewarded quality within a fixed time rather than speed.[3]

Professional writing: fast, but a single session

Noy and Zhang gave 453 college-educated professionals occupation-specific, incentivised writing tasks in a preregistered online experiment, and randomly gave half of them ChatGPT. Average time fell by 40% and graded quality rose by 18%. Inequality between workers narrowed, with weaker writers benefiting most.[2]

Participants exposed to ChatGPT were twice as likely to report using it in their real job two weeks later, and 1.6 times as likely two months later. That is a signal about adoption, not about productivity on the job. The productivity result itself comes from short tasks done once, online, for the study.[2]

Coding in the lab: one task, large effect

Peng and colleagues, then at GitHub and Microsoft, recruited 95 professional programmers through Upwork in May and June 2022. They randomly assigned 45 to use GitHub Copilot and 50 to work without it, and asked everyone to build an HTTP server in JavaScript as fast as possible.[1]

Among those who finished, the Copilot group took 71 minutes on average against 161 for the control group. That is 55.8% faster, with a 95% confidence interval of 21% to 89%. The difference in success rates, 7 percentage points, was not statistically significant. Less experienced developers benefited more. The authors note that they did not examine code quality, and that one standardised task may not generalise to large collaborative projects.[1]

Coding at work: three companies, smaller effect

Cui and colleagues analysed three randomised experiments that Microsoft, Accenture and an anonymous Fortune 100 manufacturer ran in the ordinary course of business, covering 4,867 developers. The tool was the 2022–23 version of GitHub Copilot, an autocomplete assistant. The paper was published in Management Science in 2026.[5]

Pooled across the three, developers who adopted the tool completed 26.08% more pull requests per week (standard error 10.3%). Two qualifiers matter. First, around 30–40% of developers given access never tried it, so the headline is the effect on those who did adopt. Second, each experiment on its own was noisy, and the estimates varied widely between firms.[5]

Quality signals were mixed. At Microsoft, pull-request approval rates rose by about 10%. At Accenture, the share of successful code builds fell by 17.4%. Short-tenure and junior developers gained more than senior ones, though those differences were not statistically significant.[5] Two of the six authors work at Microsoft, which owns GitHub; the experiments were not pre-registered, and the paper says so.

Coding in mature open-source projects: a slowdown

METR's 2025 study is the most cited counterweight. Sixteen experienced developers worked on 246 real issues in large, mature open-source repositories where they had contributed for an average of five years. Each issue was randomly assigned to allow or forbid AI tools. When AI was allowed, developers mainly used Cursor Pro with Claude 3.5/3.7 Sonnet.[6]

Allowing AI increased completion time by 19%, with a confidence interval of +2% to +39%. Developers had forecast a 24% time saving beforehand. Afterwards they estimated they had saved 20%. Experts in economics and machine learning had predicted savings of 39% and 38%.[6]

The authors spell out what the study does not show. It is not evidence that AI fails to speed up most developers, or developers outside these repositories, or developers using later tools.[6]

In February 2026 METR reported a follow-up with 57 developers across 143 repositories. Raw estimates now pointed to speed-ups: 18% for returning developers (CI −38% to +9%) and 4% for new ones (CI −15% to +9%). But 30% to 50% of developers said they were withholding tasks they did not want to do without AI. Pay had also dropped from $150 to $50 an hour.[8]

"our data is only very weak evidence for the size of this increase."

— METR, We are Changing our Developer Productivity Experiment Design (24 February 2026).[8]

METR believes developers were likely sped up more in early 2026 than in early 2025. It also treats the new estimate as a probable lower bound and is redesigning the study.[8] The 2025 slowdown remains a valid result for its tools and setting. What changed is how hard such a result has become to measure.

Knowledge work across a whole week

Dillon and colleagues ran a six-month randomised experiment across 66 firms and 7,137 knowledge workers. It tested Microsoft 365 Copilot, an assistant built into email, meetings and document tools. In the second half of the experiment, the 80% of treated workers who used the tool spent about two fewer hours a week on email and worked less outside regular hours. Beyond those time savings, the researchers did not detect shifts in the quantity or mix of workers' tasks from giving AI to individuals.[7]

AI Chronicle

Who benefited

The most consistent pattern across studies is compression: AI helps less experienced or lower-performing workers more than top performers.

• In customer support, the least skilled and least experienced agents gained 30%, while top agents saw small quality declines.[4]
• Among BCG consultants, the bottom half of performers gained most on inside-frontier tasks.[3]
• In professional writing, inequality between workers decreased.[2]
• In both coding studies, less experienced developers gained more, though the field differences were imprecise.[1][5]

A 2026 randomised experiment by Cruces and colleagues tested this outside firms. It recruited 1,174 adults aged 25 to 45 for an incentivised business problem-solving task. Without AI, participants with more education outscored those with less by 0.548 standard deviations. With AI, the gap fell to 0.139. When AI was then removed for a follow-up module, treated participants did no worse than controls. But a sizable education gap re-emerged.[9]

So AI narrowed differences in task execution while the tool was available. It did not erase differences in unassisted skill. That distinction matters for training. A junior worker who looks as productive as a senior one with AI may not yet have the judgement to spot when the AI is wrong.

Compression is a tendency, not a law. The Dell'Acqua paper cites other research in which higher performers benefited more. The METR slowdown fell on highly experienced developers working in code they knew intimately, where the tool had the least to add.[3][6]

AI can also stand in for some of what a colleague provides. In a preregistered experiment with 791 Procter & Gamble professionals during a one-day product-development workshop, individuals working with AI matched the solution quality of two-person teams working without it.[10] That was a single day, in one company, on early-stage innovation tasks, and the authors flag both limits.

Where AI hurt performance

Losses appear in four forms across these studies. Each is a warning about where a team's own pilot should look.

Tasks outside the tool's competence. The BCG outside-frontier result is the clearest case. The task looked no harder than the others, yet consultants with AI were 19 percentage points less likely to get it right. The group given extra prompting guidance fared worst.[3] The danger was not only wrong answers. AI-assisted recommendations were rated more coherent and persuasive whether or not they were correct:

"AI improves presentation and argumentation even when the underlying analysis is flawed."

— Dell'Acqua et al., Organization Science (2026).[3]

A polished wrong answer is harder to catch than a clumsy one. Review steps need to check the reasoning and the source data, not the prose.

Experts on work they already do well. In customer support, top agents' conversation quality slipped slightly even as they followed AI suggestions more over time.[4] In METR's 2025 study, experienced maintainers on mature codebases were slower with AI.[6]

Quality signals that move the other way. At Accenture, more Copilot use came with a lower share of successful builds.[5] Volume and quality can diverge, so a study that only counts output can miss a cost.

Self-reports that run ahead of reality. METR's developers believed they had saved 20% when they had in fact taken 19% longer. Peng's participants, in contrast, underestimated their measured gain (35% estimated against 55.8% measured).[6][1] Either way, how fast people feel is not a measurement.

From faster tasks to a faster week

Task-level gains are large. Whole-week effects measured so far are small. Both can be true at once.

Bick, Blandin and Deming ran nationally representative US surveys. By late 2024, 23% of employed respondents had used generative AI for work in the previous week, and 9% used it every workday. Between 1% and 5% of all work hours were assisted by it. Respondents reported time savings equivalent to 1.4% of total work hours.[11]

In Denmark, Humlum and Vestergaard surveyed 25,000 workers across 7,000 workplaces in 11 highly exposed occupations. Adopters reported saving about 3% of their work hours. Most chatbot users (85%) said they moved the saved time to other tasks.[12] Dillon's Copilot experiment found time saved on email but no detectable change in the mix of tasks.[7]

The arithmetic explains much of the gap. A 40% cut on a task that fills a tenth of someone's week saves 4% of the week at most. That is before any time spent prompting, checking and correcting. Not everyone uses the tool, and not every task suits it. How saved time turns into earnings or employment is a labour-market question, covered in our jobs evidence review.

How much it generalises

No single study has every feature you would want. Lab studies randomise cleanly but use one short task. Field studies capture real work but measure quality only partly, and some lack random assignment. Recent tools are under-studied because good experiments take years to publish.

AI Chronicle

Read across the matrix and three findings recur in studies with different strengths:

1. On well-specified tasks the tool handles well, people are usually faster. This shows up in the lab, online, in the field and across firms.
2. Less experienced workers tend to gain more. It appears in support, consulting, writing and coding, with exceptions.
3. There is a real risk on tasks the tool handles badly. That covers both wrong answers and slower work, and the risk is hard to see in advance.

Three things do not travel:

1. The size of any one estimate. The 55.8% and the 15% measure different outcomes in different settings. Neither predicts your team's number.
2. Results for older tools. Most of these studies used 2022–2023 tools, and METR's own follow-up shows how quickly tool use changed. That cuts both ways: newer tools may help more, and they may also fail differently.
3. A one-session result as a forecast for a job. Real jobs mix tasks inside and outside the frontier, and include coordination, review and judgement that single-task studies leave out.

Demos and vendor case studies sit even further from this evidence than the weakest study here. Our piece on what AI agent demos leave out covers the same problem for agents.

A worked example: how to read a productivity claim

Hypothetical. The vendor, study and numbers below are invented to show the method. They do not describe any real product or result.

A vendor's white paper says: "Analysts using our assistant were 40% faster." Its methods appendix reports the following:

• 60 analysts volunteered from three client firms. Half were randomly given the assistant.
• Each wrote one summary memo in a single 45-minute session.
• The outcome was time to submit a first draft.
• Drafts were rated for quality by the vendor's own staff, who knew which group each analyst was in.
• The 95% confidence interval for the time saving was 8% to 72%.

Six questions turn the headline into something usable:

Now the arithmetic for a team. Suppose drafting memos takes 6 hours of a 40-hour week. A 40% cut on drafting saves 2.4 hours, or 6% of the week, at most. If checking the assistant's drafts adds an hour a week, the net saving is 1.4 hours, or 3.5%. Use the bottom of the interval, 8%, and drafting saves under half an hour. After checking, the team is slower.

Those figures are illustrative, not a forecast. The habit is the point: convert a task-level percentage into hours of a real week, subtract the new work the tool creates, and run the calculation again at the low end of the interval. Our guide to measuring AI productivity without hiding review work sets out how to collect those numbers in your own pilot.

Questions readers commonly ask

So does AI make workers more productive?

On many well-specified tasks, yes. Controlled and field studies both find speed or output gains, often large. On some tasks it makes work worse or slower, and measured whole-week time savings so far are around 1–3% of hours.[11][12] "It depends on the task" is the finding, not an evasion.

Why did METR find a slowdown when other coding studies found speed-ups?

The settings differ. Peng timed one standard task in a lab, and Cui counted output across thousands of developers using 2022–23 autocomplete. METR studied highly experienced maintainers on large, mature codebases they knew well, using early-2025 tools, and measured time on real issues.[1][5][6] Each result can be right for its setting. METR's 2026 follow-up suggests speed-ups became likely later, but its authors say they can no longer measure the size reliably with that design.[8]

Do junior staff always gain the most?

Often, not always. Less experienced workers gained most in support, consulting, writing and coding studies. Other research cited by Dell'Acqua and colleagues found the reverse, and the 2026 education-gap experiment shows unassisted skill differences return when the tool is removed.[3][9]

Are these studies out of date now that models have improved?

Partly. Most tested tools from 2022–2023, and capability has moved since. The design lessons still hold: measure quality as well as speed, compare against a real control, and test tasks near the edge of the tool's competence. A newer model can move a task inside the frontier, but it does not remove the frontier.

Should my company expect a 15% gain?

No published number transfers directly. Run your own comparison on your own tasks, measuring finished work, checking time and errors as well as speed. Our first-week pilot guide is a practical starting point.

The next step

The studies are only as good as their measurement, and so is your own pilot. Next, read How to measure AI productivity without hiding review work. It turns the lessons above into a method you can run on your team's real tasks.

Sources and evidence note

Research reviewed and all sources accessed 16 October 2026. This is a selected, critical review of widely cited studies plus major 2026 publications. It is not a systematic review or meta-analysis, and we have not pooled estimates. Intervals marked "derived" in Figure 1 are AI Chronicle calculations of ±1.96 × the standard error each paper reports. The analyst example is hypothetical. No AI Chronicle test or measurement is claimed.

Related reading

• Who is accountable when an AI agent acts?
• AI for marketing: useful tasks and required checks