30% faster. But more productive?
What does the research say about AI at work? We reviewed 177 studies. Four things stand out:
- 01Task-level speed-ups are real and well measured.
- 02The effect shrinks at every level: 30% on a task ends up as tenths of a percent for the company.
- 03What gets lost is organisational: the cost of checking, what happens to the saved time, work that keeps its old shape.
- 04In the short run the company-level effect is zero or negative. The turn comes where the work is redesigned around the tool.
The same ratio, on ever larger systems
What we know today
Productivity is measured at three different levels
This is where most of the confusion comes from. At all three levels we look at the same ratio: useful output per unit of input. What changes is the size of the system. The research shows the effect getting smaller as you move up.
L1 Task
One person, one type of task, one sitting. How much gets done, at what quality, in how much time. This is where the 30% in the title comes from, and where the methodology is strongest: randomised experiments.
L2 Worker
The same ratio over weeks and months, across the whole job. Here it matters what the saved time turns into, and how much of it goes on checking the AI's work. Measured with telemetry and multi-month field experiments.
L3 Company
In the numerator, what we sell or deliver; in the denominator, the cost of wages, tools, training and verification. This is what shows up in the P&L. It is also where the evidence is thinnest, and what there is, is the weakest.
177 sources — and the quality of evidence falls with every level
At task level there are controlled experiments: groups doing the same work with and without AI. The worker level is covered mostly by usage data. At company level what remains is surveys, a few first econometric estimates, and a lot of vendor and consultancy material.
Which designs measure which level?
Reliable studies by level
Faster = more productive?
At task level, yes, and it is one of the best-documented effects in applied economics. Three things belong next to it. In real workplaces the effect is roughly half. Where the task falls outside the tool's capability, the effect reverses. And what people perceive is consistently larger than what can be measured.
Same technology, seven measurements
What we feel, and what the clock measures
Outside the capability frontier
Do we produce more value, or the same thing faster? Mostly the same thing faster. Sometimes better: with the newer reasoning models, the quality of legal analysis improved as well. And sometimes worse: wherever the task falls outside what the tool is good at. 030
Survey data, our own included, should be divided by two or three. In the METR experiment developers expected a 24% speed-up and ended up 19% slower. In the Norwegian sample the correlation between perceived and measured change was 0.17. 025 136
What happens to the organisation once the individual is faster?
This is where most companies stall. The individual time gain leaks out of the organisation at five points, and all five are organisational.
Task fit
The gain depends on whether the task sits inside the tool's capability. Nobody can see the boundary in advance; training made things worse on tasks outside it. 004
The price of checking
This much human rework was needed per “solved” agent task, a third of the task time. Half of the agent PRs that passed the tests could not be merged. 123 121
Where the time goes
Correlation between time saved and change in output across 5,512 Korean workers (self-reported). In a Singapore hospital, documentation time fell 15% and time per patient did not change: the minutes went into attention to the patient. 051 159
Bottleneck
If the constraint was in the process, speeding up the task only moves the queue elsewhere. This is the logic of the framework; the corpus has no direct measurement of it.
Missing complementary investment
This share of US workers receives employer AI training, while 39% use such tools. In European firm data, each percentage point of training investment adds about 5.9% to the AI effect. 153 104
Where an individual licence changes nothing further
Wide but shallow
There is agreement today that without redesigning the work there is no company-level effect. The uncomfortable fact: there is no causal study of that at all. What exists is vendor and consultancy case material, with eight- to thirty-three-fold multiples against estimated baselines. The electrification story, that factories had to be rebuilt around the dynamo and it took decades, is a good metaphor. As evidence it is thin. 150 151
What we gain — and what we can lose
The loss side is real and measurable too. It shows in three places: reliance, expertise and technical debt.
The labour-market position of juniors is an open question: job postings are tilting towards senior roles, while the heaviest AI-adopting firms hired 12% more entry-level staff. Both signals are low-confidence, so we treat it as contested. 141 133
Three questions decide whether the organisation got more productive, or just produces more
If all we have is a survey, the number overstates the effect roughly two to three times. The gap between perceived and measured gains is the best-proven fact in the field.
Does the gain have a name in the P&L? If the time went into leisure, waiting or backfill, we gained at task level and nothing at company level.
An individual licence on its own does not change the structure of the work. Redesign is the mandatory complement, and nobody has yet measured its effect causally.
Hungary: half the adoption
…and more disappointment, on the same foundations
If we are 20–30% faster, why doesn't it show in company results?
Because three numbers have to be multiplied. The first is the task-level speed-up, say 30%. The second is how much of the work AI touches at all. The third is the share of firms that use it deeply, with the work built around the tool. Everywhere it has been measured, that last number is between 2 and 16%. The product is tenths of a percent. That is the expected result; there is no mystery in it.
With the same logic, the OECD projects a contribution of 0.2–1.3 percentage points a year in the G7 over ten years, assuming adoption grows. Drag the sliders: in knowledge-intensive sectors 50–80% of the work is affected, in manufacturing 10–30%. 107
The first causal company-level estimates
177 studies, filterable
For every study: who was measured, with what, how, what effect was found, and what its weakness is. Confidence is the corpus's own grading: randomised and field experiments > usage data > official statistics > surveys > vendor material.
How this was made
The corpus holds 177 sources, collected in two waves (August 2026). Every study was extracted with the same schema: population, type of work, technology, methodology, measure, effect size, quality impact, subgroups, limitations, confidence. The levels (L1/L2/L3) are never mixed: a task-level randomised experiment is not evidence of company productivity. Negative and contradicting evidence carries equal weight. Developer research (Anthropic, OpenAI, Microsoft, Google) is treated as primary but interested-party evidence.
Every number on this page was checked against its study note on 3 September 2026. Where the note grades a study medium or low confidence, the text says so. If a number looks wrong, or a study is missing, write to us.
Ákos Janza · n20k · info@n20k.hu · n20k.hu
This page is an independent review of publicly available research. Every study referred to is identified in the list above by its title and year; the findings, figures and effect sizes belong to their original authors and publishers, and we claim no authorship of them. Where we quote, we quote briefly and with attribution, as permitted for purposes of criticism and review. Everything else on this page — the selection, the three-level framework, the confidence grading, the charts and the wording — is our own reading of that research, offered in good faith and without warranty of completeness or accuracy. It is general information, not advice; decisions should rest on the original studies and on professional counsel. Access is free of charge. Our own text and charts may be reused under the Creative Commons Attribution 4.0 licence (CC BY 4.0): please cite this page as the source, e.g. “n20k, Where is productivity? (2026), n20k.hu/research/where-is-productivity/”. This licence does not extend to the underlying studies, which remain subject to their own terms. If you are an author or publisher and believe a summary misrepresents your work, or you would like a reference removed, write to info@n20k.hu and we will correct it promptly.