AI productivity statistics: what the research actually shows
AI productivity statistics from nine studies and official reports, compared side by side: where AI made workers faster, where it did not, and the real range.
Klevere AI Team
AI Strategy Team
Search for AI productivity statistics and you will find numbers ranging from a 55% speed-up to a 19% slowdown, sometimes in the same week of headlines. Both figures are real. A controlled experiment found developers using GitHub Copilot finished a coding task 55.8% faster, while a later randomised trial found experienced open-source developers took 19% longer when they were allowed to use AI tools.
This post puts the most credible published figures side by side, explains why they disagree, and gives you a realistic range to plan with. Every number links to the study or official report it came from, and none of them are Klevere's own data.
Quick answer
Controlled studies show generative AI makes workers 14% to 56% faster on well-defined tasks, with the largest gains for less experienced staff. Economy-wide estimates are far smaller: roughly 1% aggregate productivity in the US so far, and no measurable change in earnings or hours in Denmark. Gains depend heavily on task fit, training and workflow redesign.
AI productivity statistics side by side
The list below summarises nine of the most cited, methodologically transparent sources. Each entry gives the source, the headline figure, the year and the sample or scope, so you can judge how far it generalises.
What the experiments show: big gains on the right tasks
The strongest evidence for AI productivity gains comes from field and lab experiments where researchers can compare people with and without AI on the same work. The customer support study is the best known. When a large software company rolled out a generative AI assistant to its support team in stages, agents with access resolved 14% more issues per hour. The finding that matters most for managers is who gained: novice and low-skilled agents improved by 34%, while the most experienced and skilled agents saw minimal impact.
The consulting experiment tells a similar story with an important twist. On a creative product task that GPT-4 handled well, consultants using AI completed 12% more work, 25% faster and at 40% higher quality. Those below the median skill level gained the most. The researchers call the boundary between tasks AI handles well and tasks it does not the jagged frontier, because it is uneven and hard to see from the outside.
Software development produced the most dramatic early number. In the GitHub Copilot experiment, the treated group finished the task 55.8% faster. The task was narrow and well specified, writing a small web server from scratch, which is close to ideal conditions for code completion tools.
Three patterns repeat across these experiments. First, gains concentrate on tasks that are well defined, text heavy and easy to check. Second, less experienced workers usually benefit more, because the AI effectively packages the knowledge of top performers. Third, quality as well as speed can improve, but only when the task sits inside the model's capabilities.
Where AI productivity gains did not appear
A data post that only shows wins would mislead you, so the negative and neutral findings deserve equal weight. The sharpest counterexample is METR's randomised trial. Sixteen experienced developers worked on 246 real issues in large open-source projects they knew well, and each issue was randomly assigned to allow or forbid AI. With AI allowed, they took 19% longer. They had predicted a 24% speed-up, and even after the study they believed AI had made them about 20% faster.
That gap between perceived and measured productivity is the most useful single lesson in this whole body of research. Self-reported time savings can be optimistic. METR's follow-up study found estimated speed-ups for most groups, and the authors think developers are now likely to be faster with AI than in early 2025, but they also describe their own data as only very weak evidence because of selection effects and the difficulty of timing work when people run several AI agents at once.
The consulting study shows a second kind of failure. On a business analysis task deliberately designed to sit outside GPT-4's capabilities, consultants using AI were 19 percentage points less likely to get the right answer than colleagues without it. People trusted plausible output on a problem where the model was wrong.
Government pilots add a third. The UK Department for Business and Trade gave 1,000 staff Microsoft 365 Copilot and compared them with a control group. It found small time savings, mostly on writing, yet no evidence that those savings led to improved productivity. Time saved on drafting an email does not automatically become more output if nobody changes what people do with the freed time.
Economy-wide evidence: why aggregate gains look small
Once you move from controlled tasks to whole labour markets, the numbers shrink sharply. The St. Louis Fed's nationally representative surveys found that workers who used generative AI saved 5.4% of their work hours, but because many people did not use it at all, that equalled 1.4% of total hours across the workforce. The authors translate this into roughly a 1.1% increase in aggregate productivity, a model-based estimate rather than a direct measurement.
Denmark offers the most rigorous economy-wide test so far. By linking adoption surveys to national administrative records, Humlum and Vestergaard found precise null effects on earnings and recorded hours, ruling out effects larger than 2% two years after adoption. The authors' interpretation is that time saved is partly absorbed by new work created around the tools, so gains have not yet shown up in pay or hours.
The UK government's cross-government trial sits between the two. More than 20,000 civil servants reported saving an average of 26 minutes a day over three months, almost two weeks a year when scaled up. That is a meaningful figure, but it is self-reported and comes from a single tool rollout, so treat it as an upper estimate of time saved rather than proof of output gained.
Why the numbers disagree
The spread between a 56% speed-up and a 19% slowdown is not a sign that the research is broken. The studies measure different things.
The realistic range to plan with
Pulling these together gives a defensible set of planning figures for most businesses:
How we compiled this
We selected studies and official reports that publish their method, sample size and dates: peer-reviewed or working-paper experiments, central bank surveys and UK government evaluations, with a Danish national dataset representing the EU. We excluded vendor marketing claims without a stated method. Figures are quoted as each source reports them, and we opened every source page to confirm the number before including it. Where a study has since been updated, as with METR, both results are shown.
What this means for your business
The research points to a clear conclusion: AI productivity is a property of the task and the workflow, not of the tool. Buying licences and waiting for gains is the approach most likely to produce the small or null results seen in the economy-wide data. Picking specific, high-volume tasks, redesigning how work flows around them, and measuring output rather than feelings is how organisations get closer to the experimental figures.
Start by measuring. Before any rollout, record baseline volumes, handling times and error rates for the tasks you plan to change, then compare after. Our guide to ai roi walks through how to set up that measurement so you are not relying on self-reported time savings. If you want context on how many of your staff are likely already using these tools, our generative ai usage statistics post covers adoption by role and country.
Then target the task types where the evidence is strongest. Customer support is the clearest case, which is why the figures in our ai customer service statistics round-up tend to be more consistent than those for general knowledge work. Finally, train people on where the jagged frontier sits in your own work, so they know when to trust AI output and when to check it.
How Klevere approaches it
Klevere has deployed more than 500 AI agents across over 50 projects in 12 industries, and the pattern in this research matches what we see in practice. The productivity comes from narrowing the job: an agent that handles one well-defined workflow end to end, with clear inputs, clear outputs and a human review step where errors would be costly.
We start every engagement by mapping tasks rather than roles, identifying which sit comfortably inside current model capabilities and which do not. We set a measured baseline before anything is built, design the human checkpoints around the tasks where the consulting study showed AI can confidently mislead, and report on output and quality rather than estimated minutes saved. For teams that need to build internal capability alongside the agents, our ai consulting and training work focuses on exactly that skill: knowing where AI helps and where it does not.
Frequently asked questions
How much does AI increase productivity?
Controlled studies find generative AI makes workers between 14% and 56% faster on suitable tasks, such as customer support, drafting and well-specified coding. Across whole economies the measured effect is much smaller, around 1% in US estimates and no detectable change in Danish earnings or hours so far. The gap reflects task fit, adoption rates and whether saved time is redeployed.
Does AI make experienced workers more productive?
Less than it helps newer staff. In the customer support study, novices gained 34% while the most experienced agents saw minimal impact, and a 2025 trial found experienced developers working on familiar code were 19% slower with AI. Experts benefit most when AI takes over routine sub-tasks rather than the judgement-heavy core of their work.
Is the productivity gain from AI real or just self-reported?
Both kinds of evidence exist. Randomised experiments and administrative data have measured real gains on specific tasks, but self-reported figures tend to run higher. In one trial, developers believed AI had sped them up by 20% when measured time showed a slowdown. Treat survey-based time savings as an upper estimate and measure output directly where you can.
Which jobs see the biggest productivity gains from AI?
The strongest evidence is for customer support, writing and drafting, and narrow programming tasks. These share three traits: high volume, mostly text-based work, and outputs that are easy to check. Tasks that need deep context, novel analysis or careful judgement show smaller gains, and in some experiments AI users made more mistakes on problems outside the model's capabilities.
Why is AI not showing up in productivity statistics yet?
Several reasons combine. Many workers still use AI rarely or not at all, so average time savings across the whole workforce are small. Time saved is often absorbed by new tasks or not redeployed. Organisations also take time to redesign processes around new technology, a pattern economists have seen with earlier general-purpose technologies such as electricity and computers.
Start with the tasks, not the tool
The headline AI productivity statistics are not a promise you can bank on, but they are a reliable guide to where gains are likely and where they are not. If you want to know which tasks in your business sit inside the frontier and what a realistic gain looks like for each, book a free AI audit and we will map it with you.
Sources
- NBER: Generative AI at Work (Brynjolfsson, Li and Raymond)
- Harvard Business School Working Knowledge: Humans vs machines, untangling the tasks AI can and can't handle
- arXiv: The Impact of AI on Developer Productivity, Evidence from GitHub Copilot
- METR: Measuring the impact of early-2025 AI on experienced open-source developer productivity
- METR: Uplift study update, February 2026
- Federal Reserve Bank of St. Louis: The Impact of Generative AI on Work Productivity
- GOV.UK: Landmark government trial shows AI could save civil servants nearly 2 weeks a year
- GOV.UK: Microsoft 365 Copilot pilot, DBT evaluation report
- Becker Friedman Institute: Large Language Models, Small Labor Market Effects