What does the research actually say about productivity gains from AI?

You have seen the headlines: AI makes everyone ten times more productive. The largest field studies tell a different and more interesting story: 25 to 56% faster task completion among trained people, but 19% slower work among experienced people without training (METR 2025). The difference lies not in the tools but in how people learn to use them.

The largest experiment: 758 consultants at BCG

In 2023, researchers from Harvard, Wharton and MIT ran a field experiment with 758 consultants at BCG, one of the largest consulting firms in the world (Harvard/BCG 2023). Half were given access to AI, the other half were not. The tasks were real consulting tasks, not made-up tests.

The results were decisive. The group with AI completed 12.2% more tasks, was 25.1% faster at each of them and delivered work that independent assessors rated over 40% higher in quality.

The most striking result, though, was this: those who scored lower in the skills assessment before the experiment improved by 43%, far more than those who were already the strongest. AI lifts a broad group, not just the tech enthusiasts. For a manager, that means the greatest gain lies in training the whole department, not just those who are already keen.

More studies tell the same story

The Harvard/BCG study does not stand alone. Two other large studies support the same picture:

Harvard/BCG 2023. who: 758 consultants · result: 12.2% more tasks, 25.1% faster, +40% quality

GitHub Copilot, randomised trial. who: developers · result: 55.8% faster at completing a defined task

NBER, customer service. who: thousands of customer service agents · result: +14% productivity on average, +34% among new hires

METR 2025. who: experienced developers, without targeted training · result: 19% slower with AI

Two things stand out in this table. First, the same pattern repeats across different jobs: the biggest gains are among those with the least experience. In the NBER study, new hires improved by 34% while the average was 14%.

Second, the bottom row is a warning, not a typo. It is the one that matters most for you.

The METR warning: AI can make people slower

In the summer of 2025, the research institute METR published a finding that surprised most people (METR 2025). Experienced developers working on their own projects with AI tools took 19% longer to complete the tasks than without the tools. What is more, they themselves believed they had been about 20% faster.

People felt more productive while they were actually slower. That is the most dangerous position a workplace can be in, because nobody sees the problem.

The explanation is not that the tools are useless. The participants had not received targeted training in using them in their own context: when to trust a result, when to verify, and when it is quicker to skip the tool altogether. This chimes with MIT NANDA, which reports that 95% of corporate AI projects deliver no measurable results. People get the tools but not the training.

What does this mean for your company?

Read together, the studies paint a clear picture. Three conclusions stand out:

• The gains are real but tied to training. 25 to 56% faster task completion is a realistic benchmark for trained people. Promises of tenfold productivity gains for everyone do not hold up to scrutiny.
• Train the broad group, not just the front-runners. The biggest improvement, +43% at Harvard/BCG and +34% at NBER, was among those who started at the bottom. That is precisely the group that asks for training least often.
• Measure it yourself. METR showed that people's perception of their own productivity can be off by nearly 40 percentage points. Do not rely on gut feeling, yours or your staff's. Measure a baseline before you start and measure again afterwards.

In Iceland, 80% of professionals already use AI at work but only 34% have received training from their employer (Viska 2025). If the studies above are right, a large share of Icelandic workplaces are in the METR scenario without knowing it.