Does AI really make developers faster? What the research says
From 56% faster to 19% slower. The randomised studies contradict each other. Here is how to read them, and what they mean for your business.
A vendor promises that AI will make your team twice as fast. A colleague reads that developers actually get slower. Who is right?
Both, to a degree. It depends on what you measure, with whom and on which tasks. These are the most important studies, with their limitations.
The studies
55.8% faster, on one well-defined task
In a controlled experiment from 2023, developers had to write a simple web server in JavaScript. The group using GitHub Copilot finished 55.8% faster. It was a single new, clearly described task. The researchers worked for GitHub and Microsoft. Source: Peng et al., 2023
Around 21% faster at Google
Google had 96 of its own engineers complete a task, with and without AI assistance. Those using AI took around 21% less time. The researchers themselves point to a wide margin of error. Source: Paradis et al., 2024
26% more completed tasks in practice
The largest field experiment followed 4,867 developers at Microsoft, Accenture and a Fortune 100 company. Those given access to AI assistance completed 26% more tasks. According to the researchers, less experienced developers gained the most. The study measures how much work gets finished, not its quality. Source: Cui et al., Management Science, 2026
19% slower for experienced developers
In 2025, the research institute METR had sixteen experienced developers complete 246 real tasks in large projects they knew well. With AI assistance, they took 19% longer. Strikingly, beforehand they expected to be 24% faster, and afterwards they believed they had been 20% faster. Source: METR, 2025
METR now considers that result outdated. According to METR itself, a follow-up measurement in 2026 did not yet produce a reliable new figure, partly because 30 to 50% of the developers refused to do some tasks without AI. Source: METR, 2026
What you learn from this
- The task makes the difference. Well-defined, clearly described tasks deliver the biggest gains. Complex work in a large codebase you know inside out delivers far less.
- A feeling is not a measurement. People overestimate the time AI saves them. A later METR survey found an overestimate of 40 percentage points on average. Source: METR, 2026
- Quality needs checking. In a Veracode test, 44% of programming tasks produced AI code with a risky security flaw. Source: Veracode, 2026
What this means for your business
AI does not deliver a fixed amount of time saved. It saves time when you choose the right tasks, set clear rules, have everything reviewed and measure the result.
That is exactly how we work. We pick the tasks where agents are strong, a person approves everything, and every report shows the hours saved and the review time. You will find all the figures and charts on our evidence and sources page.