Does AI-assisted software development make teams faster?
Sometimes, and by less than people feel. In METR’s 2025 trial, experienced developers took 19% longer with AI tools while believing they were 20% faster. METR’s 2026 follow-up points to a speed-up, but METR itself calls that data unreliable.
In a randomized trial of 16 experienced open-source developers on 246 real tasks, allowing AI tools made tasks take 19% longer. Before the study they expected a 24% speed-up, and afterwards they still believed AI had made them 20% faster.
What the studies on AI-assisted software development measured
- METR, 2025, randomized trial. 16 developers worked on repositories they knew well, averaging over a million lines of code, mostly with Cursor Pro and Claude 3.5 and 3.7 Sonnet. Result: 19% slower. Limits, in METR’s words: a snapshot of early-2025 tools in one setting, a small group, and it may not apply to less experienced developers or unfamiliar code. METR now marks the page as out of date.
- METR, February 2026 update. 57 developers and more than 800 tasks, with tools like Claude Code and Codex in use. The estimate was -18% for returning developers and -4% for new ones, on the same time measure where 2025 showed +19%, so this time AI tasks looked faster. Both ranges include no change at all, and METR says many developers refused to work without AI or held back tasks, so the data is “an unreliable signal”.
- DORA, 2025. A survey of nearly 5,000 technology professionals. 90% use AI at work, more than 80% believe it raised their productivity, and 30% have little or no trust in AI-generated code. AI use went with higher delivery throughput and lower delivery stability. This is a survey: it shows what people report and what moves together, not cause and effect.
- Stack Overflow, 2026. Of 14,304 developers who answered the trust question, 48% trust AI output when they can easily check it, and 6.6% trust it with important decisions. Of 13,162, 9.6% use AI answers as they are. The people who answer this survey choose to, so treat it as a view of Stack Overflow’s audience.
What changes when a senior engineer works with AI
- Boilerplate, tests for code that already exists, and small refactors go faster. In the 2026 Stack Overflow survey, debugging and refactoring were among the tasks developers found AI most useful for.
- Reading an unfamiliar library or codebase gets quicker, because you can ask it questions, including the ones you would not ask a colleague twice.
- More code gets written per day, which means more code to review per day. The reviewer did not get faster.
What does not change
- Someone has to decide what to build, and what not to. The model will happily build both.
- Every change still needs review by a person who can explain what it does.
- Security checks, data rules and payments still need testing by hand, because AI code often skips the checks nobody asked for.
- A team with weak habits gets worse faster. DORA’s summary of its own 2025 report: “AI doesn’t fix a team; it amplifies what’s already there.”
When AI slows the work down
METR’s 2025 setting is a good guide: large, mature codebases that the developer already knows well, with a high bar for quality. There, writing a careful prompt and checking the result can take longer than writing the change. AI also costs time when its output is almost right, because a subtle bug in code you didn’t write is slow to find. The code looks confident, which does not help.
How I work with it
Claude Code is one of the tools I use in my work. The rule I hold myself to is the one Simon Willison wrote down: don’t commit code you couldn’t explain to someone else. It sounds easy until the diff is long and it is late.
I sell this, so read it with that in mind. On the Full plan, I join your team, your standups and your repo for $4,995 a month: a flat monthly price, one task at a time, worked on every working day, pause or cancel anytime. If your team is already shipping well without AI, there is no reason to hire anyone for the AI part alone.
An example: the Reddit reader I built with Claude Code, and the bug a test caught
Read next
Sources
- METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, 10 Jul 2025
- METR, We are changing our developer productivity experiment design, 24 Feb 2026
- Google Cloud, Announcing the 2025 DORA Report, 23 Sep 2025
- Stack Overflow Developer Survey 2026, AI data, 6 Oct 2026
- Simon Willison, Not all AI-assisted programming is vibe coding, 19 Mar 2025