Why AI coding tools haven't made your team faster (yet).

Everyone uses it, and everyone feels faster

Google's 2025 DORA research found that 90% of software professionals now use AI at work, spending a median of about two hours a day with it. Over 80% say it has made them more productive. Other large surveys put developer usage above 90% too.

Feeling faster and being faster are different things, though. Only about a quarter of developers in the DORA research said they trust AI output a lot or a great deal.

The trial that measured it

In 2025 the research group METR ran a randomised trial with sixteen experienced open-source developers working on 246 real tasks in codebases they knew well. With AI tools, tasks took 19% longer. Before starting, the developers expected AI to make them about 20% faster, and afterwards they still believed it had.

That's a small study of very experienced people on their own mature projects, and METR's 2026 follow-up was inconclusive in both directions. But it shows something every engineering leader should take seriously: a developer's sense of speed is not a measurement.

Where the time actually goes

Writing and testing code is usually estimated at somewhere between a quarter and a third of the software lifecycle. Specifying work, waiting on decisions, reviewing, integrating, deploying and fixing take the rest. One large monitoring study found the median developer spends under an hour a day actually coding.

So even if AI doubled coding speed, the whole process could only improve by a fraction of that. The other stages decide how fast the team ships.

The review bottleneck

Faros AI studied more than 10,000 developers across 1,255 teams. On teams with high AI adoption, developers completed more tasks and merged nearly twice as many pull requests. But:

Source: Faros AI, reported in the evidence review linked below.

The work didn't disappear. It moved from the people writing code to the people reviewing it, often the most senior and busiest developers on the team.

AI amplifies what's already there

DORA's 2024 research found that more AI adoption was linked to slightly lower delivery throughput and noticeably lower stability. Its 2025 research found the throughput link had turned positive, but stability problems remained. DORA's conclusion is that AI acts as a mirror and a multiplier: cohesive teams with good practices get faster, and fragmented ones see their weaknesses magnified.

What the teams that do get faster have in common

From the research and from building production systems with coding agents every day, the pattern is consistent:

1. They write things down for the agent. Context files describing the architecture, conventions and commands, kept in the repository and updated like code.
2. They specify before they generate. A short, testable task description beats a vague prompt every time.
3. They let tests do the first review. Architecture tests, characterisation tests around legacy code and property-based tests catch bad output before a person has to.
4. They keep changes small. Agents can produce huge diffs. Teams that win cap the size and split the work.
5. They point agents at the chores first. Test backfill, dependency upgrades and migrations are ideal agent work and free senior people for design.
6. They measure delivery, not feelings. Lead time, review time and change failure rate from the pipeline, before and after.

None of this needs a new tool. It needs the process around the tools to change, on your own codebase. That's what the Agentic Engineering programme does.