Your team feels faster with AI. Has anyone timed it?

Ask a team that adopted AI tools six months ago whether they are faster now. Most will say yes without pausing. Ask how they know, and the room gets quieter.

The best evidence we have on that gap came from METR, a research group that ran a randomised trial in the first half of 2025. Sixteen experienced open-source developers worked through 246 real tasks in code they knew well. Some tasks allowed AI tools, some did not. With AI, they took 19% longer. Asked afterwards, they estimated AI had made them about 20% faster.

A 39-point gap between what people felt and what the clock said.

METR was careful about its own limits. Sixteen people is a small group. They were developers working on large codebases they had maintained for years, which is a hard setting for an assistant that knows none of that history. The researchers warned against stretching the result to everyone. We would add one more caution: your accounts team is not an open-source maintainer.

Still, the shape of the finding matches what owners describe to us in very different businesses. The work feels lighter. The week does not get shorter.

Where the time goes

Atlassian's 2025 developer experience survey found people saving around ten hours a week with AI, while half of them lost ten or more hours a week to friction elsewhere in the organisation. Finding information was the biggest time sink.

Put those together and the illusion makes sense. AI removes the visible effort: the blank page, the first draft, the typing. It adds less visible effort: explaining context to the tool, checking what it produced, fixing the parts that were confidently wrong, and hunting for the facts it did not have. The first kind of effort feels like work. The second kind feels like admin, so nobody counts it.

Which part a company memory removes

Some of that hidden effort is re-explaining. Every new session starts blank. The person has to tell the tool, again, who the client is, what the house style is, which product was discontinued, what the approval rule is for discounts. That is pure overhead, and it repeats daily across every person using AI.

A written company memory takes most of that away. When the rules, verified facts and past corrections sit in shared files that the AI reads before it starts, the setup is done once and kept current, instead of done by everyone every morning. We run our own work this way, mostly so that nobody has to retype the house rules at the start of every task.

It does not remove review time. If anything, a firm should expect review to stay roughly constant, because a person still has to check output before it goes to a client or into the books. It also does not fix over-trust. A tool that sounds sure of itself gets checked less, memory or no memory. That is a habit problem, and approval gates exist because habits slip.

There is also a reason the feeling is so convincing. Writing a first draft is the part of most tasks that people dislike. Remove it and the task feels lighter, even if checking now takes longer. Owners should expect their teams to report the feeling honestly and still be wrong about the clock. That is no criticism of anyone. It is why you measure.

A quick way to find out

Pick one recurring task that someone on your team now does with AI. A weekly client update, a quote, a reconciliation note. For the next five times it is done, record three numbers on a sticky note:

1. Minutes from start to a first draft.
2. Minutes spent checking and fixing it.
3. Minutes spent giving the tool context it should already have had.

If you have an old timesheet or a memory of how long the task took before, compare. If you do not, do two runs without AI. You will learn whether the gain is real, and the third column will tell you how much of the remaining cost is a memory problem.

Our operations drag check helps put numbers on the friction around a task like this. Book 20 minutes at pgco.world/book.