Your staff are more productive with AI. Only one in ten says the business is feeling it.

A survey of 81,000 AI users found the productivity gains are real. It also found only about one in ten said the benefit reached their employer. Why the value leaks out before it gets to the business, and what it takes to keep it.

A managing director told me last month that her team had gone all in on AI. Everyone's using it, everyone says they're faster, the mood is good. Then she said the quiet part out loud: the P&L looks exactly the same as it did a year ago.

She isn't imagining the gains, and she isn't doing it wrong. She has run into one of the most consistent findings in this year's research, and it is worth understanding before you sign off on another round of licences.

The gains are real. They are just not landing where you think.

Earlier this year Anthropic published the largest first-person account we have of what AI is doing to people's working lives: a survey of 81,000 of its own users, asked in their own words how their work had changed. A wave of analysis over the past week has gone back to it, and one finding deserves far more attention than it has had.

The productivity signal was not subtle. On a one-to-seven scale, the average self-rated gain was 5.1. Only about three per cent reported no effect or a negative one. Then the awkward part: when those same people were asked who was capturing the benefit, most pointed at themselves. Faster tasks, wider scope, a few hours back in the day. Only around one in ten said their employer or their clients were getting more out of them.

If that were a single survey you could wave it away. It is not. A round-up of this year's studies lines them all up pointing the same way. Gallup found 65 per cent of workers in AI-using organisations saw a personal productivity lift, while only 12 per cent strongly agreed AI had changed how their organisation works. A separate study of roughly 6,000 executives across the US, UK, Germany and Australia found 89 per cent reported no measurable impact on productivity over three years.

Read those together. People feel more productive. Organisations cannot find the productivity. Both things are true at the same time.

Why the value leaks out

Here is the part that matters if you are running a mid-market business. The value does not reach the organisation by default. It reaches the individual by default, and it stays there unless someone deliberately decides otherwise.

Think about what actually happens when a task that took an hour now takes twenty minutes. The forty minutes does not appear as a line in your accounts. It gets absorbed, into a slightly bigger scope, a side project, a job that quietly expands to fill the time it used to take. None of that is laziness. It is the natural result of handing people more capacity inside roles and targets that were designed for the old pace.

Nothing breaks, so nothing prompts a change. The work was built for narrower roles and a slower clock, and a productivity gain nobody captures at the organisational level is, to the business, indistinguishable from no gain at all.

What capturing it actually looks like

This is unglamorous, and it has almost nothing to do with the tools. Pick one process. Before you point AI at it, write down what it costs you today at the task level, in plain terms, so you have something honest to compare against later. Then make the decision most teams skip entirely: where does the recovered capacity go?

• Into more of the same output, if you have demand sitting there waiting for it.

• Into better quality or shorter turnaround, if that is what wins you the next client.

• Into moving people onto work the business values more than the task you just automated.

Pick one on purpose. Then change how that team is measured so the gain has somewhere to show up. If your metrics still describe the job as it was before AI touched it, the surplus will keep leaking to wherever it leaked last time.

The bit nobody puts in the deck

The same survey found something leaders should sit with. The people using AI the most were also the most worried about being replaced by it. The anxiety rose with exposure, and the heaviest users named the fear roughly three times as often as the lightest. That is not fragility. It is a rational read of their own situation, and it tends to predict who you lose.

If your most capable people quietly conclude that the productivity they created is being used to argue them out of a job, they will take it somewhere it is valued. Capturing the gain and keeping the people who produce it are the same project, not competing ones.

The work is yours, not the tool's

The clean version of this story is the one you hear at every vendor briefing: buy the tools, capture the margin. It does not survive contact with what 81,000 users actually said. The gains are large and they are real. They also default to the individual, they show up as wider scope more than raw speed, and they arrive wrapped in a worry that grows with your best people.

The tools do not do the next part for you. Deciding where the recovered capacity goes, redesigning the roles around it, measuring the thing you actually want, keeping the people who made it possible. That is the work, and it always was. AI has just made skipping it more expensive.