In 2025, a research group called METR ran a careful experiment on AI productivity. Experienced software developers worked on real tasks, some with AI tools and some without. With AI, they took 19% longer. Afterwards, the same developers estimated that AI had made them about 20% faster.

That gap between how productive AI feels and how productive it is should worry every business owner whose team has started using it. This article explains where the illusion comes from, what it’s costing you, and why governance, the unglamorous business of rules, review and measurement, is what turns AI use into real gains.

More output is not the same as better output

AI makes producing things extremely cheap. A report, a proposal, a 1,500-word blog post, a polished email: each can be generated in seconds. So people generate more of them.

The trouble is that somebody still has to read it, check it, fix it and act on it. Researchers at BetterUp Labs and the Stanford Social Media Lab have a name for AI output that looks finished but isn’t: workslop. Their survey of 1,150 US employees, written up in Harvard Business Review in September 2025, found that 40% had received workslop in the previous month. Recipients estimated it made up about 15% of what landed on their desks, and each incident took nearly 2 hours to deal with.

So the person who sent it saved half an hour, and the person who received it lost two. On a team spreadsheet that looks like a productivity gain for the sender. For the business, it’s a loss.

The busy fool problem, now with better technology

Every business has known a busy fool: someone who’s always flat out and never quite moves anything forward. AI can turn a whole team into busy fools, because it removes the friction that used to slow down low-value work.

Back to the METR study. These were experienced developers working on code they knew well. They felt faster because the tool was doing visible work: suggesting, generating, explaining. The time went on reviewing suggestions, correcting them and waiting. None of that registered as effort in the way typing does.

The study has its limits. It was 16 developers, using early-2025 tools, on large codebases they knew intimately. METR ran a follow-up with newer tools in early 2026, and the results pointed towards a modest speed-up, though METR itself described them as too uncertain to rely on. The tools are improving. The lesson about perception still travels well: people are poor judges of their own AI productivity, and “it feels faster” is not a measurement.

What are you actually risking?

When AI use isn’t governed, the costs are easy to miss because they show up somewhere other than where the AI was used.

  • Quality drift. Client documents get longer, blander and more alike. Errors slip through because everyone assumes someone else checked.
  • Rework. Colleagues and managers spend their time fixing output that looked done. The workslop research found 42% of recipients saw the sender as less trustworthy afterwards.
  • Data exposure. Without approved tools, staff use personal accounts, and client information ends up in places you never agreed to. Our shadow AI article covers this in detail.
  • Compliance gaps. Under the EU AI Act, every business that uses AI has to take measures to support its staff’s AI literacy. Ungoverned use makes that very hard to show.
  • Client trust. Clients can usually tell when they’re reading unedited AI output. They notice it most in exactly the documents where you wanted to impress them.

The governance gap is a leadership problem

It’s tempting to treat this as a staff problem. Look closer and your staff are doing exactly what you’d expect people to do with a powerful tool and no guidance.

The data backs that up. The AI Economy Ireland 2026 report from Trinity College Dublin and Microsoft Ireland found that only 44% of organisations have a formal AI policy. Those that do were 10 times more likely to report major productivity gains: 30% against 3%. Internationally, BCG’s AI at Work 2025 survey of more than 10,600 workers found that only a quarter of frontline employees said their leaders gave them enough guidance on AI.

That policy figure is a correlation, and I wouldn’t claim the policy document itself does the work. What it shows is that businesses where leaders made deliberate decisions about AI are the ones getting results. Deciding is the leader’s job. Nobody else in a small business can do it.

What governance actually looks like

For a small business, governance fits on a couple of pages and takes a few hours to set up.

  • A one-page AI policy. Which tools are approved, what data never goes into them, and who to ask when you’re unsure.
  • A review rule. Anything leaving the building that AI helped create gets checked by a named person before it goes. The person who sends it owns it.
  • An AI tool register. A list of every tool in use, who uses it and what data it touches.
  • Practical training. People trained on their actual tasks, including when not to use AI.
  • Measurement. For each process where AI is used, a before-and-after figure for the time the whole job takes, including review and rework.

Each item closes off one of the risks above. The oversight layer of the Trust Stack sets out how they fit together.

The real productivity gain comes after governance

Once the rules are clear, something useful happens. People stop wondering whether they’re allowed to use AI and start getting good at it. Reviews get faster because reviewers know what to look for. And because you’re measuring the whole job, including the checking, you can see which uses genuinely save time and which only move the work onto someone else’s desk.

That’s where the real AI productivity gain lives: in fewer, better uses of AI, applied consistently, with the time saved measured honestly. It’s less exciting than “everyone uses AI for everything”. It’s also the version that shows up in your margins.

What to do this week

  1. Ask your team where AI saves them time. Then ask where they spend time fixing AI output, theirs or anyone else’s. The second answer is usually more revealing.
  2. Pick one process and measure it properly. Time the whole job from start to finished, checked output, with and without AI.
  3. Write the review rule. One sentence: “Anything AI helped create is checked by a named person before it leaves the business.”
  4. Book an hour to draft your one-page policy. You’ll have a usable first version by the end of it.

If you’d like help with the leadership side, our AI Leadership course is built for owners and senior managers who need to set the direction and the rules. You can also request our free AI Readiness Scorecard for a review of where you stand.