Promising a human in the loop has become the standard way to make AI adoption sound responsible. In most small businesses that human is the owner, reading every draft, quote and customer reply the agents produce before it goes out. We call that oversight, and it’s a real job, one that comes down to catching whatever the machine got wrong. Reviewer is the job AI is most likely to shrink through its own success, so an owner’s real decision is what to stop checking and what to do with the attention that frees up.

Sangeet Paul Choudary argues in Harvard Business Review this month that companies are building exactly this job into their organisations. As the systems improve, the need for the checking falls away, and the people doing it gradually lose the skill and attention they’d need to catch the errors that remain.

We’ve already seen how little protection the title gives. In September 2025, xAI let go of around 500 of the generalist tutors who had been checking and labelling Grok’s output, saying it wanted specialists instead. That was one company’s decision, but the logic travels. The corrections we make most often in a small business are the ones we’ll eventually write into the agent’s instructions, so the more conscientiously we review, the faster the reviewing job shrinks.

Bainbridge’s irony

The other half of the problem was described more than forty years ago. In 1983 the psychologist Lisanne Bainbridge published a short paper called Ironies of Automation, written about control rooms and cockpits. Her central irony still holds. The more reliable the automated system, the less often the person watching it has to step in, and so the worse they get at stepping in when it matters. People are bad at watching something that’s almost always right, however conscientious the watcher.

Language models make this harder still. A failing plant usually has a gauge or an alarm somewhere that shows it, but a failing model produces a paragraph every bit as fluent and confident as the ones it got right. Anyone reviewing that output ends up mostly grading whether it reads well, and after a few hundred clean approvals the eye stops looking for anything more. I wrote recently about the queue of agents waiting on us, after Anthropic found its own agents spending four-fifths of each session idle while a person caught up, and approvals like these make up a good part of any such queue.

When the reviewer is the owner

Choudary’s remedy is to redesign around the new problems AI makes it possible to solve and to move people towards higher-value work, and for a company with thousands of staff that’s sensible advice. In a five-person firm the attention at stake is mostly the owner’s. We’re the ones approving every draft and quote, and calling it oversight.

I ran Vertical Leap on Michael Gerber’s principles for the best part of seventeen years. His technician, in The E-Myth, carries on doing the work instead of building the system that does it, and ends up working in the business rather than on it. Spending the morning approving agent output is the same trap in a new costume, and it feels like management because we’re signing things off.

Some of those approvals are doing real work, and some of them are there because they make us feel involved.

Control and comfort

Take a quote. If the agent has priced it inside your margin rules and your standard terms, your reading it adds reassurance and nothing else. If it offers an unusual discount or promises a delivery date you haven’t confirmed, that’s the one that should reach you. The first is a comfort approval and the second is a control approval, one that guards against a cost you can name. Go through every approval you gave an agent last week and sort them the same way, and if you’re like most of us, you’ll find more comfort than you expected.

The control approvals belong in the rules the agent works under, and where possible in checks that sit outside it, such as a margin calculation that stops a quote being sent, because an instruction in a prompt can still be ignored. The formal version of this is a bank of test questions with known answers, re-run whenever anything changes. A small firm’s version can be twenty questions you already know the right answers to, run again each month. The comfort approvals we can simply let go, which is harder than it sounds when they’ve become the way we keep in touch with our own business.

This sits awkwardly with something I argued a week ago. In Process maps show the flow; capability maps show the judgement I left pricing review with a person, and I’d still do that, with a condition I didn’t stress enough at the time. A review step is only worth having if the reviewer is weighing the decision, and Bainbridge tells us that gets harder the longer the system runs clean. So the review needs a trigger. If the share of quotes flagged as exceptions doubles, or our costs change, that’s the moment to stop approving on autopilot and look again at the margin rules. Without something like that, the review becomes a rubber stamp with our name on it.

I’m not sure how far the number of approvals an owner needs to give will fall, and it’ll vary from one business to the next, but I’d expect it to fall faster than we’re planning for.

Where the attention goes

Once the approvals stop eating the day, the question is what to do with the time. For most owners the answer won’t be a new market or a reorganisation. It’ll be the sales conversation we keep postponing, the product decision that’s been waiting since spring and the customer we haven’t phoned in a year, the parts of the business where our judgement and our relationships are what the customer is paying for.

The human in the loop will be us for a while yet, but it can be a loop we’ve chosen, with the rest of the day given back to the work that made us start a business in the first place.