How do I check AI output when I run the business on my own?

Do not rely on rereading every draft, because that check fades as the drafts improve. Decide the key figure before you see the AI’s version, have the draft show where each figure came from so checking takes seconds, and check only the two or three details that would cost you money if wrong. Do the checking as a separate task rather than alongside other work.

Does telling yourself to always double-check AI actually work?

Not reliably. A 2010 review of research on automated decision aids by Parasuraman and Manzey, published in Human Factors, concluded that automation bias affects experts as well as novices and cannot be prevented by training or by explicit instructions to verify the aid’s recommendations. Checks need to be built into the work rather than left to intention.

Why do people stop checking AI over time?

Because the AI earns their trust. In a 2022 Harvard Business School field experiment, 181 recruiters given an AI that was right about 85% of the time were less accurate than recruiters given one that was right about 75%, because they spent less time on each case and followed the AI more readily. The effect was driven by the more experienced recruiters.

What is the most effective way to review AI-generated work?

Make your own judgement before you see the AI’s answer, and make the AI’s answer cheap to verify. In a 2021 study by Buçinca, Malaya and Gajos, having people decide first reduced how often they accepted wrong AI suggestions, though it did not eliminate it, and people disliked the designs that worked best. A 2023 Stanford and University of Washington study found overreliance fell when explanations made the answer easier to check.

Is keeping a human in the loop enough to make AI safe for a small business?

Not on its own. A human reviewer is only as good as the attention they can give, and the research on automated advice shows that attention drops under competing work and as the system becomes more reliable. A review step needs to be quick, narrow and built into the workflow, and some tasks are better left unautomated because the only check that would catch their errors is a careful read every time.

How to check AI’s work when you are the only one checking.

Most advice on using AI safely ends with “always have a person review it.” When you work alone, that person is you, at nine in the evening, on the fortieth draft. The research on what happens to that review is not encouraging, and it points to a different kind of check.

To check AI’s work when you are the only reviewer, do not rely on reading each draft before it goes out. Reading is the check that fades first, and it fades faster as the drafts improve: in a Harvard Business School field experiment, 181 professional recruiters given an AI that was right about 85% of the time ended up less accurate than recruiters given one that was right about 75%, because they spent less time on each case and went along with it more readily. Checks that hold up are built into the work instead. Decide the figure that matters before you see the AI’s version, have every figure in the draft shown next to the place it came from so checking takes seconds, and check the two or three things that would cost you money if wrong rather than rereading the whole draft. Even then some mistakes get through, so only hand over work where the ones that do would be survivable.

Almost every guide to using AI in a business ends on the same reassurance: keep a human in the loop. The phrase assumes a reviewer with time and attention set aside for the job. In a company that might be a role. In a one-person consultancy or bookkeeping practice it is you, reading drafts in the gap between the client work and bed.

That review feels like a safeguard because, in the first week, it is one. You read every line, you catch the wrong rate, you fix the client’s name. The question is what it looks like in month four, when the drafts have been fine for weeks and there are thirty of them waiting.

Telling yourself to check does not make you check

The research on this predates language models by decades, because pilots, doctors and plant operators have been working alongside automated advice for a long time. In 2010 Raja Parasuraman and Dietrich Manzey reviewed it in the journal Human Factors, drawing on studies from aviation, medicine, process control and military command rooms. Two of their conclusions matter here.

The first is that automation bias, following an automated aid’s advice when it is wrong or missing a problem because the aid did not flag it, shows up in experts as well as novices, and cannot be prevented by training or by explicit instructions to verify the aid’s recommendations. The second is that complacency appears under multiple-task load: when other work competes for the operator’s attention, the automated part is what stops being watched.

None of those studies involved a chatbot, and a cockpit is not an inbox. But the two conditions they describe are an accurate picture of a solo operator’s evening. The instruction “always check the AI’s output” is precisely the intervention the review found does not work, and reviewing drafts while also answering messages and finishing the day’s real work is precisely the setting where watching stops.

The better the drafts get, the less you look

The more uncomfortable finding comes from a field experiment by Fabrizio Dell’Acqua at Harvard Business School’s Laboratory for Innovation Science, written up as a working paper in 2022 under the title “Falling Asleep at the Wheel.” He hired 181 professional recruiters to assess 44 job applications each, on one criterion that could be scored objectively. Some got no AI. Others got recommendations from an AI that was right about 85% of the time, or from one that was right about 75% of the time, and were told which they had.

Both AI groups did better than the recruiters working without one, which is worth saying. But the recruiters with the weaker AI ended up more accurate than those with the stronger one. They spent longer on each application, clicked through more of it, and were less likely to simply go with the AI’s pick. The recruiters with the better AI had less reason to stay alert, and so they did not. The effect was driven by the more experienced recruiters, the people you would expect to be best at overruling a machine.

This is one study, of recruiters, on a narrow task, and it has not settled anything on its own. But the pattern should change how you think about a solo setup. The risky period is not the first fortnight, when the drafts are rough and you read everything. It is later, once the system has earned your trust and your reading has turned into a glance. If you are a lawyer or an accountant who knows the work cold, the finding suggests your experience is not what will catch the mistake.

The checks that work are the ones you will like least

There is better news in research on designing the review itself. Zana Buçinca, Maja Barbara Malaya and Krzysztof Gajos tested what they called cognitive forcing designs in a study of 199 people, published in 2021. In one of them, participants had to make their own decision before seeing the AI’s suggestion, and could then change their mind. Compared with simply showing the AI’s answer alongside an explanation, these designs significantly reduced how often people went along with the AI when it was wrong.

Two limits came with that result, and both apply to you. The designs reduced overreliance but did not remove it. And when the AI was wrong, people performed best in the conditions they preferred least, trusted least and found most difficult. The check that works feels like friction, which is exactly why a busy person working alone will drop it first if it depends on willpower.

A second study explains which kind of friction is worth keeping. Helena Vasconcelos and colleagues at Stanford and the University of Washington ran five experiments with 731 people in total, published in 2023, and found that people behave as if checking were a cost they weigh. Overreliance went up as the task got harder and as the AI’s explanation got harder to follow, and it went down when people were paid more for getting the answer right. Their reading is that an explanation reduces overreliance when it makes the AI’s answer cheaper to verify, and that earlier studies finding no effect may have used explanations that did not. Both studies used lab tasks (planning meals, solving mazes) with participants recruited online, not professionals doing their own work, so treat them as a guide to design rather than a forecast.

What a check looks like in a one-person business

Put together, the evidence says a review that relies on your attention will fade, and a review that is cheap and built into the work has a chance of lasting. In practice that means four things.

Commit before you look. Before opening a drafted quote, write down the total you expect from the call note. It takes ten seconds. If the draft agrees, move on; if it does not, that is the one to read properly. This is the decide-first design from the 2021 study, shrunk to fit an evening.
Put the source next to the claim. A draft that shows which line of your price list a figure came from, or which email a date was taken from, can be checked in seconds. A draft that makes you reconstruct the figure yourself takes minutes, and at minutes per draft you will stop.
Check the fields, not the prose. Decide in advance which two or three things in each output would cost you if they were wrong (the price, the deadline, what you have promised the client) and check those every time. Rereading the whole draft is what turns into a glance by month four.
Do the checking as its own task. Clearing drafts while also replying to messages is the multiple-task setting where the old research found watching stops. Ten minutes on the drafts and nothing else is shorter than it sounds and holds up better.

None of these depends on resolve, and that is the point. They have to be built into the place the draft appears, which is the part a chat window leaves entirely to you.

What this concedes

These checks cost time, and that time comes out of the saving. A quote that drafts itself in seconds and then takes a minute to verify is still faster than writing it from scratch, but the gap is smaller than the demo suggested, and it should be counted before anything is built rather than discovered afterwards.

They also do not make the output safe. The best design in the 2021 study reduced overreliance without eliminating it, and nothing in this research suggests a well-built review catches everything. Some mistakes will reach a client. The question is whether the ones that do can be apologised for and fixed.

And it cuts against what we sell. A build of ours is designed around a person making the decisions that matter, and the evidence above says that person is the weakest part of the system unless the checking is made cheap and kept narrow. Where it cannot be (a contract clause that changes meaning with one word, a tax figure nobody will look at again until it is wrong), the only check that would work is a careful read of every output, every time. If that is the job, it should stay with you.

Common questions

How do I check AI output when I run the business on my own?: Do not rely on rereading every draft, because that check fades as the drafts improve. Decide the key figure before you see the AI’s version, have the draft show where each figure came from so checking takes seconds, and check only the two or three details that would cost you money if wrong. Do the checking as a separate task rather than alongside other work.
Does telling yourself to always double-check AI actually work?: Not reliably. A 2010 review of research on automated decision aids by Parasuraman and Manzey, published in Human Factors, concluded that automation bias affects experts as well as novices and cannot be prevented by training or by explicit instructions to verify the aid’s recommendations. Checks need to be built into the work rather than left to intention.
Why do people stop checking AI over time?: Because the AI earns their trust. In a 2022 Harvard Business School field experiment, 181 recruiters given an AI that was right about 85% of the time were less accurate than recruiters given one that was right about 75%, because they spent less time on each case and followed the AI more readily. The effect was driven by the more experienced recruiters.
What is the most effective way to review AI-generated work?: Make your own judgement before you see the AI’s answer, and make the AI’s answer cheap to verify. In a 2021 study by Buçinca, Malaya and Gajos, having people decide first reduced how often they accepted wrong AI suggestions, though it did not eliminate it, and people disliked the designs that worked best. A 2023 Stanford and University of Washington study found overreliance fell when explanations made the answer easier to check.
Is keeping a human in the loop enough to make AI safe for a small business?: Not on its own. A human reviewer is only as good as the attention they can give, and the research on automated advice shows that attention drops under competing work and as the system becomes more reliable. A review step needs to be quick, narrow and built into the workflow, and some tasks are better left unautomated because the only check that would catch their errors is a careful read every time.