An AI tool gives someone on your team an answer. It is fluent, well structured, and sounds completely sure of itself. It is also wrong. What happens next does not depend on whether the person can spot the mistake. It depends on whether they feel it is their place to argue with it.
Most of the time, the answer is fine. That is the trap. When the tool is right nine times out of ten, people quietly learn that checking it is wasted effort, so they stop looking properly. By the time the tenth answer turns up, confident and incorrect, nobody is really reading any more. The skill you needed, catching the wrong one, has gone soft from lack of use.
So the thing worth building is not trust in the AI, and it is not suspicion of it either. It is the habit of disagreeing with it well: knowing when to push back, feeling allowed to, and making it cheap to do.
A team must be able to challenge AI
There are two ways this goes wrong, and they pull in opposite directions.
The first is over-trust. People treat the output as a finished piece of work rather than a starting point. In a poll of more than a thousand managers earlier this year, seven in ten said they had seen at least one AI-related mistake from someone on their team in the past twelve months. The cause, in the researchers' own words, was over-trust and under-scrutiny. The errors were not exotic. They were ordinary things: a wrong figure, missing context, a confident claim nobody checked.
The second is quiet refusal. Someone got burned once, decided the whole thing is unreliable, and stopped using it without telling anyone. This looks like the safe failure, and it is the less dangerous of the two, but it is still a loss. You are paying for a tool the team has privately written off, and you will not find out until you ask.
The genuinely risky team is the first one: the team that uses the AI for everything and cannot disagree with it. They are fast, productive, and one plausible wrong answer away from sending it to a customer. A team that ignores the AI wastes your money. A team that cannot argue with it spends your credibility.
AI confidence makes disagreement feel harder
It helps to understand why over-trust is the default rather than a failing of careless people.
The tool is fluent, and fluency reads as competence. It gives a clean answer in one go, with no visible working and no hedging, which is exactly how a confident expert sounds. It is often right, so deference gets rewarded. And most people are busy, so handing the thinking over is a relief, not a risk they are weighing up.
There is a quieter cost underneath all this. A study this spring found that the people who passively accepted whatever the AI produced grew less confident in their own judgement over time, while the people who challenged it and changed things kept theirs. Disagreeing with the tool was not only safer in the moment. It was the thing that kept their own thinking in working order. The conclusion the researchers drew was blunt: the problem was never AI use itself, it was passive acceptance.
That reframes the job. You are not trying to make people distrust a useful tool. You are trying to stop them switching their own judgement off.
Make disagreeing normal, easy, and expected
Scepticism does not arrive on its own, and telling people to "be critical" does nothing. It has to be built into how the work runs. Three things do most of the work.
Name who is allowed to say no. People defer to the AI partly because they assume someone more senior, or the person who bought it, must know it is right. Say out loud that the person doing the task is expected to override the output when it is wrong, and that doing so is the job, not a complaint about the tool. Permission is most of it.
Make saying why cheap. If overriding the AI means filling in a form or justifying yourself in a meeting, people will stop bothering and wave things through instead. A one-line note next to the change, "swapped this figure, the AI had last year's", is enough. The point is a light record of where the tool gets things wrong, not a tribunal.
Reward the catch, never punish it. The first time someone overrides the AI and turns out to be right, that should be a good day for them, visibly. And a good-faith override that turns out to be unnecessary must never be treated as a mistake, or you teach everyone to stop checking. You want people erring towards looking, not towards waving through.
One more thing matters more than any process: leaders have to be seen being overridden. If the owner's AI-drafted email gets corrected by someone junior and the owner says "good spot" rather than getting defensive, the whole team learns that arguing with the output is safe. If the owner treats their own AI drafts as above question, so will everyone else with theirs.
We make review part of daily AI work
We use AI drafts every day, for code, for writing, for reports, for sorting incoming work. None of it ships because the model produced it. It ships because a person who is allowed, and expected, to say "no, that is wrong" has looked at it and signed it off.
The clearest version is code review. When one of our tools drafts a change, the reviewer's job is not to wave through anything that looks plausible. It is to reject anything they do not understand well enough to stand behind, even when the AI sounds sure, especially when the AI sounds sure. We are the ones who get the support email after launch, so "the model wrote it" is never an answer we would accept from each other. The disagreement is not friction in the process. It is the part of the process that makes the speed safe to keep.
People still need time to review
Scepticism has a cost, and it is worth naming so you do not overcorrect. If everyone re-checks everything from scratch, you have thrown away the reason you brought the tool in. The aim is calibration, not paranoia.
So you decide where the scrutiny goes. Anything irreversible, anything a customer sees, and any specific factual claim gets read properly. The low-stakes, easily-undone work can run lighter. You will still miss things, because no team catches every error, and a culture that pretends otherwise just makes people hide the misses. The realistic goal is that the important mistakes get caught by someone who felt able to say so, before they leave the building.
If you are weighing up an AI tool for the team, the most useful question is not how good its answers look in the demo. It is how easy your people will find it to disagree with those answers on a normal Tuesday. A tool, or a culture, that makes overriding the output awkward will quietly train everyone to stop trying, and that is the expensive failure.
A team that cannot disagree with the AI is not a team that has adopted it well. It is one that has handed over its judgement and not noticed. Decide who is allowed to say no, make it easy and expected, and let them practise on the small stuff so the habit is there when it counts. If you want to work out where AI genuinely helps in your business, and where it still needs a person ready to push back, that is the sort of thing we talk through on a short call.
