AI never says "I'm not sure." It says the wrong thing with exactly the same confidence as the right one. That single trait is behind nearly every AI horror story in our profession.

There's a particular kind of wrong that only AI does, and it's the one that'll catch you out.

A junior who's unsure hedges. "I think it's this, but let me check." That uncertainty is a gift — it tells you exactly where to look. AI doesn't do that. Ask it something it only half-knows and it hands back a confident, fluent, perfectly formatted answer that happens to be wrong. Same tone as when it's right. No flicker of doubt.

It once gave me a tax rate quoted to the decimal, with a plausible-sounding justification, that was simply last year's figure. If I hadn't already known, I'd have used it.

That's the trait to burn into your team's heads: AI is fluent, not accurate. It will invent a piece of legislation, cite a case that doesn't exist, and quote a made-up statistic — all in the register of a confident senior manager. It isn't lying. It has no idea it's wrong, which is worse, because it can't warn you.

The fix isn't distrust. If you re-check everything from scratch, AI has saved you nothing. The fix is a named habit — the Reconciled Review Method: AI-assisted, human-verified. Every piece of AI work gets a review step that's a real step, with an owner, not a vague "someone will glance at it." On anything factual, you verify the claims, figures and references against source before you rely on them. That's not bureaucracy. It's the same reason you review a junior's file before it leaves the building.

There's a quieter danger underneath, and it worries me more. If your team leans on AI for the thinking, not just the typing, the skill quietly erodes. You end up with people who can prompt but can't judge whether the answer's any good. So the review habit does two jobs: it catches today's errors, and it keeps your people sharp enough to catch tomorrow's.

The signature on the work is still yours. The tool doesn't carry the risk — you do. So the review trail, not the model, is what actually protects the firm.

Pick one workflow where you already use AI and add a verification step that can't be skipped:

Done once, it stops being a chore and becomes reflex — the same reflex you already have for reviewing anything before it goes to a client.

Don't rely on telling people to be careful — build the check into the workflow so it can't be skipped, and give it an owner. Then do the thing that actually shifts behaviour: show them. Next time AI confidently invents a figure or a reference, share it with the team. One vivid example of a fluent, authoritative, completely wrong answer does more than any amount of "please double-check." Once they've seen it lie with a straight face, they never fully trust it again — which is exactly where you want them.

You can't stop AI being confidently wrong — but you can make it show you where to look. Run this over any AI-drafted output before you rely on it:

It turns a wall of confident prose into a checklist. You still do the verifying — but now you know precisely which lines to distrust.

Add the "AI-assisted, human-verified" step — with a named owner — to one workflow where you use AI. Then hit reply and let us know how you got on, and tell us the best confidently-wrong answer AI has ever handed you.

You've got the colours and the check habit. Next week we write them down: the one-page AI policy, plus the engagement-letter wording that now needs to match it.