How would you know if it's wrong?
Everyone says to check what AI produces. I've said it on this site myself. But "check it" isn't much use on a Tuesday afternoon with forty drafts to get through, so this guide is about what checking actually means.
You're not alone in being unsure
In the government's 2026 survey of 3,503 UK businesses, 56% of those using AI had concerns about bias and accuracy. The same survey found 81% apply some human oversight, but that doesn't tell us how good the oversight is.
Why a quick read isn't a check
AI writes confidently whether it's right or wrong. A well-written paragraph with a made-up figure in it reads just as smoothly as a true one, and that's what makes it different from a colleague's rushed first draft.
That's why a read-through mostly catches mistakes in how something sounds. It won't catch a wrong date unless you go and look at the date.
Check according to what a mistake would cost
Not everything needs the same attention, and treating it all the same is how checking quietly stops. I'd sort each task by asking what happens if it goes out wrong.
| If it went out wrong | How to check | Example |
|---|---|---|
| Embarrassing, easily put right | Spot-check a sample each week | Tidying up internal notes |
| A customer or colleague is misled | A person reads every one, against the source | An email quoting a date or an amount |
| Money, law or someone's job is affected | A person who knows the subject does it properly and signs it off | A contract, a pay decision or anything sent to a regulator |
These are examples, not rules. The same task can sit in a different row in a different business.
An example: chasing unpaid invoices
Say AI drafts your payment reminders. The tone is low-risk, because a slightly stiff email won't hurt anyone. The facts aren't low-risk, because chasing a customer who has already paid costs you goodwill.
So the person checking doesn't reread the wording. They compare three things with your accounts: the name, the amount and the date. It takes seconds, and it's the part that matters.
The checker has to know the subject
This is the part that gets missed. If AI drafts something you couldn't have written yourself, you may not be able to tell when it's wrong.
A UK tribunal made a similar point about lawyers in UK and Munir [2026] UKUT 81. It said a supervisor stays responsible for a junior's work, and that it matters not how citation errors come about. That case is about legal professionals, not your office, but I think the principle travels: the person who signs it off owns it, whoever or whatever drafted it.
When something gets through
It will happen eventually, with or without AI. Put it right with the person affected first. Then ask where the check failed, rather than who failed. If people get blamed for finding mistakes, they'll stop telling you about them.
This guide is general information, not legal advice.
