A WorkLighter how-to

How to check AI's work before it reaches a customer

Last updated 7 July 2026

You check AI's work by reading it for the things it gets confidently wrong before a customer ever sees it: numbers, names, dates, prices, and any promise you didn't make. An AI assistant (a chat tool at about £20 a month) drafts fast and well, but it states a wrong fact as surely as a right one, so the last read is always yours. It drafts, you check, you send.

Why AI is confidently wrong

The thing to understand about an assistant is that it doesn't know what it doesn't know. It writes by predicting likely words, so when it's missing a fact it doesn't stop and ask, it fills the gap with something that looks right: a price you never quoted, an opening time you've never kept, a policy you don't have. It says it with exactly the same confidence as the true parts. That isn't lying, it has no idea it's wrong, which is precisely why you can't skip the check. The good news is that its mistakes cluster in a few predictable places, so checking is quick once you know where to look.

What should you always check?

Read every draft for four things. Numbers first: prices, totals, VAT, dates and times, because a wrong figure is the costliest mistake and the easiest for the assistant to invent. Then names: yours, the customer's, a product or a place it might have guessed. Then promises: any delivery date, turnaround or offer that commits you to something you didn't decide. And last, the facts about your business: hours, policies, what you do and don't offer, which it will fill from nothing if you haven't told it. Everything else, the tone and the wording, is low-risk. Those four are where a mistake actually reaches the customer.

How do you make the assistant flag its own guesses?

You can get a long way by telling it to admit what it doesn't know. Ask it to leave a clear marker wherever it would have to guess a fact, instead of guessing. Then your check becomes filling in the markers rather than hunting for invisible errors.

When you draft anything for me that a customer will see, follow this rule: if you would have to guess a price, a date, a time, a name, or any fact about my business, do not guess. Write [CHECK THIS] where the fact should go, and I will fill it in.

Never invent a figure, a policy or a promise. If something important is missing and you cannot write a good draft without it, ask me one question instead of making it up.

What is a quick pre-send routine?

Before anything goes out: read it once for the four things above, fill any [CHECK THIS] gaps, and read it once more as the customer will. Ten seconds for a short message, a minute for a quote. It feels like a faff for the first week and then it's a habit, the same glance you already give a text before you send it. The point isn't that the draft is untrustworthy, it's that you're catching the rare confident error cheaply, not rewriting good work.

Where AI still needs you

It can't check itself. Asking it "is that right?" often just gets a confident "yes," because it has no independent view of the truth. It doesn't know your prices, your diary or your policies unless you told it, and it won't know when it's wrong. The check isn't optional and it doesn't go away, but it's seconds, and it's the thing that lets you use AI on customer-facing work at all. Draft with it freely; send nothing unread. Skipping it is the first of the AI mistakes small business owners make. To see how a full setup handles this, a complete recipe is free to read.