What AI gets wrong, and how to catch it
On this page
AI tools are useful and they're wrong more often than they sound. The worst part is that wrong answers look exactly like right ones. Same confident tone, same tidy formatting.
You don't need to distrust everything. You need to know where it slips, and have a few quick checks ready.
How it goes wrong
It makes things up. This is sometimes called hallucination, and it's the big one. Ask for a business license rule, a price, a quote, or a link, and it may give you a very believable one that doesn't exist. It doesn't know it's wrong. It's producing text that sounds right.
It does bad math. It can be fine with simple sums and then botch a longer one, or miscount items in a list. If a number matters, don't take its word for it. Put the figures in a spreadsheet.
It's out of date. The tool learned from information up to some point. Prices, hours, laws, and software features change. Unless it's looking things up live, it might be describing last year's world.
It agrees with you. If you ask "Isn't this a good idea?" it tends to say yes. Ask it to argue against your plan instead and you'll learn more.
It misreads what you wanted. You asked for a friendly reminder and got a legal-sounding notice. The fix is to be specific, and to look at the first result before assuming the rest will be right.
It forgets. In a long conversation, it can lose track of things you said at the start. If it starts contradicting itself, start a new chat and paste in the important facts again.
Five checks that catch most of it
- Check every number, name, date, and link. Open the link. Look up the date. These are the things most likely to be wrong and easiest to verify.
- Ask where it got it. Then go and look at the source yourself. If it can't point to one, or the one it gives doesn't exist, treat the claim as unverified.
- Test it on something you know. Before trusting a tool with a task, give it ten examples where you know the right answer. See how many it gets right. That tells you more than any review.
- Ask twice. Ask the same question in a different way, or in a new chat. If you get different answers, neither one is safe to use.
- Have a person look at anything important. If money, health, legal standing, or someone's job is involved, a human with the right know-how should check it.
None of these take long. Most take a minute.
What to hand over, and what not to
| Job | Fine to let it draft | Always check | Don't hand over |
|---|---|---|---|
| Customer replies | First draft of a routine reply | Facts, prices, promises | Complaints, refunds, upset customers |
| Writing | Social posts, menu descriptions, job ads | Claims about your business | Anything you'd sign |
| Numbers | Explaining what a formula does | Every calculation | Payroll, taxes, final books |
| Research | A list of questions to ask a vendor | Every fact and source | Legal or medical conclusions |
| Scheduling | A first-pass shift grid | Hours, rules, who can work when | Final schedule, unreviewed |
If you're unsure which column a task belongs in, use the stricter one.
Things never to let it decide
- Legal questions: whether a contract is safe to sign, whether you're allowed to do something
- Medical advice, or anything about someone's health
- Tax and payroll decisions
- Who to hire, fire, or discipline
- Anything sent to a customer without you reading it first
It can help you prepare for these, for example by listing questions to ask your lawyer or accountant. It shouldn't be the one deciding.
Build the habit
The goal isn't to be scared of the tool. It's a habit: draft with AI, check by hand, send under your own name. Teach that to your staff too, and write the rule down where they'll see it. Something short, like "Read it. Check the facts. Then send."
If you run an AI helper that talks to customers, check its replies regularly. Read a sample every week for the first month. You'll spot problems the tool never will.
For examples of where AI helps in the back office, read using AI for scheduling, bookkeeping help, and research.