Can I just ask the AI to double-check its own answer?

It helps a little and proves nothing. A second pass by the same tool can catch a typo or an inconsistency, but it can also confirm a wrong answer with the same confidence. For facts that matter, check against the source: your price list, your calendar, the client file, the regulator's own page. A second tool is a different kind of check, but it isn't a substitute for looking at the original.

Does a better tool mean I can stop checking?

No. Newer tools tend to get things wrong less often, which is a reason to check less often on low-stakes work, not a reason to stop checking where a mistake is expensive. The checking rule should follow what a wrong answer costs you, not how good the tool is this month.

How do I find out where my own setup gets it wrong?

Read a batch of its drafts against the real records. Take twenty replies or summaries, put each fact in them next to the source, and mark every mismatch. Count what you find and note which type of fact it was. After one afternoon you will know which details need a person and which can be left alone.

It sounds right and it's wrong: where AI mistakes cost you

Why does it sound so sure when it's wrong?

Generative AI writes the most likely next words given what you gave it. It isn't looking facts up unless it has been set up to, and even then it can misread what it finds. Fluent text is the thing it does best, so fluency tells you nothing about accuracy.

A person who isn't sure usually shows it. They hedge, slow down, say they'll check. A draft from an AI tool does none of that. A made-up phone number and a real one look identical on the page.

It also knows nothing about your business that you haven't put in front of it. Ask it to quote a job and it will produce something quote-shaped, with figures that look plausible, whether or not they are yours.

Where does it actually cost you?

In a small business the expensive errors cluster in the same few places:

Prices and scope. A quote that undercuts your rate, or promises something you don't offer, is a commitment the moment it is sent.
Dates and availability. A reply that confirms Thursday when your Thursday is booked.
Names and details. The wrong client's name, the wrong job address, a detail carried over from a different matter.
Figures in a summary. A monthly report where one number is off and the rest are right, which makes the whole thing look trustworthy.
References to rules and obligations. A confident statement about what a regulation requires, which may be out of date or simply not so.

Two cases that were decided in public, an airline chatbot's wrong advice and a solicitor's list of cases that didn't exist, are set out in what AI won't do for your business. Neither involved a small firm, and the lesson carries down to one: the business answers for what it sends, however the words were produced.

What these have in common is that the error looked ordinary. Nobody was careless in the usual sense. The draft read well and went out.

Which mistakes are cheap and which aren't?

Ask two questions of any piece of output: does it leave the building, and does it contain a fact someone will rely on?

Stays inside, no one relies on it. A first-pass summary of your own meeting notes, a rough list of ideas. A wrong detail costs you a moment when you spot it.
Stays inside, but you'll act on it. A weekly numbers summary, a task list built from emails. A wrong figure sends your decisions in the wrong direction, so these need a spot-check.
Leaves the building with a fact in it. A quote, a booking confirmation, a reply to a client about their matter. A wrong detail is now yours in writing.

The third kind is where a person should always look before it goes. The first kind is where you should stop worrying.

What review rule doesn't eat the time you saved?

The usual objection is that checking everything takes as long as writing it. That's true if you check everything the same way. It stops being true once you sort the work by stakes.

A rule that holds up in a small business:

Approve before sending anything containing a price, a date, a promise or a client's name or details. This takes seconds when the draft is good and the fact sits next to the source.
Spot-check internal summaries and reports. Read a few against the original each week at first, then less often as the results hold.
Release routine wording you have already approved, such as an acknowledgement that an enquiry arrived. The decision was yours when you approved the wording, so it can go automatically.

Then make the fact easy to check. A good setup puts the figure and its source side by side, so approving means glancing at one line rather than rebuilding the answer. If checking is awkward, people stop doing it, and that is the real risk.

How do you find out where yours goes wrong?

Don't guess. Measure it once. Take twenty recent drafts, replies or summaries from whatever you are using. Put every specific fact in each next to its source, and mark each mismatch.

You are looking for patterns rather than a score. Perhaps it is always the date format, or always the second client on the list, or the figure pulled from the older of two documents. Each pattern points to a fix: better instructions, a cleaner source document, or a rule that a person looks at that one field.

Most of the repeat errors come from the inputs, not the tool: two versions of the price list, a folder with last year's template in it, instructions that never said which source wins. Fix those and the error rate drops without changing tools. When automation breaks goes through what to do when something has already gone wrong.

What about regulated work?

If you work under professional obligations, such as law, accounting or migration advice, the stakes on the third kind of output are higher, and a person's sign-off is part of the job rather than a precaution. This is general information, not advice. Check what your professional body and regulator expect of you when you use AI tools, and confirm your own obligations.

What to do next

Pick the one AI-assisted task you use most and pull twenty recent outputs.
Check every price, date, name and figure against its source and write down each mismatch.
Sort your tasks into the three kinds above: stays inside, stays inside but acted on, leaves the building.
Set the rule for each kind: approve, spot-check or release.
Fix the inputs behind any repeat error before you blame the tool.

If you'd like a second pair of eyes on where this applies in your business, the AI Tools Assessment covers it, or you can book a free fit chat to start.