What AI still gets wrong, and how to catch it
Anyone selling you AI who won't tell you where it fails is selling you something else. Here's the honest list, and what to do about each.
It invents specifics with total confidence
This is the big one. AI will produce a plausible-sounding fact, figure, date, name, quote, statistic or citation that is simply not true, and it will say it in exactly the same tone as everything it gets right.
There's no tell. No hedging, no uncertainty in the voice. The wrong answer sounds identical to the right one.
What to do: treat every specific number, name, date or quoted fact as unverified until you check it. Not the general reasoning: the specifics. If a draft mentions a statistic and you can't find the source in thirty seconds, cut it.
The practical rule: AI is reliable for shaping and drafting, unreliable for facts you didn't supply.
It doesn't know what happened recently
Every model has a cutoff date, and things that happened after it are invisible unless the tool is actively searching the web. It also doesn't know anything about your business that you haven't told it.
What to do: supply current information rather than expecting it. If something depends on a current price, a current rule, or a current fact about your industry, provide it or verify it separately.
It agrees with you too readily
Push back on a correct answer and it will often fold and give you a worse one. It's built to be helpful, and that leans toward agreement.
This is dangerous when you're using it to check your own thinking, because it'll tend to confirm whatever you seem to want.
What to do: ask it to argue the other side explicitly. "What's the strongest case against this?" And be suspicious of instant agreement on anything you were genuinely unsure about.
It's confidently mediocre at maths
Not arithmetic (modern tools handle that fine) but multi-step calculations with real consequences: margins, tax, compound figures, anything where a small error compounds.
What to do: check anything involving money in a spreadsheet. Always. It takes two minutes and the failure mode is expensive.
It writes in a recognisable way
Certain rhythms, certain phrases, a particular tidiness. Plenty of people can now spot AI writing, and a customer who spots it on your site draws a conclusion about how much you care.
What to do: always rewrite the opening and closing lines yourself. Cut anything that sounds like nobody. Read it out loud. The sentences you'd never actually say are the ones to fix.
It can't tell you what your customers want
It can tell you what businesses in general tend to do. It has never spoken to your customers, seen your area, or watched someone hesitate before booking.
What to do: use it to generate options, not to decide. The deciding needs information only you have.
It fails silently on the boring stuff
Broken links. Wrong contact details carried over from an example. A price it helpfully updated when you didn't ask it to. Placeholder text that made it into the final version.
These are the errors that actually reach customers, because nobody proofreads the parts that seem routine.
What to do: before anything goes live, check the four things that matter most: contact details, prices, links, and your own name. In that order.
The way to think about it
AI is a fast, tireless, well-read assistant with no judgment and no stake in the outcome. It will do an enormous amount of work well and occasionally do something wrong with complete confidence.
That's a genuinely useful thing to have. It is not a thing you leave unsupervised.
The businesses getting real value out of this are the ones who put a human check on the last ten percent. The ones getting burned are the ones who assumed the last ten percent didn't need one.