The two ways businesses get AI wrong, and they are opposites.

Finance people especially. They have spent a career being the person who catches the error, and now someone wants them to hand the numbers to a thing they cannot see inside. They do not know what it can do, they have heard it makes things up, and they are right to be careful.

What has worked every single time is not a demo and not a training day. It is introducing it as a check, not a doer. Nothing leaves their hands. The AI reads the reconciliation before it goes out and flags anything that looks off. It reads the supplier invoice and says whether the totals add up. Nobody is afraid of a second pair of eyes, and the first time it catches something they would have missed, the conversation changes.

Then one small task. Not the month-end, one piece of it. The bit they hate. Then another. I have never seen this fail, and I have never seen it take long. Within a couple of weeks the same person who would not touch it is asking what else it can do. The addiction starts with one taste, and the taste has to be small.

The other camp is a pleasure to work with. The owner has decided AI is the future, every person on the team has a plan and a licence, and everyone is being challenged to use it. That is exactly the right instinct and it is genuinely rare.

Then you read the weekly reports. They are long, they are confident, they all sound the same, and nobody has actually read them, including the person who sent them. The same cliches, the same dashes, the same structure every other business is also producing. The trust that was so hard to earn in the first camp has been given away for free in the second.

The dangerous part is not the writing style. It is the numbers underneath. One example I keep running into: a report that buckets sales by day, produced by a tool that runs on UTC. Every “day” is ten hours out from a Brisbane day. The totals are plausible, the chart looks fine, and every single figure is wrong. Nobody catches it because it looks like a report.

The AI did not lie. It was never told which timezone the business runs in, never told what a day means here, and never had its first ten outputs checked by someone who knew the real numbers. Trust was extended before it was earned.

Treat the AI like a new hire. A good new hire does not get the month-end on day one and does not get their work waved through unread. They get a small task. Someone checks it. They get a bigger one. When they get something wrong, someone tells them, and it goes in the handbook so it does not happen again.

That handbook is the point. Every correction you make should be written down somewhere the AI reads every time: which timezone, which words we never use, how a report is laid out, what a good email from us looks like. I wrote about how I keep mine in plain files every tool reads. My own rule about never using a particular kind of dash exists because of exactly the reports described above.

Once that file exists, the two camps meet in the middle. The scared team has a reason to trust, because they can see what the AI has been told and they checked its early work. The all-in team has a reason to slow down, because they can see how much it had not been told.

1. Start with a check, not a task. Have it review something a person already did. Nothing leaves your hands.
2. Check its first ten outputs by hand. Against the real numbers, by someone who knows them. Not the first one, the first ten.
3. Write down every correction. One line, in a file it reads every time. Timezone, currency, names, words you never use.
4. Treat “looks right” as unverified. A report that reads well has proven nothing. Tie one number back to the source before you trust the rest.
5. Read what you send. If you would not put your name on it in your own words, it is not ready.
6. Then give it the next small task. Not the whole job. The next piece.

Do that and the scared camp stops being scared within a fortnight, and the all-in camp starts producing reports someone actually reads.