After enough conversations with companies that tried AI and gave up, the list of reasons gets very short. It is almost never the model or the subscription. It is four habits, repeated identically everywhere.
The second one is the expensive one
The first answer is, by construction, an average one: safe, balanced, unsurprising. It is the starting point, not the result. Anyone who accepts it as final reaches the reasonable conclusion that AI writes mediocre text — which is true of the first answer and false of the third.
The fourth is the only one that costs you outside the company
Bad prose stays inside the company and gets rewritten. A wrong figure in a quote goes out to a client. The practical rule is simple: every number, date and proper noun in anything leaving the company gets checked at the source, however confident it sounded. Ask the model to flag explicitly what it could not verify — and do not assume it does so by default.
The model does not know when it does not know. You have to ask where the number came from.
Terms used here
- Hallucination
- A plausible false statement, produced with the same confidence as a correct one. This is why you ask for the source.
- Iteration
- One round of correction inside the same conversation. The useful result usually arrives on the second or third.
Where to start
- 1Write two lines about the company and paste them at the top of every request for a week.
- 2Never accept the first answer. Ask for a second version even when the first looks fine.
- 3Before anything leaves the company, check every figure against its source.
