What should you never hand to an AI model?

The judgement. Give a model labour you can check: work where you could say, before seeing the result, what a right answer looks like. Keep the decisions: the price, who the product is for, what you say when it goes wrong, and what to switch off.

The pricing page I copied

Years ago I asked a model what my product should cost. It was a membership-renewal tool for small clubs, the kind with a secretary who sends 2 emails a year. I'd built most of it by asking a model what a product like it should have, and when I reached the price, I asked that too.

The answer was excellent: 3 plans with sensible names, monthly and annual prices, a paragraph on why the middle one would sell best, a line about a charity discount. It looked like a real company's pricing page, and I copied it almost word for word. (I'm not proud of this.)

It looked like a real company's pricing page because it was an average of real companies' pricing pages. There were 4 decent tools in that category. The model had told me, very well, what they charged. It couldn't tell me what my customers would pay, because it had never met them, and neither, to be honest, had I. The product closed 5 months later.

Labour and judgement

• Labour: read this document and pull out the dates. Draft a reply I'll read before it goes. Compare these 2 lists. You can check the result.
• Judgement: what to charge, which customers to serve, what to say when it broke, when to stop. Nobody can check those for you, so they stay with the person who'll answer for them.

Use the model as research, never as your mouthpiece. Ask it what others charge, then go and ask your customers what they'd pay.

Who Pays for the AI? has the chapter. Nobody Bought It has the clubs tool's autopsy.