Should this be AI at all?

1. 01We have evidence of the user need that is not a demo, a hackathon win, or an executive request.Support tickets, session recordings, workaround spreadsheets — artifacts users made before we offered anything.
2. 02We have mapped the cost of each error type: wrong answer, missed answer, and slow answer, separately.One number for “errors” hides the design decision. A missed fraud alert and a false one cost different people.
3. 03A deterministic alternative — rules, search, a form, a template — was seriously costed, not strawmanned.If the honest comparison is “the dropdown wins,” ship the dropdown.
4. 04The success metric is named, measurable, and is not engagement.Time-to-done, correction rate, task completion. “Weekly AI interactions” rewards a feature users can’t escape.
5. 05We have chosen precision or recall deliberately, and can say which errors we are accepting on purpose.Tuning happens either way. The only question is whether design chose the direction or the defaults did.
6. 06Kill criteria are written down: the measurable results that mean we stop, with a date attached.A feature with no kill criteria has no success criteria — just momentum.
7. 07Someone in the room argued for “no” at full strength, and the room engaged instead of waiting them out.If the no case took under five minutes, it wasn’t made. Assign it to your best builder, not your usual skeptic.
8. 08We can name the user for whom this feature makes things worse, and we accept that trade explicitly.
9. 09The team building it would use it weekly for their own work — and can say for what.“I’d use it” is cheap. “I’d use it Tuesday to do X” is evidence.
10. 10We know what breaks if the model provider changes pricing, latency, or behavior next quarter.A feature whose economics die at 2x inference cost is a bet, not a plan. Name the bet.
11. 11The person who says “go” is the person accountable for the kill criteria six months from now, by name.

Discussion prompts

• If the model never improved past today’s quality, would we still build this? If the answer depends on a future model, we are scheduling hope.
• Who in this room believes the deterministic version is actually better, and have they said so out loud yet?
• What is the strongest version of the case against building this — stated by someone who wants to build it?
• Which error will we tolerate on purpose: the confident wrong answer or the unhelpful refusal? Whoever hasn’t answered hasn’t designed anything yet.
• If this feature hits its kill criteria in six months, who pulls the trigger — and do they know that’s their job?
• Is anyone here excited about the user’s problem, or only about the technology? Point to what you were excited about last month.