If you're using customer conversations to validate your AI product, the conversations are probably lying to you, and you're probably helping them do it.
Rob Fitzpatrick's The Mom Test is built on a single uncomfortable observation: people will tell you your idea is great because they want to be kind, and that politeness destroys companies. The fix is to stop asking about your idea entirely and start asking about their life — their actual workflows, the last time a thing went wrong, what they do when it does. Past behavior and concrete specifics are hard to fake. Enthusiasm and hypotheticals are free, which means they're worthless.
The agent layer adds a new distortion: founders are now so deep in capability thinking — what the agent can do — that customer conversations drift even further from reality. You end up stress-testing your architecture with someone who just told you they'd "definitely use that," which is the Mom Test failure Fitzpatrick warned about, wearing a demo hat. The discipline holds exactly the same: find the broken Tuesday, the workaround they're embarrassed about, the task someone's doing manually because nothing else survived contact with their actual situation. Commitment and advancement — a deposit, a changed schedule, a real next step — still separate signal from social noise, whether your product runs on Python or politeness.
- Ask about their life and past behavior, never your concept
- compliments from a demo are as data-free as compliments from your mother
- a genuine next step from the customer is worth more than a room full of enthusiastic nods.
