Agents can now run, transcribe, and synthesize a hundred customer conversations a week — which means a founder who asks flattery-shaped questions no longer gets a little false confidence, but an industrial supply of it.
The Mom Test, Rob Fitzpatrick's short book on customer conversations, starts from an uncomfortable premise: people you talk to — including your own mother — will lie to be nice, and the lies get worse the moment you pitch. His fix is structural, not motivational. Ask about the customer's life and their past behavior, never about your idea or their hypothetical future. Treat compliments as noise carrying zero information. And measure a conversation by whether the other person gave something up — time, money, reputation, an introduction — because commitment and advancement are the only signals that survive politeness.
The temptation in an AI-native studio is to point an agent at user feedback and ask it to find validation — and it will, because agents are even more agreeable than moms. The translation is to encode Fitzpatrick's discipline into the harness itself: prompts that extract specific past events rather than opinions, evals that discard compliments before synthesis, and pipelines that log only commitments — a renewal, a workflow change, a payment — as evidence. The human keeps the judgment call the book reserves for founders: deciding what the pattern of real behavior means for what to build next.
- never let an agent ask about your idea when it can ask about the customer's last week
- filter compliments out of every synthesis pipeline before they compound
- count commitments, not kind words, as evidence.
