If you're using agents to run discovery conversations at scale, the failure mode isn't volume — it's baking your assumptions into every question your system asks. Fix the question design first.
Rob Fitzpatrick's The Mom Test starts from an uncomfortable truth: people will lie to your face to protect your feelings, and they'll do it cheerfully and in detail. The fix isn't better rapport — it's better questions. Ask about someone's life, their actual past behavior, the last time a problem cost them something real. Never ask whether they'd use your product or whether the idea sounds good. Compliments feel like signal but carry none. The only real signal is commitment — time, money, reputation put on the line before you've asked for it.
The moment you deploy an agent to run customer interviews or discovery calls, you've industrialized whatever question design you shipped. A leading question asked once is a mistake; asked ten thousand times it's a dataset poisoned at the root. Fitzpatrick's discipline — anchor every question to past behavior, never to hypothetical enthusiasm — becomes a system design constraint, not a conversational nicety. Your agents also can't read the awkward pause or the subject change that signals you've hit something real, so the human review layer has to be watching specifically for advancement signals: did the conversation produce a next concrete step, or just warmth?
- Design your agent's question logic around past behavior, not hypothetical approval
- treat compliments surfaced at scale as noise until commitment follows
- put a human on advancement review because agents can't feel the room.
