AI Product Leader

Where the jobs come from

You can't automate the interview. You can automate everything around it.

Nothing replaces a live customer interview — I tested that conclusion hard before accepting it. A synthetic persona generates what's plausible, and plausible is the opposite of what you go into discovery for. But most companies already sit on an archive of real conversations with real customers: sales calls, demos, onboarding, support. Those are authentic words from real people, collected for a different purpose. The model isn't a generator there. It's an extractor.

Given the discipline to store those conversations properly and a lawful basis to use them, transcripts give up Jobs at scale — the context, the trigger, the success criteria, the alternatives that were weighed. That produces a map instead of a handful of points, which is what stops live interviews being spent on reconnaissance.

Where it stops. It will not surface hidden, unarticulated jobs — those still come out live. The archive is also biased towards people who made it to a conversation at all. And with no lawful basis to use the recordings, the method doesn't run.

Where the mechanics come from

Shop for mechanics on a donor market.

Benchmarking direct competitors converges a whole category on the same dashboard. Everyone studies everyone, so nobody finds a mechanic that isn't already in the room. That's a local optimum, reached expensively.

So I look for a more mature market whose users behave the same way, and go shopping there instead. A performance media buyer works several screens with their own money at risk in real time, decides within the day, and distrusts somebody else's automation. So does a day trader — and trading terminals are decades of product selection already paid for by a bigger, older, wealthier market. Thresholds instead of reports. Watchlists. Acting straight from the alert rather than being sent to an interface. Real-time P&L. Stop-loss limits. Deliberate information density where the design instinct says to simplify.

Where it stops. An analogy is not an identity — decision horizons, switching costs and regulation all differ, and part of the segment genuinely works slowly and in batches. Every transfer is a value hypothesis, and it goes through the same cheapest-test-first validation as anything else.

How the value gets delivered

Silence is a feature.

There have been two eras of value delivery. In the first, the user had to come in and do the work. In the second, products started doing the work and delivering a result — reports, digests, recommendations. Both still charge the user attention: somebody has to read the thing and decide whether it mattered.

The third shift isn't more automation, it's less attention. The product stays quiet while everything is fine, and speaks only when an action is required — carrying the action, not just the news. A recurring digest is actively harmful: if nine messages in ten need nothing, it trains people to ignore the tenth. Zero false alarms beats completeness. And the metric isn't open rate, it's the share of messages followed by an action.

Where it stops. Staying quiet means occasionally missing a weak signal. I take that trade knowingly: a missed alert costs one mistake, while trained-in indifference costs the entire channel — and that one is irreversible.