We changed who's in the room for early AI discovery: the experience owner and the model expert, nobody else at first. A hackathon proved the pattern.
A recent hackathon taught us more about process than product. The pairings that moved fastest turned out to be a designer working directly with an applied AI engineer from first contact, nobody else in the room yet. The engineers had every piece of data they needed before anything got built, and the scope estimates, for once, held. We’ve since made that pairing our default for early AI discovery, and we wrote down the risk in the same breath: put the full cast back in the room too early and old habits return, with the AI expertise minimized back into a resource instead of a partner.
I want to be precise about the claim. This is sequencing, not exclusion. Product management, research, and engineering leadership all come back into the room the moment the question changes from what’s possible to what’s worth shipping. Early discovery answers the first question, and the first question is a 2-person conversation: the person who owns the experience and the person who knows what the model can actually do, in direct contact, no interpreters.
Every seat you add to early discovery adds a translation layer, and AI work dies in translation.
Translation kills AI work for a specific reason: feasibility in AI fails at a strange grain. Not whole features but this particular prompt against this particular data. A designer sketching against an imagined capability wastes a sprint. A designer sketching next to the person who can test the capability in an afternoon wastes nothing. Every intermediary between those 2 people rounds the weirdness off, and the roundoff is where scope surprises live.
This continues a case I made in Issue 13: AI inverted the economics of product development, so discovery has to move to the front of the process. This week’s addendum is about casting. Once discovery leads, it matters who holds the flashlight, and the smallest room holds it steadiest. The industry version of this mistake is everywhere right now: AI initiatives staffed like platform migrations, 9 stakeholders deep before anyone has established what the model can do on real data.
It’s a pilot, not a doctrine. The risk cuts both ways, and we logged that too: a 2-person room can fall in love with a demo nobody needs, which is exactly the failure the returning cast exists to catch. So the discipline runs in both directions. Small room first, full room fast behind it, and we’ll report back once the pattern has survived a full roadmap cycle.
The teachable part
Start AI discovery with the smallest room that holds the experience owner and the model expert. Add seats when the question changes from what’s possible to what’s worth shipping.