If you can build something in a day, why spend a month in discovery?

That question made us uncomfortable, because we had a well-developed discovery process and were proud of it. But it led to a more useful question: if AI can co-create specs with you, what do you actually need before you start?

Our answer, once we tried it: domain knowledge, empathy, and common sense. The people in the room already had all three. Most of our pre-planning existed to move those three things between people, not to create them.

The process we had was designed for human builders. Stakeholders spent weeks specifying solutions. PMs wrote detailed PRDs, user stories, and acceptance criteria. Design produced pixel-perfect Figma handoffs. Engineers estimated, then built. Every artifact existed to survive a handoff from one person to the next, and the whole path took weeks to months.

The process we moved to starts with a PM, a designer, and an engineer spending 60 to 90 minutes working with AI — we’ve used Agent OS and BMad, among others. The AI asks clarifying questions. The humans supply the knowledge, the empathy, and the judgment. The spec emerges from that dialogue, and the same AI then executes it with engineering oversight. A working demo exists in a day. We put it in front of real users and iterate on actual usage.

The demo also turned out to be a better discovery artifact than any document we used to write. The people who planned the feature get sharper ideas from using it than they ever got from reviewing it, and engineers take the finished product further because feedback arrives about something concrete.

The economics drove the decision more than the novelty. When shared context can become working software in a day, a month of talking before building is a cost you have to justify. We couldn’t.

The obvious objection is that you should validate before you build. We found we were still validating — with working software in front of users, instead of assumptions written into Jira.

The question that changed the most was our own. We stopped asking how to document work for engineers and started asking what the AI needed to co-create the spec with us. That change removed most of our discovery phase — roughly 80 percent of the elapsed time. Discovery didn’t go away. It now happens around something that runs.

Long-form discovery is still available to us. It has just become something we choose when a problem warrants it, instead of the default cost of starting.