The classic product playbook doesn't work for AI

One pattern I keep seeing across genAI and ML projects, especially on teams shipping this kind of work for the first time, is that they’re running the classic product playbook:

PM defines the problem (and sometimes a solution shape) → Engineering builds it → Ship → done ✅

That playbook doesn’t work for AI.

You end up shipping something that captures a fraction of the possible value, the results underwhelm, and the project gets quietly cancelled. Or you end up taking forever, trying a bunch of tweaks during the “build it” phase, because AI systems have a lot of possible levers, rather than working through them systematically.

The better rhythm for these projects requires deeper product and engineering collaboration, and much more iteration:

1. PM defines the problem.
2. Both define success metrics and how the system will be measured and evaluated.
3. Engineering builds V1 as a feasibility check and baseline.
4. Both analyze what V1 reveals, then prioritize experiments by expected impact.
5. Repeat, so that you end up with a series of experiments and clear improvement in success metrics over time.

Both is the important word here. The relationship between product and engineering is collaborative, and the shared expectation has to be that V1 is a starting point rather than a finish line.

This does not mean the project takes forever. The iteration loop should be super fast. But it needs to be intentional, and it can’t be skipped.

This way of working is new, and this expectation for product people is new too. But it’s a powerful and satisfying way of building, because you see clear evidence of improvement over time.