After 10 years of dev, I made the transition to AI PO. Here’s what I wish someone had told me before starting.
Technical skills remain an advantage
Being a developer is a huge asset as an AI PO. Three concrete reasons:
You validate feasibility in real time. When a stakeholder asks “can we do X with AI?”, you can answer precisely rather than with a vague “I’ll look into it.”
You communicate with the technical team as equals. Developers trust you because you understand their constraints. Less friction, more velocity.
You write better specs. A spec written by someone who has implemented features is much more useful than one from a PM who’s never opened an IDE.
The main mental shift
The hardest change: you no longer produce code, you produce decisions.
As a dev, your value was measurable and visible — I implemented X, fixed Y, here’s the commit. As a PO, the value is less tangible — I identified that X was the right thing to do, I avoided implementing Y which wouldn’t have worked.
This requires genuine self-work, especially if you were a prolific developer.
Mistakes I made
Mistake 1 — Wanting to control everything technically. At first, I wanted to validate every technical choice, every prompt, every integration. That’s counterproductive. Your role is to frame, prioritize, and remove blockers — not to code the agents yourself.
Mistake 2 — Confusing speed with haste. AI projects have a learning curve. The first 3 weeks seem slow because you’re building foundations (evals, architecture, test data). Resist the pressure to deliver features before having these bases.
Mistake 3 — Underestimating data work. AI agent quality depends 60% on the data and examples you give it, and 40% on the prompt. I spent too much time optimizing prompts before having quality data.
Skills that make the difference
Knowing how to write evals. How do you objectively measure if your agent is better than before? Without rigorous eval, you’re navigating blind. It’s the most underestimated skill in the field.
Understanding costs and latency. Every LLM call costs time and money. An effective AI PO optimizes both dimensions continuously, not just at the end.
Communicating uncertainty. AI is probabilistic. “It will work in 95% of cases” is a valid deliverable. Knowing how to communicate that to stakeholders without alarming or lulling them — that’s an art.
Stéphanie Caumont
AI Product Owner · Learn more