Français
Deutsch
The experiment: 30 days with Claude as co-PO
In January 2025, I ran a radical experiment: for one month, I systematically involved Claude in all my Product Owner decisions on the EQUANS Site Factory project. This lessons-learned report documents what worked, what failed, and what I learned about the real limits of AI as a work partner.
The experimental protocol
I defined three delegation levels for Claude:
- Level A — Consultation: I ask Claude's opinion before making a decision
- Level B — Co-drafting: Claude drafts, I review and validate
- Level C — Delegation: Claude handles a task end-to-end
Results by task type
User story writing (Level B → Level C)
After 2 weeks of learning the project context, the quality of Claude-written user stories was rated "very good" by the development team in 80% of cases. I moved to Level C for maintenance user stories and kept Level B for new features.
Backlog prioritisation (Level A)
Claude provided prioritisation analyses based on estimated impact and technical complexity. Useful as a starting point, but its recommendations consistently lacked political and commercial context. I never moved beyond Level A here.
Sprint review preparation (Level C)
Automatic sprint review deck generation (Jira data → slides) was the experiment's greatest success. Time saved: 2h per sprint. Quality rated equivalent to my manual version.
Stakeholder communication (Level B)
Claude drafts, I adjust tone and content based on my knowledge of the individuals. The most cost-effective use case in terms of effort-to-quality ratio.
The main lesson
"An AI assistant is not a Proxy PO. It's a PO amplifier. It multiplies your productivity on tasks you already master well. It doesn't compensate for gaps — it reveals them."
Areas where AI was useless or counter-productive: sponsor negotiations, culture-related team decisions, crisis management. These situations require human presence, organisational memory and situational empathy that no LLM possesses.