« Will your agent challenge mine? » — How an RTE and a PM co-built an AI-native practice in real time (Pluxee)

Pluxee is a global employee benefits and engagement company operating in 28 countries, with product teams organized in a SAFe Release Train. This is not a story from a consultancy or a lab — it is ours: two people on the same Release Train, facing a real problem together, who ended up somewhere neither of us had planned.

At the start of a new PI, we identified a concrete risk: not enough time and capacity to prepare the discovery topics needed for the next PI Planning. We had different instincts about how to respond. One of us started building a Rovo agent designed to challenge discovery content, support its formalization, and verify its readiness against Pluxee’s governance rules. The other was running a hands-on PoC with BMAD, an AI-assisted product framework being evaluated across the organization. We compared notes, shared feedback, and kept iterating — separately and together.
BMAD did not fully deliver on the quality we needed. So we doubled down on Rovo. With less than three weeks to be PI-ready, we each created an agent — built independently, with different architectures and different prompts.

And the topic we were racing to prepare? An extremely high potential feature : integrate Generative AI into an existing product, which would improve user experience and unlock major savings for Pluxee!

That is when the turning point happened. When one of us offered to give feedback on the other’s draft, the immediate reaction was: « You mean you’ll ask your agent to challenge the output of mine? » Neither of us had thought of it that way until that sentence was said out loud. It reframed everything.

What followed — a cross-review combining human judgment and two agents in conversation, a discovery artifact good enough to go to PI Planning, and a rethought model for how an RTE and a PM can work with AI under real pressure — is what we will unpack together in this session.

We will be honest about what did not work: the prompts that failed, the BMAD comparison, the moments where an agent confidently produced the wrong thing. And we will share the one transferable pattern that emerged — something any PO, PM, coach or RTE can try the week after the conference, without a budget or a data science team.
The story is still live. We will tell you how it ended in September.