Client: a leading US benefits-administration fintech (anonymised)
My role: Senior Product Manager, GlobalLogic — leading discovery through engineering handoff
Shape of the product: an agentic AI “nudge” platform that helps FSA participants spend their pre-tax benefits before forfeiture deadlines
The problem
Every year, benefits participants forfeit unspent FSA balances — bad for participants, and a retention and satisfaction problem for the platform serving them. The client wanted an intelligent nudging capability: right message, right participant, right moment — at scale, across many plan configurations.
What made it hard
- Plan diversity. Spending rules, carryover provisions, grace periods, and transaction limits vary per plan — a nudge that's accurate for one plan is misleading for another. “Amount at risk” is a calculation, not a field.
- Message fatigue. Dozens of candidate nudge scenarios competed for a participant's attention. Send everything and you train people to ignore you.
- Proving it works. Without attribution, nobody can say whether nudges changed spending behaviour — which kills the business case.
What I did
- Ran structured discovery across 20+ candidate nudge journeys with client product leadership; drove prioritisation to a disciplined monthly cadence (a hard cap on sends per participant) rather than a fire-everything model
- Designed the product around a multi-agent architecture — separate agents for participant scoring, message generation, and outcome attribution — and defined how they hand off to each other
- Pushed for a closed learning loop as a first-class requirement: campaign outcomes feed back into targeting recommendations, with a human-review step the client can toggle, and configuration history retained so unsuccessful changes can be reversed
- Owned the PRD, use-case scenarios, user flows, and orchestration documentation; resolved 35+ rounds of annotated client feedback into a single aligned spec for engineering handoff
Outcome
Discovery converted into an engineering-ready, client-signed specification — with prioritisation logic, agent responsibilities, and success metrics (delivery, open, click, and spend-impact reporting) defined before a line of production code.
What I'd tell another PM
Agentic AI products live or die on the boring parts: data availability per use case, attribution design, and the review loop between AI recommendations and human owners. Model choice was the least contested decision in the whole engagement.