PayPal Compliance AI Assistant
An AI assistant that cut compliance analysis time by 90% — and uncovered hidden insights that helped Product ship better, faster, and within regulatory bounds.
// the problem
At PayPal UK, the compliance team was drowning in a weekly ritual: manually pulling data from fragmented, inconsistent sources, cleaning it up, and then analysing reviews and approval/rejection decisions on product and feature proposals submitted by Product and Engineering. The data was messy, the sources were siloed, and the analysis was slow. More importantly, by the time insights reached Product teams, the window to act on them was often closed.
// why this matters
Compliance isn't just a gate — it's a signal. Patterns in what gets rejected, what gets flagged, and what sails through contain real product intelligence. But that intelligence was buried in manual process and never surfaced in a form Product could act on. The cost wasn't just time lost. It was good products shipping late, and the occasional bad decision slipping through because no one had the bandwidth to look closely enough.
// how it works
An AI assistant built to ingest the messy weekly data feeds from multiple fragmented sources, clean and normalise them automatically, and surface structured analysis of reviews and approval/rejection patterns — including trends, anomalies, and flags that human reviewers had been missing. The output: a clean weekly report that replaced hours of manual work, and a pattern layer that made compliance insights actionable for Product teams rather than opaque to them.
// trade-offs i made
The biggest decision was what to optimise for first: speed of delivery or quality of analysis. I chose speed — a scrappy first version that handled 70% of the data cleaning reliably was more valuable than waiting for 100% coverage. The 30% edge cases were flagged for human review rather than forced through an imperfect pipeline. That decision to embrace a human-in-the-loop model rather than chase full automation was the right call: it built trust with the compliance team faster than a black-box solution would have, and the edge case patterns over time informed the next iteration.
The other significant trade-off: surfacing insights that hadn't previously been visible created new organisational questions about who owned them and how they fed back into Product planning. Good tooling generates good problems. That's a feature, not a bug.
// what i learned
90% reduction in time spent on data cleanup and analysis. Hidden compliance patterns surfaced for the first time. Product and Engineering teams able to ship better, compliant products faster — with earlier visibility into what would and wouldn't pass review rather than finding out at the gate.
The more important finding: the tool revealed how much intelligence had been sitting in the compliance data all along, invisible because no one had the bandwidth to look at it systematically. Automation didn't just save time — it changed what was knowable.
// what's next
A real-time dashboard rather than a weekly report. Predictive flagging — using the pattern history to score new proposals before they enter the review queue. And better feedback loops between the tool's outputs and the compliance team's evolving judgment, so the model improves with each cycle.