AI agents are only as effective as the data feeding them. This session brings together practitioners from Stripe, FinWise Bank, and Hummingbird to work through what it actually takes to build a reliable data foundation for AI in financial crime compliance, covering input quality, context engineering, governance, and how to make AI-adjudicated decisions defensible to regulators and banking partners alike.

Key takeaways

• Data quality is the ceiling, not the model: Feed a model richer data and the outputs improve immediately. The model itself is rarely the limiting factor in agent performance.

• Observability belongs at every stage: Track input data quality, agent decision steps and output behavior separately. Drift in any one of them can quietly degrade alert quality without a visible break.

• Agents need to know when to stop: Design systems to escalate or halt when required data is missing, because AI will fill gaps with something plausible if you let it.

• Data age matters, but context determines which data stays relevant: Historical transaction patterns can be valuable; stale KYC fields usually aren't. The question is whether the age of a given data point still serves the decision.

• Documentation is what saves you in an exam: If you can't explain your data selections, error rate thresholds and change management process, you won't have answers when regulators ask.

• Governance requires cross-functional ownership: Pre-deployment standards, testing requirements and post-deployment monitoring all need alignment across compliance and engineering. Neither can govern AI alone.