Revolut · 2025
Agentic Platform
Humans supervising autonomous agents inside regulated financial workflows.
The problem
Defining interaction patterns for how humans supervise and collaborate with AI agents executing autonomous multi-step tasks in high-stakes financial workflows.
What I shipped
Designed interaction patterns for human-AI supervision over autonomous multi-step agents at Revolut. Focus areas: agent definition UX, plan inspection, intervention/handoff, audit trails, and graceful failure. Patterns shipped influenced Revolut's broader agentic platform for financial workflows. Also enhanced KYC with AI + Selfie 3.0 → −20% onboarding time while remaining compliant.
What I'd change
The supervision patterns assumed operators would read an agent’s plan before approving it; in practice they skimmed. I’d invest earlier in risk-weighted summaries — surfacing only the steps where an agent deviates from precedent, not the whole plan.