CHAPTERS
Why financial crime is a top priority—and where AI can help
Stefano Amorelli introduces Qonto and frames financial crime as a massive global problem, with trillions laundered annually. He sets the goal: show how Qonto uses Claude in a security- and compliance-first way for sensitive, high-stakes investigations.
The investigation lifecycle today: alerts → manual casework
He walks through how a suspicious transaction becomes an alert and then a labor-intensive human investigation. Investigators must collect data from many sources, compile evidence, and make judgment calls with legal consequences.
Three AI layers: general chat, predictive systems, and agentic AI
Stefano distinguishes between simple document Q&A, the “native AI” in alerting (predictive ML + rules), and the opportunity: agentic AI to assist the second line of defense. He focuses on where genAI can add the most value—complex, multi-source investigations.
Model choice: why Opus 4.7 for long-context investigation work
He explains why Qonto selected Claude Opus 4.7: investigations require reasoning across large, scattered evidence. Long-context reasoning quality matters more than marginal cost savings in this high-stakes domain.
Benchmarking long-context reasoning: GraphWalks as a signal
Stefano highlights GraphWalks as a benchmark aligned with investigative work: connecting facts spread throughout a document rather than nearby snippets. He positions Opus 4.7 as leading for this kind of cross-document/cross-context linking.
Data access is the hard part: MCP meets security and compliance concerns
He pivots to the key challenge: LLMs are only useful if they can access the right data and tools. MCP unlocks this, but triggers valid concerns from security/compliance—so Qonto builds explicit boundaries and a secure harness around MCP.
The data-source mess in fincrime: OSINT, KYC/KYB, dashboards, and actions
Stefano describes the investigative reality: many internal/external systems, different formats, and even multimodal data. The system must unify access while controlling permissions and enabling selective automated actions.
Security-first architecture: Cowork plugin → MCP gateway → federated MCP servers
He presents the core architecture: Claude Cowork as the analyst UI, a central MCP gateway enforcing controls, and multiple downstream MCP servers that connect to internal/external APIs. The gateway becomes the control plane for authentication, authorization, and auditing.
Control-plane details: SSO/OAuth, PASETO tokens, RBAC, and audit trail
Stefano details the concrete mechanisms that make the system acceptable in regulated contexts. Short-lived PASETO bearer tokens, role-based access control, and comprehensive audit trails ensure access is restricted, attributable, and reviewable.
Implementation mechanics: Terraform policy, ContextForge gateway, Kubernetes deployment
He shows how the governance becomes operational: RBAC policies are versioned (Terraform in GitHub), the gateway uses an open-source base (ContextForge), and MCP servers run on Kubernetes—reachable only via the gateway. Tool calls are instrumented for observability and audit.
What analysts experience: Cowork dashboards, inline widgets, and guided actions
A demo illustrates the user impact: analysts can run an investigation and get an interactive, AI-generated dashboard in one interface. Instead of tab-sprawl across tools, Cowork provides consolidated views and actions to speed decision-making while keeping humans in control.
Operational monitoring: Grafana + ClickHouse for auditing and performance
He explains how Qonto monitors real usage in production: audit events, tool calls, and authorization flows are stored and visualized. This supports investigations, compliance reviews, and performance tuning.
Designing a complex Cowork plugin: orchestrator skills, XML prompts, and verification
Stefano shares prompting and plugin structure lessons: break prompts into modular skills, use an orchestrator to route tasks, and add a meta-skill to verify results. XML-structured prompts and explicit tool scoping improve reliability and efficiency.
Evals for trust: tool selection/order, grounding, and reasoning quality
He addresses the core stakeholder question—“Can we trust it?”—by outlining evaluation targets. Beyond correct final answers, they evaluate tool-use behavior, hallucination/grounding, and the quality of reasoning (often with LLM-as-judge).
Scaling adoption with governance: the MCP gateway as a company flywheel
He concludes with organizational impact: once the secure gateway and first plugins exist, other teams can quickly build new plugins and add MCP servers. This accelerates AI adoption while keeping audit, RBAC, and identity controls consistent.
Get more out of YouTube videos.
High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.
Add to Chrome