What it is
Reference architecture for finance-vertical Claude integrations. The interesting bits are the patterns for compliant LLM use in regulated domains — exactly the territory of an AI-Act-aware GRC platform.
What’s worth studying
- Privileged-source system prompts. Account scoping, allowed actions, and customer-tier risk classification are rendered into the system prompt by the platform, never by user chat. Same pattern as Strix’s scope injection.
- Structured outputs as the default. Compliance-relevant content is constrained to JSON schemas, not free-form prose. Less room for the model to drift.
- Conservative tool surface. The agent can read more than it can write; writes route through human approval. Maps directly onto Article 14 (human oversight).
Drill-down
The full per-doc analysis lives below — these are the original numbered analyses, rendered as styled HTML. Pick a section to study deeper.