BCBS 239 — Risk Data Aggregation and Risk Reporting — Regulatory Reference

BCBS 239 does not explicitly reference AI agents, but its Principle 3 — accuracy and integrity of risk data — requires banks to demonstrate that every data input to risk reports is traceable, complete, and has not been altered without record. When AI agents read or aggregate risk data, each access becomes a contribution to the aggregation pipeline. Banks must be able to show regulators exactly which data an agent read, under what policy, and at what time. AutoPIL's source registry and tamper-evident audit chain satisfy this by logging every agent-level data access decision with a cryptographic record before any data enters the agent's context window.

BCBS 239 applies directly to global systemically important banks (G-SIBs) designated by the Financial Stability Board. National supervisors have also extended equivalent requirements to domestically systemically important banks (D-SIBs) in many jurisdictions, including the EU under EBA guidance. For AI deployments, this means any large institution subject to BCBS 239 must treat AI-driven risk data aggregation the same way it treats traditional ETL pipelines — with full lineage, integrity controls, and documented governance. Banks outside the G-SIB or D-SIB lists may still face equivalent obligations under their domestic stress-testing or risk reporting frameworks.

Principle 2 covers data architecture and IT infrastructure. It requires G-SIBs to design and maintain a data architecture that supports accurate risk data aggregation under normal and stressed conditions. For AI deployments, this means the underlying infrastructure governing what agents can access, and how, must be part of that documented architecture — not an ad hoc overlay. AutoPIL's agent registry and source registry satisfy this by making every AI agent a registered, governed entity with explicit policy bindings. The registry documents which agent accesses which data source, under which policy, creating the machine-readable data architecture map BCBS 239 supervisors expect to see during review.

AutoPIL addresses BCBS 239 data lineage requirements through two mechanisms. First, the source registry records every data source as a governed object with sensitivity classification, owner, and access controls — providing the data provenance layer Principle 3 requires. Second, the tamper-evident audit chain writes a cryptographic record of every agent access decision: which agent, which source, which policy version, ALLOW or DENY, and a hash-chained timestamp. This chain cannot be retroactively altered without detection, which supports the independent validation requirement under BCBS 239's supervisory review principles. AutoPIL policy IDs FS-BCBS239-P2-001 and FS-BCBS239-P3-001 map directly to these two principles and can be referenced in regulatory submissions.

BCBS 239 is supervised by national regulators, not enforced directly by the Basel Committee. In the EU, the ECB conducts BCBS 239 thematic reviews of significant institutions and has issued supervisory findings requiring remediation plans with defined timelines. Banks that cannot demonstrate data lineage or risk data integrity face Pillar 2 capital add-ons, public supervisory disclosure, and restrictions on stress-test model reliance. As AI-driven risk aggregation becomes more common, supervisors have signaled that AI-specific data flows must be subject to the same lineage documentation as traditional pipelines. Gaps in agent-level audit trails are now a documented finding category in ECB risk data reviews.