Audit & compliance for agentic finance.
Autonomous systems do not enter a regulatory vacuum — they enter a dense existing regime. The working question for any deployment is not "is this allowed?" but "which of our standing obligations does this system now carry, and where is the evidence?" This page maps the major anchors.
SR 11-7: the incumbent discipline
The Federal Reserve's SR 11-7 guidance (with OCC 2011-12) has governed model risk at U.S. banking organizations since 2011, and its three pillars translate directly to agents: robust development and documentation, independent validation with effective challenge, and ongoing monitoring through the model's life. What agents change is scope — an agent is a model that acts, so validation must cover behavior and constraint-adherence, not just predictive accuracy. Supervisors have signaled consistently that AI systems in decision roles fall inside existing model-risk expectations, not outside them.
The EU AI Act: the extraterritorial layer
The EU AI Act entered into force in August 2024 with obligations phasing in through 2026–27. AI systems used in certain financial contexts face requirements that read like a regulatory restatement of the oversight framework: risk-management systems, data governance, technical documentation, automatic event logging, human oversight, and demonstrated accuracy and robustness. Any institution serving EU clients or markets should assume the documentation bar it sets becomes the de facto global baseline — as GDPR's did.
Recordkeeping: the oldest obligation is the sharpest
Books-and-records regimes (SEC 17a-3/4, CFTC equivalents, MiFID II) already require reconstructing how orders came to be. When an agent originates the order, the record must include what the agent knew, which constraints it evaluated, and which version of it acted — the tamper-evident evidence trail of Control 5. The practical standard: append-only, integrity-protected, and reconstructable by someone who did not build the system.
What an audit-ready deployment produces
- A validation dossier: rationale, testing record (including failed variants), boundary-scenario results, and sign-off by an independent reviewer with authority to say no.
- A constraint register: every runtime limit, where it is enforced, and proof it binds outside the agent's own process.
- The kill book: criteria, owners, response times, and dated drill records.
- Drift dashboards with threshold history and every graduated response taken.
- Hash-chained decision logs mapping each action to code, config, and data versions.
References: Federal Reserve SR 11-7 / OCC Bulletin 2011-12; Regulation (EU) 2024/1689 (AI Act) public texts and Commission implementation timeline; SEC Rules 17a-3/17a-4; MiFID II recordkeeping provisions. Editorial reference only — not legal or compliance advice.