AI Agent Governance for Banks, Asset Managers, and Fintechs
Financial services firms operate under overlapping regulations (SOX, OCC guidance, FFIEC, MAS TRM, PRA) that require provable controls on any system that touches financial data or executes transactions. AI agents are no exception.
What Goes Wrong Without Controls
A mid-size bank deploys 12 agents to automate trade reconciliation, compliance reporting, and customer onboarding. Within three months:
- An agent processes a batch of 400 reconciliation entries without approval — one entry contains a $2.3M discrepancy that propagates to downstream reports
- An auditor asks for evidence that agents only accessed approved data sources during Q3 — the team has application logs but nothing the auditor can independently verify
- A model provider outage causes an agent to retry with a different model that has not been approved for production use — no policy prevented the fallback
Each of these is a control failure that produces a finding. Proof-first architecture prevents all three.
Regulatory Mapping
| Regulation / Standard | Requirement | Claw EA Control |
|---|---|---|
| SOX (ITGC) | Access controls, change management, monitoring | Scoped tokens, approval gates, proof bundles |
| OCC Bulletin 2023-35 | Third-party risk management for AI/ML | Gateway receipts prove which models were called and under what policy |
| FFIEC Handbook | IT audit, business continuity, operations | Tamper-evident logs, kill switch |
| MAS TRM | Technology risk management (Singapore) | Egress allowlist, two-person rule |
Recommended Control Stack
Budget Controls
Per-run and per-day spending limits enforced at the policy layer. Prevents runaway costs and unauthorized financial exposure.
Approval Gates
Human-in-the-loop sign-off before high-value actions execute. Receipt proves the approval occurred and who approved.
Tamper-Evident Logs
Append-only, hash-linked audit trail. Modification of any entry invalidates the chain. Auditor-verifiable offline.
Scoped Tokens
Capability tokens bound to identity groups. Agents can only access systems their token permits.