Top 10 Controls for Agentic AI in Financial Services
What regulators, standards bodies and incident post-mortems emphasize for agentic systems in regulated environments, and the artifact each control should leave behind.
Financial services carry a unique combination of fiduciary duty, strict regulators (SEC, FINRA, OCC, NCUA, FCA, MAS, FDIC, The Federal Reserve) and high-value targets. Agentic AI multiplies both the upside and the blast radius. The controls below cover what regulators, standards bodies and incident post-mortems emphasize for AI systems in regulated environments.
1. Written intent & scope declaration
Every agent must have a written and approved purpose or intent: what it can do, what data it touches, what systems it calls, what it must never do. This is the agent’s charter.
- Why: Supports what examiners expect under SR 26-2 and the EU AI Act Art. 9 in terms of model risk management; anchors every downstream control.
- Artifact: Intent manifest reviewed by model risk and compliance stakeholders.
2. Governance and accountability
Name an accountable human owner per agent. Inventory every agent in a central registry with its intent manifest, risk tier and categories, agent type and last approval date.
- Supports alignment with: 2023 Interagency Guidance on Third-Party Relationships: Risk Management (Federal Reserve, FDIC and OCC); DORA (EU); FFIEC IT Handbook.
- Registry fields: owner, reviewer, risk tier, risk category, agent purpose type, last red-team date, last review date, kill-switch status.
3. Human-in-the-loop gates for material actions
Define thresholds (dollar amount, customer impact) above which the agent must stop and get human approval.
- Example: An agent drafts a wire; a human releases it. The agent proposes a trade above $X; a human confirms it.
- Trap: Because humans have limited attention, too many approval requests lead to fatigue and mistakes. If escalating to a human too often, tighten policies so routine actions don’t need approval.
4. Govern least-privilege tool & data access
An agent’s permissions should match its declared intent. The identity and credential plumbing (OAuth, cloud IAM, SPIFFE, row-level security) lives in your IAM stack; the control is making sure what the agent can touch stays aligned with what it is supposed to do, and that drift gets caught.
- Baseline: Scoped, short-lived credentials per agent; no shared service accounts; separate credentials per environment.
- Governance layer: Compare declared intent against observed tool and data access; flag over-privilege and unused scopes; require sign-off for any expansion.
5. Policy-as-code on every tool call
Encode allow or deny rules (OPA/Rego, Cedar, or equivalent) in front of every tool the agent can invoke.
- Why: Rules in a prompt are suggestions and models are not perfectly aligned. In either case the model can ignore them. Only code in front of the tool can enforce a policy every time.
- Examples: PII egress, sanctioned counterparties, trading-window restrictions, cross-tenant isolation.
6. Immutable, reconstructible audit trail
Log every prompt, model response, tool call, input, output, policy decision and human approval (timestamped, hashed, masked, encrypted, and WORM-stored). Regulators will ask you to reconstruct a specific decision.
- Minimum fields: agent ID, user ID, intent, tool, inputs, outputs, policy verdict, model + version.
- Reconstruction: Reconstruct decisions from the logs. You should be able to walk through any past decision (inputs, tool calls, policy verdicts, outputs) and explain why it happened.
7. Red-team & adversarial evaluation
Run prompt-injection, jailbreak, data-exfiltration and tool-abuse evaluations as a gate any time an agentic workflow changes materially. Treat it like penetration testing: scheduled, scoped, signed off.
- Why: Every change in the model, tools, or prompts can change outcomes. New vulnerabilities emerge that need to be tested.
- Cadence: Every model change, every tool addition, at minimum quarterly.
8. Data classification & PII/NPI boundaries
Classify data flowing into and out of the agent (public, internal, confidential, NPI, PCI). Enforce at the retrieval and output layer. You should mask, redact or refuse, rather than hoping the model behaves.
- Required for: GLBA, GDPR, and state privacy laws.
- Examples: An agent taking screenshots of client data for a financial advisor writing a report, or an agent writing account numbers into an unapproved Slack channel.
9. Model & prompt version observability
Know which model and prompt version made each decision, run evals continuously, and be able to roll back or disable fast when something regresses.
- Baseline (all agents): Record model, prompt and tool-schema version on every decision; keep a change log; have a one-click rollback or kill path.
- High-risk tier (money movement, trades, customer-facing decisions): Pin versions, gate changes through regression evals, no silent auto-upgrades of production endpoints.
10. Continuous monitoring, drift & anomaly detection
Monitor live: refusal rate, tool-error rate, cost per task, latency, output distribution shift and policy-violation attempts. Alert on anomalies the way you would for a payments service.
- Kill switch: Every agent must have a one-click disable path that works without a deploy.
- Detection: Use a dedicated threat detection tool that can monitor unusual agent activity, alert the responsible team, and define when to stop the agent.
Alexis Yelton is co-founder and CPO/CAIO of TenetGraph. She has worked across AI, model training, and agentic systems. Before TenetGraph, she was AI Lead at ActZero and VP of Data and Machine Learning at Genvid Technologies. She holds a PhD and completed postdoctoral work in bioinformatics at UC Berkeley and MIT.