Resources
Examiner-Ready Agentic Banking: An AI Agent Governance Checklist for Credit Unions and Community Banks
No U.S. banking rule is written specifically for AI agents yet, so examiners can be expected to apply existing third-party risk and internal control guidance. Be ready to show an inventory of each agent and its permitted actions, board-approved limits, vendor due diligence, human approval points, a complete audit trail, and a tested way to shut it off.
This checklist supports exam preparation. It is not legal or regulatory advice.
What regulators have said about AI agents so far
There is no AI-agent-specific rulebook yet. Institutions preparing for exams lean on existing third-party risk and internal control guidance, plus what regulators have published about generative and agentic AI scope.
- On April 17, 2026, the OCC, Federal Reserve and FDIC issued updated interagency model risk management guidance, rescinding the 2011 guidance (OCC 2011-12). The guidance states: "Generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance." It says it is expected to be most relevant to organizations over $30 billion in total assets, that it does not set enforceable standards, and that non-compliance will not result in supervisory criticism. The agencies plan a request for information on model risk management, including banks' use of AI, generative AI and agentic AI. Source: OCC Bulletin 2026-13.
- That guidance's "model" definition excludes deterministic rule-based processes and software without statistical, economic or financial theories. That is useful context when a policy layer sits outside the AI model; it does not exempt any product from other applicable guidance. Source: OCC Bulletin 2026-13.
- Interagency Guidance on Third-Party Relationships: Risk Management (June 2023) applies to all third-party relationships. It is organized around a life cycle: planning, due diligence, contract negotiation, ongoing monitoring and termination. It says rigor should match risk and criticality. Sources: OCC Bulletin 2023-17; Federal Reserve SR 23-4; FDIC FIL-29-2023.
- The OCC's Community Bank Guide to Third-Party Risk Management (May 2024) gives worked considerations and examples for each life-cycle stage. Source: OCC Bulletin 2024-11.
- NCUA's AI resource page lists 07-CU-13 and 01-CU-20 as resources for due diligence on AI vendors. It names challenges beyond traditional vendor management: algorithmic decision-making, fair lending, member data privacy, operational resilience and model risk. Source: NCUA AI resource page.
Why an AI agent that acts needs more than a chatbot review
An agent that can execute transfers or change records raises the criticality of the vendor relationship. Chatbots that only answer questions present a different risk profile than agents that complete transactions. Interagency third-party risk management guidance says rigor should match risk and criticality (OCC Bulletin 2023-17). Require stronger controls, a deeper audit trail and tighter contracts when the agent can move money. See controls to require before an AI agent moves money and nine ways AI systems in banking get exploited, including prompt injection and invented actions.
The exam-readiness checklist
Work through each group below. Keep the evidence named for each item. Where a row cites regulatory guidance, the source is listed with a link.
1. Inventory and scope
- List every AI agent in pilot or production, with a named owner.
Evidence to keep: Agent inventory spreadsheet or register
- Document the actions each agent can take and the systems it touches.
Evidence to keep: Action and system map per agent
- Record which customers or members each agent serves and any channel limits.
Evidence to keep: Scope and channel description
- Note whether the agent can execute transfers or change records, or only answer questions.
Evidence to keep: Criticality rating tied to permitted actions
2. Board and management oversight
- Board-approved use of AI agents, including risk appetite and dollar or action limits. (NCUA 07-CU-13)
Evidence to keep: Board minutes or policy approval
- Named accountable owner for AI agent risk (management), with a reporting cadence to the board or supervisory committee. (Interagency TPRM guidance)
Evidence to keep: RACI or ownership memo; reporting calendar
- Clear escalation path when an agent fails, invents an action, or hits a limit. (NCUA 07-CU-13)
Evidence to keep: Incident and escalation procedure
3. Vendor due diligence
- Financial condition and business continuity review of the vendor. (NCUA 01-CU-20)
Evidence to keep: Financial review file
- Independent control reports (for example SOC 2 Type II) reviewed, not filed unread. (NCUA 07-CU-13)
Evidence to keep: SOC 2 Type I and Type II review notes
- Security, data handling, and training commitments, including whether customer data is used to train AI models. (Interagency TPRM guidance)
Evidence to keep: Security questionnaire and written commitments
- Subcontractor map and incident history. (OCC Community Bank Guide)
Evidence to keep: Fourth-party list and incident summary
4. Contract terms
- Data ownership, return, and deletion on exit. (Interagency TPRM guidance)
Evidence to keep: Signed contract clauses
- Audit rights for the institution and its examiners or auditors. (Interagency TPRM guidance)
Evidence to keep: Audit and access clauses
- Incident notice timelines and severity definitions. (Interagency TPRM guidance)
Evidence to keep: Incident notification schedule
- Change notification for model or policy updates that affect the institution. (Interagency TPRM guidance)
Evidence to keep: Change-notification clause
- Exit and termination plan, including how to shut the agent off and recover data. (Interagency TPRM guidance)
Evidence to keep: Termination and transition plan
5. Controls on what the agent can do
- Authenticated request only: the agent acts only when an authenticated customer or member asks.
Evidence to keep: Architecture description; session handling
- Policy limits enforced outside the AI model (immutable policy layer).
Evidence to keep: Policy configuration demo or screenshots
- Human approval thresholds for higher-risk amounts or new payees.
Evidence to keep: Approval matrix
- Institution-held kill switch with a named owner and a tested response time.
Evidence to keep: Kill-switch runbook and test log
For control design detail, see controls to require before an AI agent moves money.
6. Audit trail and evidence
- Each action logged with the request, AI interpretation, policies evaluated, decision, and exact transaction executed.
Evidence to keep: Sample audit export covering the full decision chain
- Logs are immutable and available to compliance and internal audit in real time or on short notice.
Evidence to keep: Access procedure and retention settings
- Ability to reconstruct who asked, what the agent understood, and what executed for any sampled case.
Evidence to keep: Reconstruction walkthrough for internal audit
For a failure example when the trail is missing, see The Audit Ghost.
7. Testing, monitoring and change control
- Pre-launch testing, including prompt injection and invented-action scenarios.
Evidence to keep: UAT test results
- Ongoing monitoring of agent activity and exceptions.
Evidence to keep: Monitoring dashboard access and review cadence
- Change log for AI model or policy updates that affect production behavior.
Evidence to keep: Change log and approval records
- Periodic review of limits, inventory, and vendor performance; incident response plan tested. (Interagency TPRM guidance)
Evidence to keep: Periodic review minutes; IR test summary
Pre-launch tests should include scenarios from nine ways AI systems in banking get exploited.
Questions to ask an AI agent vendor before signing
Use these vendor-neutral questions across the third-party risk management life cycle. For category-level comparison rather than named vendors, see compare agentic AI vendors by category.
Planning
- Which transactions and servicing tasks does your agent complete end to end in production today?
- How do you classify criticality for an agent that can move money or change records?
Due diligence
- Where are limits, payee rules and approval thresholds enforced, and can the AI model change them?
- What does the audit trail record for each action, and can we export it for examiners?
- Is our data isolated from other institutions? Is customer data used to train AI models?
- Which independent control reports do you hold (for example SOC 2 Type II), and can we review them?
- Which subcontractors touch our data or our customers?
Contract negotiation
- Who owns our data and conversation history, and how do we get it back if we leave?
- What are incident notice timelines, and what counts as a security or operational incident?
- How do you notify us of model or policy changes that affect our production agents?
Ongoing monitoring and termination
- Who can pause or shut off the agent, and how fast does the kill switch take effect?
- What does an orderly exit look like, including data return and agent shutdown?
What evidence to have ready for an exam
Keep a working file you can hand to internal audit or an examiner without rebuilding it under time pressure. Cadence below is suggested, not prescribed by a rule.
| Evidence item | Who owns it | Where it lives | Refresh cadence (suggested) |
|---|---|---|---|
| AI agent inventory and permitted actions | IT / vendor management | Vendor management file or GRC tool | After every new agent or scope change |
| Board approval of use, limits and risk appetite | Management; board or supervisory committee | Board packet and minutes | At approval; at least annually |
| Vendor due diligence package (financial, legal, SOC 2, security) | Vendor management | Third-party risk file | Before signing; on renewal or material change |
| Contract with audit, incident, change and exit terms | Legal / vendor management | Contract repository | At signing; when amended |
| Policy configuration and approval thresholds | Compliance / risk | Paygent Central or equivalent control console | At launch; when policies change |
| Sample and live audit trail exports | Compliance / internal audit | Audit export store | Sampled monthly or per exam request |
| Kill-switch owner, runbook and test results | Operations / IT | Operations runbook | Tested at least annually |
| UAT and prompt-injection test results | IT / risk | UAT evidence folder | Before go-live; after material model or policy changes |
Guidance at a glance
Only documents from the facts checked for this page. Open the source before you rely on a paraphrase.
| Document | Issuer | Date | What it says about AI or third parties | Source |
|---|---|---|---|---|
| Interagency model risk management guidance (OCC Bulletin 2026-13) | OCC, Federal Reserve, FDIC | April 17, 2026 | States generative and agentic AI models are not within its scope; expected to be most relevant to organizations over $30 billion in total assets; does not set enforceable standards; agencies plan an RFI on AI including agentic AI. | View source |
| Interagency Guidance on Third-Party Relationships: Risk Management | OCC, Federal Reserve, FDIC | June 2023 | Applies to all third-party relationships; life cycle of planning, due diligence, contract negotiation, ongoing monitoring and termination; rigor should match risk and criticality. | View source |
| Community Bank Guide to Third-Party Risk Management | OCC | May 2024 | Worked considerations and examples for each TPRM life-cycle stage for community banks. | View source |
| Artificial Intelligence (AI) resource page | NCUA | Ongoing | Lists 07-CU-13 and 01-CU-20 for AI vendor due diligence; names challenges such as algorithmic decision-making, fair lending, member data privacy, operational resilience and model risk. | View source |
| 07-CU-13 Evaluating Third-Party Relationships | NCUA | Letter to Credit Unions | Officials are responsible for planning, directing and controlling the credit union; risk assessment and due diligence for third parties; institution remains responsible; independent control reports help evaluate vendor controls. | View source |
| 01-CU-20 Due Diligence Over Third-Party Service Providers | NCUA | Letter to Credit Unions | Officials should require a due diligence review before entering any third-party arrangement, including financial review and legal review of contracts. | View source |
How Payman AI supports exam readiness
Payman AI helps you show evidence for the control, audit and monitoring checklist groups. It does not make an institution compliant by itself.
- Paygent Central provides real-time monitoring, execution traces and audit trails, configurable risk controls and spend limits, an emergency kill switch and a UAT environment.
- Controls map to checklist groups 5 to 7: authentication-gated actions, an immutable policy layer outside the model, a complete audit trail (request, AI interpretation, policies evaluated, approved or blocked, exact transaction), physical data isolation and a bank-held kill switch. Overview: Agentic Banking.
- Customer data is not used to train AI models. Details: Security.
- Payman AI has SOC 2 Type I and SOC 2 Type II reports. See Security and the Trust Center.
- Payman AI integrates with your core and digital banking systems rather than replacing them.
Frequently asked questions
Is there a federal rule specifically for AI agents at banks or credit unions?
Not as of October 2026 based on the sources cited on this page. The April 17, 2026 interagency model risk management guidance, OCC Bulletin 2026-13, states that generative and agentic AI models are not within its scope, and the agencies plan a request for information on model risk management including banks' use of AI, generative AI and agentic AI. Check those sources for updates; this page does not forecast future rules.
Does the 2026 model risk management guidance apply to agentic AI?
The guidance states: "Generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance." It also states it is expected to be most relevant to organizations over $30 billion in total assets, and that it does not set enforceable standards; non-compliance will not result in supervisory criticism. Source: OCC Bulletin 2026-13.
What will NCUA examiners look at for an AI vendor?
This page does not predict exam questions. NCUA's AI resource page points credit unions to 07-CU-13 and 01-CU-20 for due diligence on AI vendors. Those letters cover risk assessment, due diligence, financial review, legal and contract review, and monitoring of controls. Be ready to show that work product.
What should an AI agent audit trail record?
For each action: the request in the customer's or member's words, how the AI interpreted it, which policies were evaluated, whether it was approved or blocked, the exact transaction executed, plus timestamp and identity. The trail should be immutable and available to compliance. Payman AI's decision chain is described at Agentic Banking.
Which due diligence questions matter most for an AI agent vendor?
Start with: where rules are enforced (inside or outside the model); how each action is logged for examiners; how you shut the agent off; whether customer data is isolated and whether it is used to train AI models; and how you are notified of model or policy changes. Use the full vendor question list on this page, and compare categories in the vendor comparison guide.
Who should own AI agent risk in a community bank or credit union?
The board approves risk appetite and use; management names an accountable owner; compliance, IT and vendor management share oversight. NCUA 07-CU-13 states that officials are responsible for planning, directing and controlling the credit union's affairs and remain responsible for third-party relationships. Interagency third-party risk management guidance describes board and management responsibilities across the life cycle.
Can we use a vendor's SOC 2 report as our due diligence?
It is one input, not the whole review. Interagency third-party risk management guidance and NCUA 07-CU-13 describe independent control reports as part of evaluating vendor controls. Pair them with financial review, legal review, security and data-handling commitments, and ongoing monitoring. Payman AI has SOC 2 Type I and SOC 2 Type II reports; see Security and the Trust Center.
Next step
Walk through the checklist with compliance, IT and vendor management, then see how Payman AI records the evidence trail in a live demo.