What Good Third-Party AI Risk Management Looks Like Under OCC 2023-17

The Five Visible Properties of Good Third-Party AI Risk Management

A community bank CEO with multiple AI vendor relationships has heard the framework: third-party risk management applies to AI. The framework is clear; the operational implementation varies. The CEO needs to know what good looks like in practice.

What follows is the plain-English version of the five visible properties of good third-party AI risk management.

Property 1: A Complete AI Vendor Inventory

The bank can produce, on demand, a list of every AI vendor relationship. The list includes:

• Vendor name, AI capability provided, business function served

• Risk tier

• Contract terms relevant to AI (right-to-audit, attestation refresh, indemnification)

• Diligence status and date

• Last review and exercise of right-to-audit

• Subservicer relationships within the vendor's stack

Inventory completeness is the foundation. Banks with incomplete vendor inventories cannot demonstrate the rest of the program.

Property 2: Onboarding Diligence That Goes Beyond Standard Third-Party Review

Onboarding diligence for AI vendors goes beyond standard third-party diligence. AI-specific elements include:

• Vendor's model risk management practices

• Training data sourcing and refresh

• Fairness testing methodology and results

• Explainability output and documentation

• Vendor's incident response when AI misbehaves

• Subservicer transparency

The bank produces its own written assessment of these elements, not just vendor representations.

Property 3: Contracts With AI-Specific Terms

Vendor contracts include AI-specific provisions:

• Right-to-audit specific to AI model and operation

• Vendor obligation to disclose material model changes

• Attestation refresh schedule

• Vendor's compliance obligations including state and federal AI requirements

• Indemnification carveouts and limits

• Exit transition support if relationship ends

Contracts without AI-specific terms produce the most common examination weakness.

Property 4: Ongoing Monitoring That Produces Retained Evidence

The bank monitors vendor AI on a documented cadence:

• Annual review of vendor performance against contract terms

• Right-to-audit exercise every two to three years for high-risk vendors

• Attestation review when vendor pushes material model changes

• Continuous monitoring of vendor's regulatory and operational status

• Documented response to identified concerns

• The evidence supports examiner inquiries spanning 12 to 24 months.

Property 5: A Documented Exit Plan for Each Material AI Vendor

The bank has a documented plan for ending each material AI vendor relationship. The plan addresses:

• Replacement capability identification

• Transition timeline

• Data migration and disposition

• Model re-validation if replacement is selected

• Operational continuity during transition

Exit planning is often the weakest element across community bank programs.

What Separates Substantive AI Vendor Oversight From Administrative Compliance

The five properties above can be operated administratively or substantively. The distinction matters at exam time.

Administrative operation captures the activities formally. Inventories exist; contracts are signed; reviews occur. The substance is thin.

Substantive operation produces evidence that the activities affect the bank's decisions and risk posture. Reviews identify concerns; concerns produce changes; changes are documented.

Banks that pass exam scrutiny operate the program substantively. Banks that face findings often have administrative versions of each property.

What the CEO Controls in Third-Party AI Risk Management

The CEO controls four things in third-party AI risk management.

The first is the program ownership clarity. The vendor management function, the model risk function, and the CRO's office all touch AI vendor oversight. Clear ownership prevents gaps.

The second is the investment level. Substantive program operation requires investment in personnel and external partner support. Under-investment produces administrative-only operation.

The third is the contract discipline. The CEO can require new and renewing vendor contracts to include AI-specific terms.

The fourth is the board engagement. AI vendor risk is a board-level topic, particularly for material vendor relationships.

Three Things to Work Through Before the Third-Party Risk Exam

A community bank CEO should work through three things on third-party AI risk management. The first is a current-state assessment against the five properties. The second is a remediation plan addressing identified gaps with sized investment. The third is contract policy update for new and renewing AI vendor agreements.

When this discipline is in place, the third-party risk examination moves from "explain your AI vendor oversight" to "review the program's most recent outputs," and the gap is the gap between findings and routine confirmation.

Operate the Five Properties — That's What OCC 2023-17 Actually Expects

A community bank CEO operating third-party AI risk management is operating the program OCC 2023-17 explicitly expects. The five properties named here are the documented patterns of substantive operation.

If your bank has not assessed third-party AI risk management against the five-property framework in the last twelve months, that is the conversation worth having with your Tech-Operations partner before the next exam cycle.

Five Nines Technology Group is a Tech-Operations partner for community banks and credit unions. Translating regulatory frameworks into operating discipline at community bank scale is where our team focuses.