"Vendor risk assessment AI" means two different jobs, and a U.S. financial institution usually needs both. The first is assessing vendors whose products contain AI, such as an underwriting engine, a fraud model or a document-analysis tool. The second is using AI to make your own vendor risk process faster and more continuous. This guide covers both, with the regulatory frame that applies to banks and credit unions, a practical assessment structure, and the evidence you should demand before a model touches a member or borrower decision.
TL;DR
- Outsourcing a function to an AI vendor does not outsource the accountability. Under the 2023 Interagency Guidance, the institution keeps responsibility for risk management across the full third-party lifecycle.
- Tier the diligence by use case, not by vendor reputation. An autonomous credit decision deserves far more scrutiny than a drafting assistant.
- Ask for evidence, not assurances: validation reports, independent audit results such as SOC 2, and sample outputs.
- Include fourth-party exposure. A secure vendor can still depend on foundation model providers and data aggregators you never contracted with.
- Use AI inside the process too, but keep named human owners and independent validation on every go/no-go decision.
Automate Regulatory Model Risk Governance
Examine models against 32 qualitative criteria and resolve risk Tiers with pre-deployment checklists per OCC 2011-12 guidelines.
Table of Contents
- Why AI vendors change the assessment
- What U.S. regulators expect
- A five-domain assessment for AI vendors
- Tier by use case before you send a single question
- The traps that sink AI vendor reviews
- Contract terms that matter for AI
- Using AI to run the vendor risk process itself
- Our perspective: assess us the way you would assess anyone
- Where RiskInMind fits
- A starting checklist for your next review
- FAQ
- Sources
Why AI vendors change the assessment
A traditional vendor review asks whether a provider is financially stable, secure and operationally resilient. Those questions still matter. AI adds a layer: the vendor's product may make or shape decisions, its behavior can drift after deployment, and its inputs and training data may flow through parties you cannot see.
The concern is already measurable. In the Venminder 2025 State of Third-Party Risk Management survey, 49% of respondents reported a third-party cyber incident in the prior 12 months, and organizations had begun adding AI-usage language to vendor contracts. For smaller institutions the pressure is specific: the Ncontracts 2025 TPRM survey notes that most institutions are now monitoring vendor AI usage in some form.
Regulators frame this as relationship risk more than technology risk. The practical consequence: if a vendor's model contributes to an underwriting decision, that model belongs in your own model risk management inventory, validation program and ongoing monitoring, whoever built it.
What U.S. regulators expect
The interagency lifecycle
The June 2023 interagency guidance on third-party relationships, issued by the Federal Reserve, FDIC and OCC, replaced earlier agency-specific guidance with one risk-based framework. The OCC bulletin describes five stages:
- Planning
- Due diligence and third-party selection
- Contract negotiation
- Ongoing monitoring
- Termination, with documented exit strategies
The guidance does not carve out a separate AI track. It scales expectations to the criticality of the activity, which is exactly why use-case tiering matters: a vendor model that drives credit decisions is a high-risk activity and warrants rigorous, independent validation of its limitations and outputs.
Credit unions and the NCUA
If you are a federally insured credit union, the NCUA maintains a page of resources for due diligence on AI vendors. It points to its third-party letters (07-CU-13 and 01-CU-20) and names the challenges that go beyond ordinary vendor management: understanding algorithmic decision-making, fair lending compliance, member data privacy, operational resilience and model risk. Examiners will expect your file to address those themes by name.
Treasury and NIST as the technical backdrop
The Treasury report on AI-specific cybersecurity risks in financial services advises institutions to expand their usual due diligence when a vendor offers AI systems or relies on them, and includes a section on the questions to ask. The NIST AI Risk Management Framework supplies the governance vocabulary: successful AI risk management is an operating model with clear accountability, not a checklist.
A five-domain assessment for AI vendors
General security questionnaires miss the AI-specific failure points. Structure your review around five domains, and write the questions so the answer is a document, not a yes.
| Domain | What to establish | Evidence to request |
|---|---|---|
| Data privacy and retention | What data the model sees, where it is stored, whether it trains on your data, how it is deleted | Data flow diagram, retention schedule, contractual no-training clause |
| Model training, validation and performance | Training data provenance, intended use, known limitations, drift monitoring | Validation report, performance metrics by segment, bias testing results |
| Information security | Access control, encryption, incident response, penetration testing | Current SOC 2 report, pen test summary |
| Technology integration | How the tool connects to your core, LOS and document systems, and which sub-service providers sit behind it | Architecture diagram, sub-processor list |
| Legal and regulatory compliance | Fair lending, adverse action explainability, record retention, audit rights | Policy documents, explainability samples, audit-right language |
The domain structure follows a widely used AI vendor questionnaire framework, but the evidence column is where your institution earns its examination result.
Assurances versus evidence
The FS-ISAC generative AI vendor evaluation guide is built around a workbook that financial institutions can use to evaluate and qualitatively score generative AI vendors. Practitioners draw the same line it implies. Vendor documentation is necessary but rarely sufficient. A statement that the model is unbiased is an assurance. A bias test with methodology and results is evidence. A SOC 2 report is evidence about controls, but it does not validate model performance, so do not let it stand in for validation.
Tier by use case before you send a single question
The most common waste in AI vendor review is giving a low-risk writing assistant the same forty-page questionnaire as a credit decisioning engine, or the reverse. Tiering fixes this.
A workable approach asks three questions about each use case:
- Decision autonomy. Does the model decide, recommend, or only draft for a human to review?
- Impact. Does the output affect credit, pricing, fraud holds, account access or regulatory reporting?
- Data sensitivity. Does it handle nonpublic personal information, financial statements or bank statements?
Autonomous, high-impact, sensitive-data uses go to the top tier: full five-domain review, independent validation, and tight contract terms. Human-in-the-loop drafting tools with no member data can sit in a lighter tier with annual attestation.
If you want a quick way to place a vendor model on the model risk spectrum, our OCC 2011-12 model risk tier calculator is built for exactly that first-pass triage.
A worked example
Imagine a community bank evaluating two tools in the same quarter. The first summarizes board packets for staff. It handles internal documents, a human reads every output, and nothing it writes reaches a customer. The second analyzes applicant bank statements and produces a cash-flow score that a loan officer sees before approving a small business line. The first lands in a light tier: security attestation, a data-handling clause and annual review. The second lands in the top tier: validation evidence, bias and explainability testing, monitoring for drift, an owner in credit, an owner in technology and an owner for the data, and an exit plan. Same vendor category, entirely different diligence. That is what the guidance means by scaling to criticality.
The traps that sink AI vendor reviews
Self-approval
When the business unit that wants the tool also signs off on its risk assessment, the review becomes a formality. Effective programs separate the sponsor from the challenger. Independent validation can come from your internal model risk management function or an outside audit firm, but someone who did not choose the tool must test its intended use, limitations and performance data.
Fourth-party blind spots
Your vendor may be well controlled and still depend on a foundation model provider, an API integrator or a data aggregator. The Treasury report on AI cybersecurity risk is the best public reference on why this matters. Ask vendors to disclose sub-processors and the testing standards those parties follow, and consider requesting an inventory of the AI components in the product, similar in spirit to a software bill of materials. Then write notification duties for sub-processor changes into the contract.
One-time questionnaires
Models change after go-live. Data shifts, vendors retrain, and providers swap underlying models. The EY 2025 Global TPRM Survey reports that 31% of organizations cite AI and machine learning capabilities for enhanced due diligence as their top investment driver, and the broader argument is that AI should deliver continuous insight rather than only automate form-filling. Ongoing monitoring is a lifecycle stage in the guidance, not an optional extra.
Missing exit plans
Termination is the stage most files skip. For an AI vendor, an exit plan covers data return and deletion, model decommissioning, a fallback process if the tool disappears overnight, and who holds the decision records generated while it ran. Test the fallback before you need it.
Contract terms that matter for AI
As the Venminder data shows, institutions have begun adding AI-specific language to vendor contracts. The clauses worth negotiating:
- A prohibition on training the vendor's or a sub-processor's models on your data, unless you agree in writing
- Advance notice of material model changes, retraining or changes of underlying foundation model
- Audit and information rights, including access to validation results and sub-processor lists
- Incident notification windows that cover model failures and data exposure, not only breaches
- Record retention and exportability of decision logs and audit trails
- Exit assistance, data return and certified deletion
Your counsel should draft the final language. The point for the risk team is to know which evidence the contract guarantees you can obtain later.
Using AI to run the vendor risk process itself
The second meaning of the keyword is applying AI to third-party risk management. The realistic benefits are in the repetitive work: reading SOC 2 reports and policy documents, mapping questionnaire answers to your control framework, spotting inconsistencies between what a vendor says in different documents, and watching for changes between annual reviews.
Three cautions keep this useful rather than risky:
- Treat the AI tool as a vendor too. The system that scores your vendors is itself a third-party relationship and needs the same tiering and evidence.
- Keep a human decision-maker. AI can draft the finding, but a named reviewer approves the risk rating.
- Log everything. If an examiner asks why a vendor was rated moderate, you need the inputs, the output and the person who accepted it.
For a wider view of how institutions are applying these tools, see our overview of third-party risk management for U.S. financial institutions and our AI risk management best practices.
Our perspective: assess us the way you would assess anyone
We build an AI risk platform for credit unions, community banks and lenders, so we sit on both sides of this conversation. Institutions evaluate our software the same way we would urge them to evaluate any AI vendor, and we think that is correct.
Three convictions shape how we build. First, an AI vendor should be able to show where your data goes. Our positioning is in-house AI automation with no third-party data exposure, and we hold SOC 2 certification, but we would never ask you to take either on faith: request the report and test the claim. Second, speed and control are not opposites. Our workflows return results in under half a second, yet every output sits behind granular audit trails so that a reviewer can reconstruct what happened and why. Third, specialized beats generic. Our agents are each focused on a defined task, such as regulatory compliance or credit risk assessment, and coordinated by a central AI director we call Ava, which makes it easier to say what each component is responsible for and to validate it.
The standard we hold ourselves to is simple: a model that influences a lending decision must be explainable to your examiner, validated independently of the people who sell it, and replaceable without stranding your data.
— Raj
Where RiskInMind fits
If your assessment finds that you need AI inside underwriting, compliance monitoring or portfolio oversight, we built the platform for that environment. It covers loan application analysis, credit memo generation, CECL modeling, fraud detection for documents and regulatory reporting, with audit-ready outputs. Plans (Starter, Professional and Enterprise) are described on our pricing page, and a demo is the fastest way to put our controls through your own diligence questions. Our API documentation is public for technical reviewers.
A starting checklist for your next review
Use this order of operations the next time an AI vendor lands on your desk:
- Define the use case and place it in a tier.
- Name the business, technical and data owners, plus an independent challenger.
- Send the five-domain request, asking for documents rather than yes/no answers.
- Map sub-processors and foundation model dependencies.
- Add the vendor's model to your MRM inventory if it influences decisions.
- Negotiate AI-specific contract terms and the exit plan.
- Set monitoring triggers: performance drift, model change notices, incidents, audit report renewals.
- Calendar the reassessment, and document why each go/no-go decision was made.
For related reading, our guides on regulatory compliance and AI for banks and advanced AI risk management tools go deeper on the controls surrounding these tools.
FAQ
What is a vendor risk assessment?
It is the structured review of a third party's risks before you contract and throughout the relationship. For U.S. banks, the interagency guidance organizes it across planning, due diligence, contracting, ongoing monitoring and termination. The depth should match how critical the vendor's activity is to your institution.
Can AI write a vendor risk assessment for me?
AI can draft summaries, map answers to controls and flag inconsistencies, which saves analyst time. It cannot own the conclusion. A named reviewer should approve the rating, and the AI tool itself should be assessed as a vendor.
What should an AI vendor risk assessment include?
Cover data privacy and retention, model training and validation, information security, technology integration with sub-service providers, and legal and regulatory compliance. Add use-case tiering, evidence requests, contract terms and an exit plan.
How do credit unions approach AI vendor due diligence?
They start with the NCUA's third-party letters and its AI resource page, which highlight algorithmic transparency, fair lending, member data privacy, resilience and model risk. The same evidence-first approach applies, scaled to the credit union's size and the vendor's role.
Sources
- OCC Bulletin 2023-17, Third-Party Relationships: Interagency Guidance
- Federal Register: Interagency Guidance on Third-Party Relationships
- NCUA: Artificial Intelligence resources
- Venminder: State of Third-Party Risk Management 2025 highlights
- Ncontracts 2025 Third-Party Risk Management Survey
- U.S. Treasury: Managing AI-Specific Cybersecurity Risks in the Financial Services Sector
- NIST AI Risk Management Framework
- FS-ISAC Generative AI Vendor Evaluation and Risk Assessment Guide
- EY 2025 Global Third-Party Risk Management Survey
- AI third-party risk assessment questionnaire framework