Book a demo
Topics · REGULATORY

Credit unions Model Risk Vendor Selection | Moring

Vendor selection is where credit union model risk is largely decided, because most AI arrives through core vendors and fintech partners. Selection should tier diligence by use case, demand evidence rather than assurances, probe fourth-party model dependencies, and contract for audit rights, change notice, and exit.

On this page
  1. Short answer
  2. How it works
  3. FAQs
  4. Related reading
  5. Sources

Short answer

Vendor selection is where credit union model risk is largely decided, because most AI arrives through core vendors and fintech partners. Selection should tier diligence by use case, demand evidence rather than assurances, probe fourth-party model dependencies, and contract for audit rights, change notice, and exit.

How it works

Vendor selection is the highest-leverage model risk decision a credit union makes, because most of the sector's AI arrives embedded in core platforms, lending systems, and fintech products rather than built in-house. Choosing the vendor is therefore choosing the model risk profile: the institution acquires the model's opacity, the vendor's update cadence, and the fourth-party dependencies behind the product, all before any member sees an output. Interagency guidance on third-party relationships makes the accountability explicit: outsourcing the function does not outsource the responsibility for managing its risk.

Diligence should tier by use case, not by vendor reputation. An autonomous credit decision deserves far more scrutiny than a drafting assistant, and the questions should scale accordingly: how the model was developed and validated, what data trained it, how outputs are explained, what drift monitoring the vendor runs, and what evidence the institution can obtain continuously rather than once at onboarding. Reviewers examining an AI vendor file look for three things: the institution understood the system before signing, the contract retains oversight, audit rights, and exit options, and monitoring is ongoing rather than a checkbox at renewal.

Evidence beats assurance. Practical selection criteria for AI vendors include validation reports, independent audit results, and sample outputs the institution can test against its own member data patterns, plus clarity on where the model runs and what data residency applies. Fourth-party exposure deserves explicit attention: a secure vendor can still depend on foundation model providers and data aggregators the credit union never contracted with, and a model swap behind the interface can change behavior without any institutional decision. Selection should ask how the vendor notifies customers of model changes and whether versions can be pinned.

The model risk frame around selection is now explicit. Revised interagency guidance issued as SR 26-2 in April 2026 carries third-party expectations into model risk management, with the institution responsible for understanding models it deploys regardless of origin, and its predecessor SR 11-7 established the validation and governance pillars that vendor evidence should map to. FFIEC guidance on outsourcing supplies the due diligence, contract control, and ongoing monitoring expectations; EU AI Act obligations attach where vendors or deployments create European exposure, and vendor documentation is part of meeting them.

Close selection with contract terms that convert assurances into controls: audit and information rights, notice of material model or data changes, performance and fairness metrics reported on a schedule, incident notification duties, and exit assistance including data portability sufficient to move or rebuild. Selection documents, the scored questionnaire, the evidence received, and the gaps accepted with sign-off, become the opening pages of the model risk file for that system, and the quality of those pages is what reviewers judge when the system later misbehaves.

Frequently asked questions

How should a credit union evaluate AI vendors for model risk?

Tier diligence by use case rather than vendor reputation: an autonomous credit decision warrants far deeper scrutiny than a drafting assistant. Demand evidence, validation reports, independent audit results, sample outputs tested against the institution's own data, and probe fourth-party dependencies such as foundation model providers. The 2023 interagency third-party guidance supplies the supervisory frame, and accountability for outcomes remains with the credit union regardless of the vendor's quality.

What contract terms matter most for AI model risk?

Audit and information rights, notice of material model or data changes, scheduled performance and fairness reporting, incident notification duties, and exit assistance with data portability. Version pinning or rejection rights matter where model swaps could silently change behavior. These terms convert vendor assurances into controls the credit union can exercise, and they determine what evidence exists when reviewers later ask how the system is governed.

What is fourth-party exposure in AI vendor selection?

It is the risk hiding behind the vendor: foundation model providers, data aggregators, and subcontractors the credit union never contracted with but whose behavior shapes the product. A vendor can pass diligence while its underlying model provider changes the model, altering outputs and risk overnight. Selection should ask vendors to disclose AI supply chains, notify the institution of changes, and explain their own oversight of those dependencies.

When should model risk questions enter the vendor conversation?

Before selection, not after signing. Model risk belongs in the requirements and scoring stage, because validation evidence, explainability, monitoring, and change notice are easier to secure in competitive evaluation than in renegotiation. The institution should leave selection with a scored record of what was asked, what was answered, and what gaps were accepted, since that file becomes the opening evidence for the system's model risk management.

Sources

  1. ncua.gov
  2. crosscheckcompliance.com
  3. www.myabt.com
  4. www.riskinmind.ai
  5. www.standardc.com
  6. www.360factors.com
  7. - federalreserve.gov
  8. - federalreserve.gov
  9. - ithandbook.ffiec.gov
  10. - artificialintelligenceact.eu