Scoring AI Risk Without a Model Risk Team
How a credit union with one compliance analyst can run a defensible AI risk program
Guest Editorial by Lisa Pent, Founder and CEO, PentEdge
A $900 million credit union sat down to score its AI inventory last fall. The CCO had read the regulatory guidance. She knew that model risk management was the gold standard. She also knew she did not have a model risk team, would not have one next year, and could not justify one to a board already stretching the compliance budget.
The instinct in that situation is to do nothing. If you cannot do it the way a top 25 bank does it, wait until you can. That instinct is wrong. Nobody expects a credit union to reproduce the model risk program of a large bank. The question that matters is whether the credit union can articulate, in writing, why each AI tool sits at the risk level it sits at, and what controls match that level.
That is a question a one-person compliance shop can answer, if the framework is right.
What A Risk Score Has to Do
Strip away the jargon and a risk score for an AI tool has to do three things. It has to be repeatable, so two people scoring the same tool land in roughly the same place. Defensible, so the score can be explained to a board member, an examiner, or an auditor without a 40-page appendix. And proportionate, so the effort of scoring does not exceed the value of the score.
Most published AI risk frameworks fail the third test for credit unions. They were designed for institutions with dedicated model risk teams and assume statistical analysis a small compliance shop cannot sustain. Small institutions either skip risk scoring entirely or copy a framework they cannot run.
There is a middle path. It uses inherent risk factors a non-quant can observe, mitigation factors that map to controls the credit union already has, and a residual score that lands every tool on a defensible spectrum.
The Five Inherent Risk Factors
Inherent risk is the risk of the AI tool before any controls are applied. Five factors do most of the work.
Data sensitivity. Does the tool touch member PII, financial data, or regulated information? A scheduling assistant scores low. A chatbot trained on member call recordings scores high.
Decision impact. Does the tool influence a decision that affects a member? A meeting summarizer does not. An adverse action letter generator does. A credit decisioning model does heavily.
Autonomy. Does the tool act on its own or recommend to a human who acts? Recommendation systems carry less inherent risk than autonomous ones, and full autonomy on any member-facing decision should be a flag every time.
Reversibility. If the tool produces a bad output, can it be unwound? A misclassified fraud alert corrected within minutes is different from a denied loan the member never reapplies for.
Regulatory exposure. Does the workflow touch a regulated activity such as lending, BSA, UDAAP, fair lending, or privacy? If yes, the inherent risk is higher regardless of how the model performs in testing.
The Six Mitigation Categories
Mitigation is the work the credit union does to reduce inherent risk. Six categories cover most of it. Human-in-the-loop review. Documented testing before deployment. Ongoing monitoring of outputs. Access controls and data segregation. Vendor due diligence on the underlying model. Member-facing transparency where required.
Each category is either present, partial, or absent for a given tool. The combination of inherent risk and mitigation produces a residual score that lands the tool in low, medium, high, or critical. That is the score the board sees and the score the examiner asks about.
Why This Works in a Small Shop
Three things make this approach sustainable for a credit union without a model risk team.
It uses observation, not modeling. A compliance analyst can answer every inherent risk question by reading the vendor documentation and watching the tool work. No statistical analysis is required.
It uses controls the credit union already has. Human review, access controls, vendor due diligence, member disclosure: these are not new disciplines. They are the controls credit unions already apply to other technology, mapped onto AI.
It produces a paper trail. Every score has a written rationale. Every rationale references the inherent factors and the mitigation factors. That paper trail is what turns an exam question into a five-minute conversation instead of a finding.
Where To Start
If you have an AI inventory, you can start scoring this month. Pick the three tools that touch the highest-risk workflows. Walk each one through the five inherent factors and the six mitigation categories. Write the rationale. File it.
By the time you have scored ten tools, you will have a working framework. By the time you have scored thirty, you will have a program. Not a model risk program in the top 25 bank sense, but a real, defensible, proportionate one.
The credit union that waits for a model risk team to materialize will wait a long time. The credit union that starts scoring now will be ready when the examiner asks.
About The Author
Lisa Pent is the Founder and CEO of PentEdge, a women-owned RegTech company that built the financial industry’s first AI Risk Score™, launched at FinovateSpring 2026. Its platform, AIMS™, gives community banks and credit unions one place to detect the AI reaching into their systems, score it, and keep it examiner-ready. Lisa spent 30-plus years in financial services, including a decade at Thomson Reuters building SaaS for financial institutions.
