AI lending for credit unions: what ships, what decides, and who owns the model

Two categories, one decision

Everything on offer falls into one of two buckets, and the difference determines how much work you inherit.

Preparation AI reads documents, extracts figures, checks them against guidelines and drafts follow-ups or narratives. One vendor's agent reviews a borrower document set in 15 to 25 seconds against agency, overlay or custom guidelines and generates cited follow-ups, and it is deliberately designed never to make a credit decision. On the commercial side, preparation AI classifies borrower documents, builds spreads, calculates coverage across entities and guarantors and drafts the memo. In both cases no credit model enters your inventory, which makes the governance conversation short.

Decisioning AI scores or approves the application. That is where the approval-rate gains are, with one vendor targeting auto-decisioning of roughly 80% of consumer applications, and it is also where a model appears that your risk function has to document, validate, monitor and defend. Both are legitimate purchases. Buying the second while budgeting for the first is the mistake worth avoiding.

• Preparation AI: document review, extraction, spreading, memo drafting, follow-up generation
• Decisioning AI: scoring, approval, risk-based pricing, strategy design
• Preparation AI adds no credit model to your inventory
• Decisioning AI adds validation, monitoring, fair-lending testing and adverse-action defence

What documentation arrives with the model

This is the question that separates vendors once you get past the accuracy slide, and the answers vary more than you would expect. A custom model built on your own portfolio is powerful and it is also yours to defend. Ask for the specific deliverables in writing rather than accepting a description of the vendor's approach.

One decisioning vendor markets a deep fair-lending apparatus, including less-discriminatory-alternative model searches and adversarial debiasing, without publishing a model-risk deliverable, validation package or adverse-action artifact list. That is not a criticism of the modelling, which is well regarded, but it does tell you where the assembly work lands. Another publishes model documentation from day one alongside a seven-year tamper-evident decision log, decision replay, an override audit, adverse-action reasons mapped to ECOA and a one-click examiner export. For a credit union without a dedicated model-risk function, that difference is worth more than a point of approval lift.

• Validation documentation at implementation, and after each retrain
• Ongoing performance and drift monitoring, and who reviews it
• Fair-lending test results, including any less-discriminatory-alternative search
• Adverse-action reason mapping, at the application level
• A decision log you can replay, and an override record with attribution

How to read the numbers

Performance figures in this segment are almost entirely vendor-measured, undated and occasionally inconsistent with each other. One vendor's active-model count appears as 600-plus on its website, 1,200-plus in an April 2026 announcement and 1,500-plus in August 2026 boilerplate, which is not dishonest but does mean no figure should be repeated without a date attached. Approval-lift and automation claims rarely carry methodology.

The way through this is to stop negotiating over published numbers and insist on a backtest. The better products support it directly, including backtesting a strategy against your own historical applications and shadow-testing it with maker and checker approval before it decides anything live. That is your members, your criteria and your loss experience, and it settles in two weeks what a quarter of claim comparison will not.

One category of claim deserves to be discounted to zero. A vendor in this market states that all of its clients have passed their NCUA audits since deployment, with no methodology, sample size or third-party attestation. That cannot be checked and should carry no weight. Published artifacts can be checked, which is the standard worth applying.

Where AI has not arrived: member business lending

The consumer side of this market is mature. The commercial side is not, and the reason is structural rather than temporary. Consumer credit decisioning works because there is a large, homogeneous population of past applications to learn from. A credit union's member business loans are fewer, more varied and less comparable, so the same modelling approach has much less to work with.

What AI does do well on the commercial side is the preparation work, and that is where the current products sit: classifying a folder of borrower documents, spreading returns, calculating coverage across entities and guarantors with add-backs applied, checking a file against written credit policy and drafting a memo. Neither leading consumer decisioning vendor covers commercial at all, with one enumerating exactly six model types, none of them commercial.

For a credit union that means the AI question splits by book. On the consumer side it is a decisioning conversation with real governance consequences. On the member business side it is a document-and-analysis conversation, and the useful requirement is traceability: a figure that links back to the page it was read from, so an automated spread can be defended rather than explained.