Explainable AI in credit: the questions your auditor will ask
AI-assisted credit analysis is no longer experimental. The banks that get it through audit are the ones that can answer five specific questions about any decision.
The conversation about AI in credit has moved. Two years ago the question was whether a model could analyse a loan file at all. Now the model analyses the file, and the question belongs to risk officers and auditors: can you stand behind what it did?
That question decomposes into five smaller ones. Any bank putting AI into the credit process should be able to answer all five — for any individual decision, months after the fact.
1. What did the system see?
The inputs. Which statements, which registry extracts, which transaction history, as of when. If the analysis ran on Tuesday's data and the customer uploaded a corrected statement on Wednesday, the record needs to show that. An AI recommendation without a snapshot of its inputs is unfalsifiable — and auditors are professionally allergic to unfalsifiable.
2. What did it conclude, and why?
Not a score. A score is a number wearing a suit. What stands up to scrutiny is a written recommendation with its reasoning: which factors drove it, what would have changed it, stated in language a credit committee — and a customer, and a regulator — can read. If the reasoning can't be written down, the reasoning isn't done.
3. Who decided?
The system assists; a person signs. This is both the emerging regulatory expectation for consequential decisions and, frankly, good engineering: the model's job is to make the human's judgement better-informed and more consistent, not to replace the accountability that banking is built on. The record should show the recommendation and the human decision — including when they differ, because the disagreements are the most informative entries in the log.
4. Is it consistent?
Two files with the same facts should get the same analysis. With manual review this is genuinely hard — two experienced analysts can defensibly read one file two ways. With assisted analysis it's a property you can test and demonstrate: same inputs, same reasoning, every time. Consistency is the quiet superpower of doing this well; it converts "we believe our process is fair" into "here, run it yourself."
5. Where does the trail live?
If the answers to the four questions above are scattered across emails, model logs and someone's memory, you don't have explainability — you have an archaeology project. The trail has to live on the case itself: inputs, analysis, recommendation, reasoning, sign-off, one record.
This is the standard we built Vincu Decision against — every run logged with its model version and inputs, every recommendation written out with reasoning, every decision signed by a person. It's in production at a European bank today, and the audit questions above are answered by opening the record, not reconstructing it.
If your AI-in-credit conversation is heading toward these questions, we should talk.