Risk
What an AI score should decide, and what it should not
6 min read
Every lender we have spoken to in the last two years has asked some version of the same question: can the model just decide? The honest answer is that it can, technically, and that this is usually the wrong thing to ask it to do.
A score is a compression of evidence. It is useful in exactly the same way a map is useful — as a way to see the terrain, not as a substitute for walking it.
What a score actually is
An AI score takes the material already gathered about an application — verified identity, bank behaviour, screening results, declared information — and estimates how likely a defined bad outcome is. The definition of that outcome is yours: default within twelve months, early arrears, fraud. The model does not invent a policy; it ranks cases against the one you already have.
That distinction matters because teams that treat the score as the policy end up unable to explain a decline. A regulator, a customer, or a court asking why this person was refused is not asking for a number. They are asking for the reason the number exists.
Where it belongs in the decision
The useful pattern is three bands. Below a threshold, accept without review. Above another, decline without review. Between them, a person decides with the score and the evidence in front of them. The thresholds are configuration; they move when your appetite moves, without a release.
A model that always decides will be quietly overridden. A model that never decides will be ignored. The work is drawing the line between the two.
Whose history the model learns from
A model trained on your own book reflects your customers and your definition of failure. That is the version worth having. Without enough history, start with explicit rules you can defend, collect outcomes, and retrain. Borrowing someone else's model is borrowing someone else's definition of a good loan.
- Agree the outcome you are predicting before you train anything
- Keep a record of what the model saw when it scored each case
- Retrain on a schedule, not only when something goes wrong
- Treat a sudden change in the score distribution as an incident
Have a system exactly as you envision it
Let's talk. It's time to make a better version of your business.