AI Is Better Than Us at Scoring Credit — and Nowhere Near Ready for the Decisions That Matter

AI is already better than most humans at scoring credit. It's also nowhere near ready to be trusted with the decisions that actually matter. Both things are true, and the gap between them is where the next decade of lending will be won or lost.

I've spent two decades in credit risk — building the software, and then building models for India's thin-file MSMEs. So let me be clear about what AI genuinely does well here, because it's a lot. It reads documents, spots fraud patterns, synthesises alternative data, monitors a portfolio in real time, and prices routine risk faster and more consistently than any human team. For the 95% of lending that is mechanical, AI isn't a threat — it's overdue.

But credit isn't the 95%. Credit is the 5% that breaks the rules. And that's where the honest limits show up:

1. AI is only as good as its data — and confidently wrong without it. Point a model at a borrower it has never really seen and it doesn't say “I don't know.” It interpolates, smoothly and persuasively, into territory where it has no business having an opinion. The thin-file borrower doesn't get a careful maybe; they get a confident guess dressed as a score.

2. Models learn the past. Credit blows up on the future. A model trained on ten good years has never seen the eleventh. It's precisely at the regime change — the shock, the cycle turn, the “this time is different” — that historical patterns betray you, and that's exactly when a lending decision matters most. Judgment about what the data can't yet show is not in the training set.

3. Someone has to be accountable. “The model declined you” is not an answer a regulator accepts, a customer deserves, or a lender can defend. Credit needs a human who can explain the call and stand behind it — for fairness, for compliance, and for trust.

So the future isn't AI replacing the underwriter. It's AI doing the 95% so the underwriter can finally spend their judgment on the 5% that decides whether a loan book compounds or quietly detonates.

Automate the mechanical. Own the judgment. Confuse the two and the model will be right until the day it catastrophically isn't.

Where have you seen AI genuinely improve a credit decision — and where have you seen it confidently miss?