Most underwriting AI stops at the bank statements

When somebody says they use AI to underwrite, they almost always mean one thing: it reads the bank statements. It pulls the deposits and withdrawals and daily balances off three months of inconsistent PDFs and hands you clean numbers instead of sixty pages someone has to re-key. That saves real hours, but it’s the floor. It’s data entry that doesn’t get tired.

A step up from that, barely any shops are here, is inference. You’ve funded thousands of merchants and you know how the ones that looked like this turned out, so you let the model lean on that history and put a number on the risk. Useful. Still mostly working with what’s in front of you, the file the broker sent and the numbers on the page.

The part almost nobody uses it for is the part it’s actually best at, which is digging.

Every deal has threads in it that a standard check doesn’t pull. A background report keys on exact identifiers: run the name, run the entity, get a result, move on. But people are messier than their identifiers. A merchant who burned a funder two years ago doesn’t reapply with a sign around his neck. He reapplies after spelling his name a little differently than he did last time, and the old default sits under the old spelling, where the report isn’t looking. Or the business that defaulted is gone and there’s a new one in an associate’s name (a cousin, an old employee, somebody), so on paper it’s a clean first-time applicant, except it runs out of the same address, takes deposits from the same handful of customers, and sometimes sits in a bank account that’s older than the company is supposed to be.

Those are exactly the connections a model is good at finding, because finding them is fuzzy and cross-referenced and buried in unstructured text, which is the kind of work that breaks a rules engine and bores a person into skipping it. A human will chase one thread when something feels off on a file. A model can chase all of them on every file: name variants that are close but not exact, an address that showed up on a deal you declined last spring, a phone number attached to three businesses that are supposedly unrelated, a debit pattern you’ve seen before under a different name. None of those is proof by itself. Put together, they’re the thing the first report told you wasn’t there.

That’s where the expensive mistakes hide. They usually aren’t the deal you read wrong. They’re the deal you read right, off a file that was put together to look clean. The numbers can all be real and the merchant can still be someone you already decided not to do business with. Reading the statements faster does nothing for that. The only thing that catches it is pulling the threads, and pulling them on every deal, not just the ones that happen to smell off.

The model doesn’t make the call here either. It surfaces the thread and you decide whether it means anything. A different spelling can be a typo or it can be somebody hoping you won’t notice. But I would rather have my people looking at a flagged connection and deciding it’s nothing than never seeing it, which is what happens on most of the volume right now.

So when the conversation about AI in underwriting stays on extraction, how fast it reads, how cleanly it parses, I think it’s pointed at the least interesting thing the technology does. The statements were always going to get read; that part was never really the question. The harder question, and the one that actually decides whether you make money, is whether you can tell when a clean-looking file belongs to someone who already cost you once. We catch more of those than we did a year ago. We still miss some.

*Cosmin Panait is a co-founder of Blackbridge Investment Group, a revenue-based funder in New York.*