Underwriting After the Language Model
Credit decisions are being made on documents no human will ever read.
By Lena Vogel
The document was always the bottleneck
Small business lending stalled for decades on the same problem: the information needed to underwrite sat in bank statements, invoices, and tax filings that cost more to process than the loan earned.
Models that read unstructured documents reliably changed that arithmetic. Approval times fell from weeks to hours at lenders who rebuilt the process around the capability instead of bolting it on.
We did not make the credit model smarter. We made the inputs cheap.