MCA accounting CEO challenges Upstart's Paul Gu on AI in cash-advance underwriting

On June 12, David Roitblat, who runs AI My Advance and an accounting firm built for cash-advance shops, published a rebuttal on deBanked to something Upstart chief executive Paul Gu said on the company's most recent earnings call. Gu's position, first reported in March , was that underwriting has always been a math problem, that humans were never very good at it, and that large language models belong on the operational work around a loan rather than the credit decision itself. Roitblat agreed with almost all of it, then argued the whole thing breaks the moment you carry it into merchant cash advance. The disagreement is not really about whether AI can do math. It is about when underwriting ends. In consumer lending the decision is made once, against a clean credit file, and then monitored passively until the fixed payments either land or stop. In an advance, the decision at funding is the opening frame, not the whole film, and the data that decides whether the deal performs keeps arriving every week the merchant keeps trading. The funders who read that correctly treat the bank statement after funding as part of underwriting, not as a collections afterthought. Roitblat put it plainly: in MCA "the moment of decision is a single frame in a longer film," and the frames that matter most often come after the money goes out. He is careful to give Gu his due first. Upstart says 91% of its loans are now fully automated, built on years of structured-data modeling rather than GPT-class models, and the cleanest near-term wins for language models really are operational, things like document review, lien perfection, and title work. None of that is in dispute. His split is that consumer credit and a cash advance are two different problems wearing the same word. A consumer borrower has a credit file, a score, standardized paperwork, and a fixed term. A small business hands over a stack of bank statements with idiosyncratic memo lines, seasonality no bureau captures, owner behavior that never appears in a FICO number, and the live possibility that another funder stacked a position last Tuesday. Calling both of those the same math problem, Roitblat argues, is a category error that gets assumed rather than examined. His sharpest point is that the math-versus-language line does not survive contact with a real file. Deciding whether a $4,200 ACH debit is a vendor payment or a repayment to a stacked funder requires reading the counterparty, recognizing the other funder, and knowing the conventions of the trade. That is not a clean number waiting for a model. It is a language problem and a math problem sitting in the same line, and getting it wrong changes both the merchant's true obligations and the price of the next deal. The operator read A funder that underwrites once and files the deal away is blind on the part of the advance that actually moves. Deposit patterns shift in week three, a processor changes, a second funder turns up in the debits, and none of that is visible to a model trained on the single day the deal closed. The shops that price well feed post-funding behavior back into how they score the next file, so the same monitoring that protects the current book is what makes the underwriting sharper. There is a quieter commercial point under the technical one. Roitblat sells a monitoring product, so the piece is partly a pitch, and an editor should read it as one. But the structural claim stands without the sales layer. An advance is a continuous exposure, and a continuous exposure is either underwritten continuously or it is not really underwritten at all. A merchant can clear approval cleanly and then change behavior in ways that no closing-day snapshot would ever catch. Opinion Underwrite the film, not the frame A cash advance is not a fixed-term loan with a messier borrower. It is a moving position whose real risk is re-revealed every week the business deposits, debits, and reconciles. A funder that captures that behavior in a structured, dated record prices the next deal sharper and can answer a capital partner's questions without scrambling. The shop that treats approval as the finish line keeps relearning the same defaults the expensive way. More on what funders should be reading after the money goes out at plutusai.net .