- 1Roitblat agrees underwriting is mainly a math problem and LLMs are not the right tool for it.
- 2He argues MCA underwriting is messier than consumer installment lending and runs after funding.
- 3He says bank-statement reading and post-funding monitoring are core underwriting, not paperwork.
David Roitblat, founder and CEO of AI My Advance and Better Accounting Solutions, argues in a guest column that Upstart CEO Paul Gu's view of AI in underwriting does not carry over to the merchant cash advance industry.
What Gu said
On Upstart's Q4 earnings call, which was covered on March 23, 2026, Gu argued that humans have never been good at precisely underwriting loans and projecting cash flows, and that this is a math problem rather than a language problem. In his framing, large language models suit the operational layer around lending, such as reading documents, perfecting liens and checking property records, but not the underwriting decision itself. "No matter how many humans you have, you don't want that army of humans underwriting loans for you," he said.
Where Roitblat agrees
Roitblat concedes much of that. Underwriting is at heart a probability problem, he writes, and LLMs alone are not the right architecture for portfolio-level probability estimation. Upstart's result of 91% of loans now fully automated came from years of structured-data modeling rather than GPT-class models. He also agrees that the clearest near-term LLM wins are operational, in areas like HELOC processing, lien perfection and title work.
Where he disagrees
His objection is to the conclusion that the conversation is therefore settled. Upstart underwrites consumer installment loans, with clean data, predictable durations and borrowers who have credit files. MCA underwriting, he says, prices a daily or weekly remittance against a business's future receivables over months. The inputs are messy: bank statements categorized idiosyncratically, seasonality, stacking risk, processor changes, partner disputes and owner behavior that a FICO score misses. Consumer credit grew up with decades of structured infrastructure, he says, while MCA data is unstructured by default.
Roitblat contends that parsing a statement, reading a memo line and recognizing a third stacked funder are part of underwriting in MCA, not side tasks. Telling a $4,200 ACH debit for a vendor payment apart from one for another funder's advance, he notes, takes both language and math.
Underwriting after funding
Roitblat also says underwriting does not end at approval. He describes a system that links to merchant bank accounts after funding, tracks deposit and remittance patterns, finds what defaulted deals had in common and feeds that back into pricing the next deal. He says his platform, AI My Advance, does this.
His summary is that MCA underwriting is a math problem wrapped in a language problem wrapped in a continuous-monitoring problem. He credits Upstart with earning its view, but says it was built on a problem whose inputs were already structured.
Source reporting: deBanked.
