- 1Gu: underwriting is a big math problem, not work humans do especially well
- 2Upstart sees LLMs fitting document-heavy tasks like securing a HELOC lien
- 3Upstart's credit models predate modern LLMs
Upstart CEO Paul Gu told analysts on the company's Q4 earnings call that humans have never been good at precisely underwriting loans and forecasting their cash flows over five years, which is one reason he does not see AI as a threat to Upstart's business. The comments came in response to an analyst's question.
Math, not human judgment
Upstart's consumer credit models were built before the current wave of large language models, and the company has long counted heavy automation among its defining traits. Its underwriting rests on strong data access and math. Gu argued that AI is mainly good at replacing tasks people already do well, and underwriting is not among them.
“That’s something that has always been solved as a big math problem,” he said.
Where LLMs fit
Gu pointed to a HELOC as a better use case. Human processors must secure and perfect a lien and check property records, work he described as messy and costly because many people are needed to verify it. “Those are actually the perfect problem to throw sort of LLM-style AI against,” he said.
That is how Upstart plans to apply AI, Gu said. He noted that models from Anthropic, OpenAI, Gemini and others can solve human-type problems at scale, working around the clock and running in parallel by the hundreds. But no matter how many humans a lender has, he said, it would not want an army of them underwriting its loans.
Source reporting: deBanked.
