AI

The Fatal Flaw in Upstart CEO’s Vision for AI Underwriting in MCA: A response to Upstart’s view of AI in underwriting

David Roitblat is the founder and CEO of AI My Advance and Better Accounting Solutions, a leading authority in specialized accounting for merchant cash advance companies alongside our new innovative CRM designed to solve the critical gaps holding the industry back. To connect or schedule a call about working with AI My Advance or Better Accounting Solutions, email David@betteraccountingsolutions.com

As deBanked reported (March 23, 2026), Paul Gu, CEO of Upstart, said something on his company’s Q4 earnings call that has been quoted approvingly in a few corners of the credit world. It deserves a careful read because some important points are incorrect when it comes to the MCA industry.

Gu’s argument, as reported, runs roughly like this. Humans have never been very good at precisely underwriting loans and projecting cash flows. That problem has always been a math problem, not a language problem. The recent wave of AI, the LLMs from Anthropic, OpenAI, and Google, is good at the things humans are naturally good at: reading documents, navigating messy paperwork, perfecting liens, and checking property records. Therefore, in Gu’s framing, LLMs are well-suited to the operational layer around lending but not to the underwriting decision itself. He puts it bluntly: “No matter how many humans you have, you don’t want that army of humans underwriting loans for you.”

The deBanked piece does a real service by reporting it. But the framing carries a serious blind spot. It implies that AI underwriting is essentially a solved problem, owned by structured-data shops like Upstart, and that the broader conversation about AI in lending is mostly noise. For consumer credit, that may be true. For the MCA industry, it is not, and treating it as if it were misses the actual point.

Begin with the parts of his argument that hold up. Underwriting, at its core, is a probability problem. Given a set of inputs, what is the likelihood of repayment, and what is the distribution of outcomes if it fails? That has always been the question, and Gu’s phrasing is exactly right: “That’s something that has always been solved as a big math problem.”

He is also correct in saying that LLMs, in their current form, are not the natural tool for that math problem. They are extraordinary at reading, summarizing, classifying, extracting, and reasoning over language. They are not, on their own, the right architecture for portfolio-level probability estimation. The deBanked piece links to Upstart’s own track record, 91% of their loans now fully automated, and that result was not built on top of GPT-class models. It was built on years of structured-data modeling. Gu is entitled to point that out.

He is also right that the most obvious near-term wins from LLMs in lending are operational: HELOC processing, lien perfection, document review, title work. Anywhere a human currently spends hours reading and routing paper, an LLM can do meaningful work. His framing of those as “the perfect problem to throw sort of LLM-style AI against” is well put.

The problem is the conclusion the framing pushes the reader toward: that because the underwriting decision is a math problem, and because the new wave of AI is mostly a language problem, the conversation about AI in underwriting is largely settled. That conclusion translates poorly from Upstart’s world to ours, and it does so in a way that matters.

Upstart underwrites consumer installment loans. The data is clean, the durations are predictable, the borrowers are individuals with credit files, and the question being asked is essentially: will this person make 36 or 60 fixed monthly payments on time. That is, as Gu says, a big math problem with a long history of structured inputs.

MCA underwriting is not that problem. It is a different problem with a different shape, and the difference is not cosmetic.

An MCA underwriter is not pricing a fixed-term installment loan to an individual with a credit file. They are pricing a daily or weekly remittance against a small business’s future receivables, over a horizon measured in months, where the inputs are messy by nature: bank statements with idiosyncratic categorization, industry-specific seasonality, owner behavior that does not show up in a FICO score, stacking risk, processor changes, lease events, partner disputes, and dozens of other signals that live in unstructured form. The decision is not made once and then monitored passively. It is revisited continuously as the merchant’s behavior evolves over the life of the advance.

That is a different beast, and the math that solves consumer credit does not, on its own, solve it. The part of AI that Gu sets aside, the language part, the unstructured-data part, the continuously-observing part, is precisely the part that matters for the underwriting decision itself in MCA, not just the paperwork around it. Setting it aside is not a clean theoretical move. It is a category error when applied to this industry.

In modern MCA operations, the most important thing AI is doing is not the initial decision. It is what happens after the capital goes out.

Underwriting does not end at approval. Our new and up-and-coming platform continues to ingest merchant behavior after funding. It links directly to merchant bank accounts, monitors post-funding activity, tracks changes in deposit patterns and remittance behavior, and detects patterns across defaulted deals. It does not simply record that a deal failed. It identifies what those failed deals had in common. That information does not sit in a static report. It feeds directly back into how new deals are evaluated.

That is underwriting, too. It is just underwriting, not the single-moment, single-file decision the term usually evokes. It is a feedback system: the decision at funding is the first input, the merchant’s subsequent behavior is the second, and the model that prices the next deal is the third.

Gu’s framing treats underwriting as the moment of decision. In MCA, the moment of decision is a single frame in a longer film. The frames that matter most are often the ones after funding, where a human underwriter cannot realistically hold the pattern in mind across hundreds or thousands of merchants, while a system can. Any framework that ignores those frames does not describe MCA underwriting. It describes something else and calls it by the same name.

This is also where the LLM-versus-structured-model dichotomy starts to fall apart. Reading a bank statement well is partly a math problem and partly a language problem. Categorizing a $4,200 ACH out as a vendor payment versus a stacked advance from another funder is not pure math. It requires reading the counterparty name, recognizing the funder, knowing the industry conventions. That work sits exactly at the seam between what Gu calls “solved as a big math problem” and what he calls “the perfect problem to throw sort of LLM-style AI against.” In MCA, those two are not separable layers. They are the same workflow, and pretending otherwise produces a model of the industry that does not match how the industry actually runs.

The deeper point is that the AI conversation in lending is not one conversation. Upstart’s answer is the right answer for Upstart’s problem. It does not automatically transfer, and the confidence with which it has been quoted in the credit world suggests the transfer is being assumed rather than examined.

Consumer credit, where Upstart operates, has had decades of structured-data infrastructure built around it: bureaus, scores, standardized loan documents, regulated disclosures, and well-defined repayment behavior. A math-heavy, low-LLM approach makes sense there because the inputs were already structured by the time the modeling started.

MCA grew up differently. The data is unstructured by default. The borrowers do not have meaningful credit files in the consumer sense. The product is non-recourse against a fluctuating revenue stream. The lifecycle is short and active. The signal that matters often lives in places, bank statement memos, merchant behavior, processor data, partner disputes, that look like language problems and behavior problems before they look like math problems.

That is why the LLM-style AI Gu is comfortable assigning “around the edges” work, which, in MCA, frequently involves core underwriting. Parsing the statement is part of underwriting. Reading the memo line is part of underwriting. Recognizing the third stacked funder is part of underwriting. Watching how a merchant’s deposit pattern shifts in week three of an advance is part of underwriting. Those are not auxiliary tasks bolted onto a clean math problem. They are the problem.

The disagreement with Gu, then, is not that AI cannot underwrite. He is not really arguing that. His actual position is closer to this: LLMs are not the right tool for the underwriting math, and you should not let the hype around LLMs convince you otherwise. On that, we agree.

The disagreement is this. In MCA, the underwriting problem is a math problem wrapped in a language problem wrapped in a continuous-monitoring problem. The math is necessary but not sufficient. The language and behavioral layers are where modern AI, LLM-style and otherwise, is genuinely changing how deals are priced, monitored, and learned from. With our new and up-and-coming platform, AI My Advance, this is exactly what we do: we parse the unstructured inputs, track merchant behavior after funding, and feed what we learn from defaulted deals back into how the next deal gets priced. Treating that work as merely “operational” is not a small mistake. It is the kind of mistake that produces a confident answer to the wrong question.

Upstart has earned the right to its view. The 91% automation figure is real, and the underlying modeling is serious. But that view was built on a problem whose inputs were already structured. The MCA industry is solving a different problem in real time, and the tools that work for it are not the same tools, in the same proportions, as those that work in consumer credit. The sooner that distinction is named clearly, the better the conversation about AI in lending becomes.