What Pavan Agarwal Got Right About AI Decades Before the Industry Figured It Out

The interesting question in mortgage AI right now is not whether the technology works. It is which architecture works under the regulatory load that financial services actually carries. Pavan Agarwal made a call on that question more than a decade ago.

The call was to build AngelAi® on a deterministic foundation. Same input, same output. Every output is traceable to the rules and data that produced it. He committed to that architecture during a period when the rest of the AI field was investing heavily in the opposite approach, building probabilistic models that could not produce the same answer twice and could not be reliably explained.

For most consumer applications, that probabilistic looseness is acceptable or even useful. For mortgage decisions, it is operationally fatal. A lender cannot tell a declined applicant that the system did not feel like approving them that day. The Consumer Financial Protection Bureau cannot audit a decision that cannot be reproduced. Adverse-action notices, fair lending reviews, regulator examinations: all of these require an answer that does not move when the question is asked twice.

Agarwal had been operating inside that regulatory environment from day one. He grew up around it, in the mortgage company his father founded in 1980, helping his parents through the kind of jobs immigrant kids do when the family business is the family table: mowing lawns, painting walls, handing out flyers, working the phones. By the time he became CEO of Sun West Mortgage Company in 2008, he had worked nearly every role inside a lender. He was not theorizing about how AI might be governed in the future. He was building credit decisions that had to be defensible in the present. The deterministic call followed from the requirement.

You notice this when you meet him. He shows up at industry events in an AngelAi t-shirt and jeans, walks the floor with his team rather than just sending them, and replies to his own LinkedIn comments. At the 2022 Originator Connect conference in Las Vegas, he opened his AngelAi presentation with Queen's "A Kind of Magic" and The Buggles' "Video Killed the Radio Star," then walked the room through what his AI could do while the eighties hits played. It is an unusual style for a national mortgage CEO. It is also consistent with the work, which is built on the conviction that the technology should disappear behind the experience.

The cost of the architectural choice was real. Deterministic systems are slower to develop, less impressive to demonstrate, and considerably less interesting to talk about at conferences than the language models that dominated the AI conversation through 2022 and 2023. For several years, Agarwal's architectural conservatism looked unfashionable to observers who were not closely tracking what was happening inside regulated lending.

By 2025, the federal banking agencies and the CFPB had issued guidance establishing the standards Agarwal had been building toward for years: auditability, traceability, explainability in adverse-action notices, defensibility in front of regulators with the authority to challenge a specific decision. The standards were not a surprise to anyone who had been operating inside a lender. They were a surprise to the AI vendors who had been pitching probabilistic systems into financial services without understanding what the lenders' obligations actually were.

What Agarwal recognized, well before the rest of the field, was that the regulator's question and the operator's question are the same question. A system that cannot explain itself to a regulator also cannot explain itself to the borrower it just declined. A system that cannot reproduce its own decisions cannot be trusted by the lender that has to stand behind them. The deterministic architecture solved both problems by being a single answer to both questions.

It is not a coincidence that Sun West fully warrants every underwriting decision AngelAi produces. That arrangement would be impossible on a probabilistic foundation. A lender cannot warrant what it cannot trace. The architecture and the accountability are the same call, made for the same reasons.

The AngelAi logo is a ballerina. People ask Agarwal about the choice constantly, and he has a standard answer. "To become a ballerina is a lifetime commitment," he says. "A ballerina is someone who makes something that is, in actuality, very hard look easy. Artificial intelligence is much like a ballerina in that way. Getting a mortgage is difficult. AngelAi makes it look easy." The deterministic foundation is the hard part. The ballerina is what the borrower sees.