How Valon Rebuilt a $13 Trillion Industry From Scratch

Mortgage servicing handles $13 trillion in consumer debt, yet Linda Du called it one of the least disrupted industries, with technology that had not caught up to the 2000s. Andrew Wang framed the work as a systems-engineering challenge: large volumes of data and money movements must reconcile accurately. Legacy systems may preserve only a current snapshot of a mortgage, leaving important changes in ad hoc databases. That can prevent family members who take over a loan from accessing records they need, including state-tax information. AI may help interpret mortgage context, Wang said, but cannot recover history that was never represented correctly.

Valon chose to build both a servicer and its software from scratch. Its founders said incumbent servicers were unlikely to adopt new systems in a heavily regulated business, while buying a servicer risked inheriting legacy platforms and intellectual-property constraints. Operating as a servicer gave Valon a way to test its software against real loans and regulatory requirements; the system grew from one loan to close to a million. The company began before generative AI worked well. Because Valon could not initially obtain major government licenses, it started with non-QM loans that required state licenses. Some approvals depended on profitability, creating a chicken-and-egg obstacle. One speaker estimated stacked approvals could take at least three to five years; Valon received New York approval in about three years, described as a record, but not permission to originate mortgages there.

Valon built a compliance framework by reading federal rules and regulations across all 50 states and abstracting them into a common architecture. A speaker said they spent 18 hours a day for six months during COVID reading and annotating regulations, then tested the framework for five to six years. The servicer was designed to demonstrate that its technology could pass state exams and audits by Fannie Mae, Freddie Mac, and Ginnie Mae. Escrow shows why this structure matters to homeowners: servicers collect monthly reserves for taxes and insurance that are generally paid annually, and rising costs can trigger a sudden payment increase. A speaker said many systems recalculate over one year even when regulations and investor standards may allow changes to be spread over 36 months, a longer schedule that often requires ad hoc supervisor approval. Poor data models can also misrepresent changing insurance and mortgage products, creating errors that customer service cannot fix on its own.

Valon said its operation was about three times as efficient as alternatives, passed savings to mortgage asset managers through lower prices, and turned a breakeven business into one with a 70–80% operating margin. Its first large customer helped attract other large customers interested in the software. The market, rather than the founders, showed product-market fit: within six months of launching as a software company, Valon signed more than $200 million in deals. A close partnership with a mortgage company let it demonstrate the platform in an operating environment. Valon structures engagements around outcomes, earning when customers earn or save money. It later sold its servicing business to Carrington and shifted its focus to selling its operating system. Rhythm confirmed a transfer of four million loans to Valon, described as almost 10% of the market and the largest servicing transfer.

Valon says AI agents can automate servicing workflows that were previously too complex, while emphasizing that the system of record remains essential. Its agents can coordinate tailored outreach and plans for homeowners after disasters instead of relying on staff to divide loan lists and apply uniform processes. A speaker said Valon has dozens of agents for tasks such as escrow analysis and is testing strategies across loan portfolios, including an example involving 100 loans. Wang said mortgage AI must prioritize correctness and be willing to say “I don’t know” rather than hallucinate. Valon uses vector databases and context sequencing to retrieve relevant information for models. Because servicing links accounting, money movement, customers, regulations, and vendors, Wang said it cannot be optimized within a simple fixed boundary; Valon aims to evaluate and improve models in ways consistent with servicing practices and individual customer needs.

Valon chose servicing because it is sticky, difficult, and complicated, with the aim of expanding into other areas later; monthly contact also creates an ongoing relationship with homeowners. The platform could extend to commercial mortgage servicing, which involves property, tenant, and tenant-financial information and might support small-business financing; healthcare revenue cycle management was also named as a possible adjacent sector. Enterprise deployment, however, is a change-management challenge as much as a technology one, requiring alignment among stakeholders with different objectives. A speaker said about 75–80% of Valon’s management team had been at the company for five or more years and that the company had spent the preceding six to twelve months building change-management capabilities. The speaker expects change management to be the biggest value driver for AI over the next decade and said deployment roles require agency, comfort with ambiguity, customer empathy, perspective-taking, problem-solving, and people skills.