Key Takeaways

  • Valon Technologies raised a $150 million Series D round at a $2.3 billion valuation to modernize vertical mortgage servicing software.
  • Valon spent six years operating as a licensed mortgage servicer to prove its software in production before selling off the servicing business to become a pure software vendor.
  • The mortgage industry relies on core systems of record designed in the 1960s, prior to the creation of the internet.
  • Autonomous AI agents fail in legacy financial workflows because old architectures lack direct APIs and a verified single source of truth.

The Trojan Horse Approach to Regulated Software

Most enterprise software founders pitch products to risk-averse buyers who refuse to test unproven code on critical revenue streams. In mortgage servicing, where loans last up to 30 years and regulatory errors trigger heavy federal fines, banks will not buy core tools from an unproven startup.

Valon chose a harder path. Instead of selling software directly to banks on day one, the team spent six years operating as a fully licensed mortgage servicer. They handled real homeowner payments, managed escrow accounts, and resolved defaults directly on their own stack.

As Du explained, “The way that we decided to go to market, it took six years, is we actually built a servicer first, built the technology alongside that servicer, and then when the market made it clear that we'd kind of proven our point, we sold the servicer and today we're just a pure play.”

By absorbing the regulatory risk internally, Valon gathered real operational data and stripped away the main objection enterprise buyers had. Once the software managed billions in loans without regulatory penalties, Valon sold off the servicing book to focus entirely on software licensing.

Why AI Agents Die on 1960s Infrastructure

Silicon Valley is flooded with vertical AI startups pitching autonomous agents to enterprise back offices. Du points out that these tools fail immediately when dropped into legacy industries because the underlying systems of record are broken.

“The industry today really runs on a system of record that was built in the 1960s. So actually before the internet was invented,” Du stated. Mortgage servicing still depends on green-screen mainframes that run batch files overnight. There are no webhooks, no real-time event logs, and no clean programmatic interfaces.

LLM agents cannot complete complex actions if they cannot reliably query a database or trigger state changes through an API. “It turns out that if you don't have a system of record that is a single source of truth with very basic things like APIs and tools that an agent can call, it's kind of dead in the water,” Du noted.

Winning vertical AI requires rebuilding the unsexy database layer first. If you own the API that records the truth, you control the layer where autonomous agents run.

What to Do With This

Audit your product roadmap this week. If you are building an AI agent that sits on top of dirty spreadsheets or legacy third-party databases, stop working on prompts. Map out the three core data writes your agent must execute, and build a clean internal system of record with explicit API endpoints before writing more agent workflows.