Context
Mortgage servicing is the work of collecting payments, managing escrow, and helping homeowners after a loan is made. It covers about $13 trillion of consumer debt and still runs largely on software designed before the internet. In this episode, a16z General Partner Angela Strange talks with Valon's Andrew Wang and Linda Du about rebuilding that infrastructure. It matters to PMs because Valon chose the slowest go-to-market path, a regulated company that ran its own software first, and now deploys AI into some of the largest and most regulated enterprises there are.
The Big Idea
In a heavily regulated industry, the winning product strategy was to become the customer first: run the operation yourself, prove the software on your own loans, then sell it.
That proof gave Valon three times the efficiency of traditional servicing and a clean system of record. It is also what now lets AI agents handle work that was too complex to automate before.
Key Insights
Legacy systems lose the loan's history
- What: old servicing systems keep a snapshot of the loan today, not the full history of how it changed.
- Example: when a borrower dies and the kids take over, they can lose access to the records from before the change. They may need those records to file state taxes, so they repeat their story to the servicer again and again.
- Why it matters: Andrew's point is that AI can help find context, but the data must still be represented correctly in the system of record for anything to work as expected.
Bad data models compound for years
The systems were architected in the 1960s, so their data models do not match how things like property insurance and mortgage insurance work today. Over a decade, people force new cases into a model that does not fit them. Linda says the problem does not show up that week. It shows up five years later, when someone makes an escrow payment and is charged the wrong amount. Her conclusion: you can have the best, most empathetic call center agent, but you cannot fix that homeowner experience without fixing the infrastructure.
Rising escrow costs expose rigid software
- The issue: insurance and tax costs are exploding, and many homeowners cannot absorb a big escrow payment change quickly.
- The software gap: most systems calculate the amount over one year, even though regulation and many investor standards allow spreading it over 36 months.
- The result: the exception becomes a manual approval by a supervisor, instead of a configurable rule that can be changed on the fly when regulations change or a disaster hits.
Three paths and why two failed
Valon considered three ways in. Each had a clear problem, which is why they picked the hardest one.
- Sell software to incumbents: no servicer will use a brand new system in such a regulated field. A marquee customer would also shape the product to look like its current process, so you risk rebuilding the legacy system "with a really pretty UI".
- Buy a servicer and swap modules: this forces platform design decisions that are probably wrong, and it creates an IP problem with the existing platform.
- Start a servicer and build from scratch: the software evolved as the servicer grew, from one loan to close to a million. It is the most painful and slowest option.
Licensing was a chicken-and-egg maze
- Some regulators, including New York, expect a company to be profitable before it gets a license. Fannie and Freddie approvals have similar catches.
- Valon sequenced licenses on purpose. It started with non-QM loans because they only need state licenses, then dealt with California's stricter rules.
- Applicants must also show five-plus years of experience across collections, customer service, payments, delinquency, and foreclosure.
- Most people forecast three to five years at the minimum, since the approvals stack on each other.
- Andrew describes New York approval taking about three years, and visiting a mailroom to find a lost application so the clock would not reset.
Regulation became architecture
Valon read every federal rule (RESPA, TILA, FDCPA, GLBA, TCPA) and the mortgage, debt collection, foreclosure, privacy and escrow rules of all 50 states. Andrew calls it refactoring the legal code base: build an abstract framework that handles all of it, put it into the architecture, then test it. He spent 18 hours a day for six months on it during COVID, followed by five to six years of testing that it works as intended.
Efficiency was the product
For the servicer, unit economics was the product. Better technology meant more margin, and Valon passed that margin on to customers. Asset managers are "economic animals", so lower prices brought in loans. Linda says Valon is about three times as efficient, turning a roughly break-even business into one with a 70 to 80% operating margin. Running its own servicer also created urgency, because Valon competed directly with its future software customers, and it created safety because the system passed every state exam and Fannie, Freddie and Ginnie audit.
AI unlocks the long tail
- Before: you could only automate clearly deterministic steps, for which you already had structured data.
- Now: a much higher share of tasks can be automated, including rare and complex cases.
- Orchestration: after a disaster, a servicer can set up a group of agents to call homeowners, offer plans, and run different workflows per locale. Before, someone printed a spreadsheet and handed lists of loans to different teams.
- Champion and challenger: you can try different strategies on 100 loans, change notifications, and compare outcomes. That was not possible when each strategy needed human training.
Commodity work becomes specialized
Linda argues the real gain is customization more than raw efficiency. A servicer running on near-zero margins used to have only a cookie-cutter process it could not change. With agents, each servicer can decide how to handle a specific scenario, which turns a commoditized business into a more specialized, flavor-driven one.
Rethinking how humans use systems
Valon first built the system of record with defined UIs, and hired operators skilled at navigating those UIs. Now it builds its own harness on top of the system of record, so back-office staff can use a chat interface. Homeowners can reach the same system through voice AI or chat AI. Linda says there are now many more form factors than before.
Deployment is a change-management problem
Linda says that six years ago she would have called enterprise rollout a technology problem. Today she calls it change management: thousands of people, each with a slightly different objective function, and many layers to move through. Over the last six to 12 months Valon has built that muscle. She calls change at scale "the number one biggest value driver for the next decade" for AI. The people who thrive in this work have high agency, high ambiguity tolerance, and empathy for the customer. They can see different points of view, find the global goal everyone shares, and bridge between them. Linda adds that you need both strong problem solvers and people who are good with people.
Mental Models & Frameworks
Safety and urgency balance
Linda describes Valon's software strategy as balancing two forces.
- Safety: came from running its own servicer for six to seven years under real exams and audits.
- Urgency: came from competing in the market with the servicer, which created momentum.
- Flywheel: once the first big customer signed, the other big players called to ask if Valon was serious about selling software.
Everything is servicing
Servicing is the underlying critical infrastructure for anything with money movement, regulation, and an operational component. Andrew uses this to explain how the platform can expand.
- Commercial real estate: a commercial servicer sees tenants' financials, which could support small business financing.
- Healthcare: revenue cycle management is "just servicing for hospitals", with claims and patient data.
- Why mortgages first: it is the stickiest, hardest, and most complicated, so the rest gets easier.
Harness tuned for correctness
Valon builds its own harness on top of the system of record, with vector databases and thought given to the sequence of context. For mortgages, correctness matters more, so models are aligned to say "I don't know" instead of giving a hallucinated answer. Andrew also wants a continual learning loop that keeps running evals on real data. The loop must improve the model for the specific customer, not just for servicing in general.
Trade-offs & Nuance
Slow path versus fast sales
The route Valon chose took years of licensing, regulation reading, and testing before any software sale. The trade-off is clear: it avoids being a "pretty UI" on a legacy design and gives a proof point no pure software seller has. Linda says plainly that it is the most painful and longest option.
Keep the servicer or sell it
Founders could have kept scaling the servicer, since $10 billion plus companies do this. But keeping it means keeping the technology and the edge for yourself. Linda says every conversation ended with a wish to change the industry, so Valon sold the servicer business to Carrington and now sells the operating system to the industry.
Aligned pricing versus simple pricing
Valon structures deals around outcomes: it earns when a customer earns or saves a dollar. Linda says that alignment is a no-brainer for customers, and that losing it is when a company starts going downhill. The cost is that Valon's own revenue is tied to results it must keep delivering.
Practical Application
Own the hardest customer job
Before building for a regulated industry, ask whether your team should do the operational work first. Valon proved safety by running its servicer through state exams and agency audits. That gave sales credibility that no deck could.
Model history, not snapshots
Review your system of record for what it throws away. List the events, such as a borrower changing, that overwrite data without keeping the earlier state. Fix those first, because AI and support teams can only use history that was captured.
Make rules configurable, not manual
Find approvals in your product that exist only because the system cannot express a rule. Valon's example is escrow shortfalls: spreading a payment over 36 months is allowed, but most systems force a supervisor approval. Turn these into configurable policies.
Test strategies with champion challenger
If you have a workflow that agents can run, split a small group, such as 100 loans or users, into different strategies. Change one thing at a time, like the message or timing, and compare outcomes. Valon says this was not possible when every variation needed human training.
Staff for change management early
When planning an enterprise AI rollout, hire for both problem solving and working with people. Map the layers of the customer organization and what each group is trying to optimize before you ship, since Linda's team learned the rollout is mostly about moving people.
Questions to Consider
- Which parts of your own product's data model keep only a snapshot, and what history would a customer or support agent need that you have already lost?
- Which manual approvals in your workflow exist only because your software cannot express a rule that policy already allows?
- If you had to run the operation your product serves for a year, what would you learn that a customer interview would not show you?
- In an AI rollout at a large customer, who holds a different objective than you assume, and how would you find out before launch?
- Does your pricing reward you only when your customer earns or saves money, and if not, what would it take to get there?
Bottom Line
Valon won a $13 trillion market by doing the hard, regulated work itself first: licensing, reading every rule, and running its own servicer. That proof and a clean system of record now let AI agents take on complex servicing work, and the biggest remaining challenge is helping huge organizations change how they operate.
Concepts to Explore
Champion and challenger testing
A method where a current approach (the champion) is tested against an alternative (the challenger) on a slice of the portfolio. Andrew and Linda say agents make this practical for servicing outcomes because you no longer need to retrain human staff for each variation.
System of record
The authoritative database a business runs on. The Valon team argues AI is only as useful as the data in the system of record, which is why they built a new one instead of adding AI on top of a flawed legacy core.
Escrow
Part of a mortgage payment held in reserve for property taxes and insurance, paid out annually. It is the source of several examples in the episode, because rising insurance and tax costs make rigid escrow software painful for homeowners.
Case Studies Mentioned
Valon from servicer to software
Valon started a mortgage servicer, grew it from one loan to close to a million, and used that to prove its platform. The servicer was then sold to Carrington, and Valon now sells the software to the industry. Linda says that within six months of going to market as a software company it signed over $200 million of deals. She also says a large servicing transfer of about four million loans, close to 10% of the market, was confirmed with Valon. The lesson is that product-market fit was decided by the market, and the time spent as a servicer built the evidence.
Tools & Products
| Tool / Product | What it does | Why it was mentioned |
|---|---|---|
| Valon OS | Mortgage servicing software and system of record | The platform Valon proved on its own servicer and now sells to the industry. |
| Voice and chat AI agents | Let homeowners and back-office staff interact with the system | Examples of new form factors on top of the system of record. |
People to Follow
Andrew Wang
Valon's CEO and a former investor at Soros who spent time building reconciliation systems for mortgage money flows. He also still works as one of the most productive engineers at the company.
Linda Du
Valon co-founder who speaks on go-to-market, enterprise deployment, and change management for AI in regulated industries.
Angela Strange
a16z General Partner and host of this conversation, known for investing in infrastructure.
Notable Quotes
"You can have the world's best customer experience team and the most empathetic call center agent in the world. But if fundamentally you were charged the wrong amount of money because the system of record was flawed, then you can't fix that homeowner experience without fixing the infrastructure." (Linda Du)
"If you had asked me six years ago, I would have told you, this is a technology problem. What I know today is that this is a change management problem." (Linda Du)
"It turns out revenue cycle management is just servicing for hospitals." (Andrew Wang)
