You need courage for what's coming

A conversation with Mike Yu of Vesta

The last wave of mortgage technology companies raised billions of dollars and promised to make a loan cheaper to produce.

They went public. They put real engineers on a real problem, and the process they set out to fix genuinely deserves fixing. I respect anyone who spends their capital and their years on it. But across that decade, the cost to make a loan didn’t fall. It roughly doubled, to a record north of $12,500, against a long-run average closer to $7,000, by the Mortgage Bankers Association’s numbers. Most of those companies now trade more than ninety percent below where they listed.

Here’s the part that matters for what comes next. The technology mostly worked. The screens got modern, the data moved, borrowers got slick portals. It just didn’t make the loan any cheaper, because it digitized the old workflow instead of shrinking it. New system. Same machine.

I think this time might actually be different. Not because the demos are better, or because every lender now says the word AI. Because for the first time, the software can do the work the last generation could only route around. The last wave made the mortgage process more digital. This wave may make it smaller.

To see why, you have to understand what the work actually is. So I called someone I respect on this and asked him.

Adding Is Easy. Subtracting Is Hard.

Mike Yu co-founded Vesta, a loan origination system a growing number of lenders are moving onto. PennyMac runs on it. He spent four years inside one of the last wave’s companies before leaving in 2020 to build something better, and raised $55 million to do it, from Andreessen Horowitz and Bain and the kind of investors who don’t usually return mortgage tech’s calls.

He’s the person whose entire job is to sell you on technology. So I expected the pitch. Instead he told me the software was never the hard part.

“Adding is really easy once the work is being done more than one time,” he said. “Subtracting is really hard.”

Here’s the example he gave me. An underwriter opens a file and catches a large deposit on a bank statement that needs sourcing. The processor had already been through that same statement and missed it. Not out of laziness. The processor thought the job was to confirm the account holder’s name matched the borrower. And the underwriter is doing the very same check. “The underwriter, I promise you, is making sure the account holder matches the borrower name,” Mike said. “So what are we doing here?”

The work is built so the person downstream redoes what the person upstream already did. Add an automated tool to that line and nobody notices, because there’s already a human doing it twice. Try to remove one of the humans and you find a person attached to the job, a manager who built the team, a risk officer who asks what happens the one time it goes wrong, and a meeting to revisit it next quarter.

Adding is invisible. Subtracting has a victim.

That is why a decade of software never moved the cost. The tools got added on top. The duplicate human work underneath never got removed, and the human work is where the cost lives.

What AI Can Do That Software Couldn’t

Here is the shift, and it’s the whole reason I think the next few years break differently.

Ask Mike what underwriting most loans actually involves, and he’ll tell you it isn’t judgment. “We don’t really underwrite loans in this industry,” he said. “We make the loan fit in a box somebody else already defined.” You take the borrower’s documents, check them against Fannie or Freddie’s rules, and if it fits, you ship it. “You’re reading documents, you’re auditing documents, and you’re running a bunch of rules,” he said. “There’s no reason software can’t do all of that.”

That sentence is the difference between the last wave and this one. Reading a document and deciding whether it matches a rule is exactly what the previous generation of software could not do. So it did the next best thing. It moved the document to the right person faster, put it on a nicer screen, and waited for a human to actually read it. It routed the work. It couldn’t perform it.

AI can perform it. Mike told me Vesta already has lenders using agents to underwrite certain loan types start to finish. Not purchases yet, he is quick to say. But the nature of the task, read, check, compare, verify, is no longer something only a person can do.

He gave me a second example. For years the industry tried to automate the back-and-forth with title companies by getting everyone onto the same software. It never worked, because title is a thousand fragmented shops that run on email. “All that stuff didn’t work,” Mike said. “You can just layer AI on top of email, and you don’t need the software layer anymore.” The old approach demanded the whole world change its tools to match yours. The new one reads the mess as it is. Software needed the world to conform. AI doesn’t.

That’s what makes this wave different in kind, not degree. The last one made operations digital. This one can make them smaller.

The Hard Part Is No Longer Software. It’s Courage.

Which moves the bottleneck somewhere uncomfortable.

If the software can finally do the work, the question stops being technical. It becomes whether a company will actually remove the work. Pull out the second reviewer. Redesign the process around the exceptions instead of running every file through every step. Accept that an agent now does what a trusted person used to do, and own the risk on the day it’s wrong.

That is not a software problem. That is a courage problem. And it runs straight into the victim. The second reviewer is a person. The complexity is somebody’s whole role, and plenty of good people quietly tie their worth to being the one who handles it. Take the complexity away and you take away the reason they felt important. People fight that, sometimes without admitting it to themselves.

I’ve watched a company build a workflow specifically to kill a piece of duplicate work, and then keep doing it anyway, because the people whose job it was kept doing it out of habit and nobody had the stomach to make them stop. The tool wasn’t the obstacle. The will was. Mike calls it a cultural re-litigation companies have to go through, and most won’t until they’re forced.

So the winners of this wave won’t be the companies with the best AI press release. They’ll be the ones with the courage to subtract the work, take people off the file, and live with a leaner machine. That is a far rarer thing than a roadmap.

The Next Five Years Will Not Be Even

Mike is careful about the clock. “In eighteen months it’ll still look pretty similar,” he told me, because change moves slowly through a business this fragmented. The danger is reading that as permission to wait.

Three to five years out, he thinks it’s a different business, and the gap between the lenders who started and the ones who didn’t won’t be closeable. He explained why with a scene every operator has lived. Your vendor says the project ships June 15th. On June 13th they tell you it’s actually July 31st. “Why didn’t you tell me on May 15th you were behind? They thought they could catch up. Everyone thinks they can catch up.” And mostly, he said, they can’t. The head start compounds.

In that window, lenders are going to look wildly different from one another. Some will bolt AI onto the same old process and change nothing that matters. Some will keep paying humans to do what an agent should be doing. Some will rebuild around the exceptions and actually take the cost out. For a few years, two companies making the identical loan will do it at costs that aren’t close.

Then it converges, the way it always does. When Rocket sold push-button-get-mortgage, everyone eventually got a slick point of sale. The market drags everyone to the same place eventually. But not everyone makes it. And before it does, there’s a stretch where the difference between a company that subtracted and one that didn’t is the difference in your rate sheet, your comp, and your borrower’s cash to close. Where you hang your license during that window is going to matter more than it has in twenty years.

I’m rooting for all of it. But rooting isn’t believing, and the lenders betting on this, including the ones moving onto Vesta, may be right or may not be. Only one thing will tell you who actually pulled it off.

Cost Per Funded Loan Is the Scoreboard

So how do you tell, from the outside, which company truly changed and which one just changed its slide deck? The demo won’t tell you. The cost will.

This part is mine, not Mike’s. You watch one number. The cost to make a loan.

The biggest driver of that cost is productivity, the number of loans each person can carry, and productivity is the only thing technology is legitimately for. Not a nicer interface. Not a better portal. Loans per person. When a company truly subtracts work, productivity climbs and cost per funded loan falls, and there is nowhere for it to hide. When it doesn’t, you bought a screen.

You can already see that scale and a modern stack, on their own, buy almost nothing. By Richey May’s 2025 numbers, the lenders producing over three billion dollars a year make a loan for about $12,188. The industry average is about $12,349. A hundred and sixty-one dollars. Only subtracted work moves that number, and very few have achieved it or even made progress.

I hold vendors to exactly that standard, because I learned it the hard way. Everything anyone tried to sell me promised to increase productivity, so now I make them prove it with data, on real loans. I’ve been burned too many times. It looked great in the demo, then six months into implementation our productivity hadn’t moved, and more than once it had gone backward.

And the market is getting harder to fool. Borrowers can now drop a loan estimate into a tool, have lenders bid against it in the background, and get a verified competing quote inside a day. Even realtors, who never used to care about rate, are starting to, because referring a borrower who then finds a better deal makes them look bad. And it’s easier than ever for borrowers to get multiple quotes. What wins in that world is price, and price is cost.

Ask for the Benchmark

A friend of mine invests in distressed corporate debt at D.E. Shaw. He told me how they decide what to buy.

They ignore the story.

The marketing. The decks. The turnaround narrative. All of it.

Then they read the financials.

Because the financials don’t lie.

You can do the same, and you don’t need permission. When the next company walks you through its AI, don’t argue the pitch. Ask for its benchmark cost data. Most lenders benchmark against STRATMOR or Richey May (or they should). Ask where they land on fulfillment cost per loan and G&A cost per loan. They don’t have to show you their profits or anything private, just the benchmark. A company that has actually taken its cost down will show you in a heartbeat. One that’s bolting AI onto the old machine will tell you it’s confidential, or change the subject back to the platform. Either way, you have your answer.

The demo shows you what they bought. The benchmark shows you whether it worked.

Read the number.

Rich Weidel
CEO, Princeton Mortgage