Mortgage AI has a Proof Problem

MBA Annual was full of robots, AI evangelists & vendors promising to save lenders a fortune. Actual evidence of those savings? Ehhh, not so much.

Which makes Karen Postiglioni's column today especially timely. Before mortgage spends another quadzillion dollars on AI, maybe we should figure out what's actually broken, measure the results & make sure we're not just automating dysfunction?

Plus, GSE repurchase dollars are climbing, there's a bit of a shakeup at Lower, & I've got a little more MBA Annual gossip for ya, including party reviews & a humanoid robot that I swear to God was stalking me. 🤖

Also, I caught the flu 🤒 at Annual, so apologies for not being able to send an edition on Wednesday. It’s been brutal!

Anyway, let’s get on w/ it!

By Karen Postiglioni 

I spent the last few days in Chicago for MBA Annual, and I think it's safe to say we're officially living in the future.

There were two humanoid robots there. One even had its own badge… better creds than me.

At another event, facial recognition technology identified me in photos and sent me every picture I showed up in. Front-facing, side profile, didn't matter. It found me.

My Life360 app even alerted me in real time to a reported fight or assault (yikes! I still love you, Chicago) happening just a couple blocks away.

I'm convinced we're not far from wearing AI glasses at conferences that tell us someone's name, company and probably the last three conversations we had before they've even walked up to say hello.

And in mortgage, an industry facing no shortage of economic headwinds, everybody’s looking for an advantage. A way to differentiate, reduce costs, improve margins, increase productivity or just survive until conditions improve.

I'm all for it. In fact, I think we need more experimentation and more willingness to challenge how we've always done things.

But now is the time to get creative, not desperate. 

Because there's a difference between adopting technology and improving a business.

And with all the AI solutions entering our industry, I think it's worth revisiting some basic operational disciplines that can get lost in the excitement.

First, start with the problem, not the solution.

Not every problem needs AI. Sometimes it needs a better process, cleaner data, fewer handoffs or someone to figure out why your processors are still maintaining that spreadsheet nobody wants to admit exists.

Before evaluating a solution, can you clearly articulate the problem, quantify its impact and identify what success would look like?

If not, why are we shopping?

Second, establish the baseline before touching anything.

You can't measure improvement if you don't know where you started.

How many touches does a loan require today? How long does the process take? What's the defect rate? How much rework are you creating? What does it actually cost? 

Figure out your KPIs tied to what you’re changing BEFORE you change it. And make sure you're comparing apples to apples. Loan complexity, volume and product mix matter.

Third, show your math. All of it.

A vendor tells you its solution will save 20 minutes per loan. Awesome.

But what happens to those 20 minutes?

Did you reduce overtime? Increase capacity without adding staff? Improve turn times? Increase pull-through? Eliminate rework? Or did you just give someone 20 more minutes to do something else?

There's value in all sorts of operational improvements, but time saved isn't automatically expense eliminated.

And the cost of the solution isn't just the subscription. It's implementation, integration, training, maintenance, oversight and whatever other technology you now need to make it work.

Projected ROI is one thing. Realized ROI is another.

Fourth, measure adoption, not just deployment.

I've seen plenty of so-called successful implementations that weren't particularly successful operationally.

Is anybody actually using the solution? Do employees trust it? Are they finding workarounds? Are they performing the same work twice because they don't trust the output?

Did we ask the people doing the work what they needed before deciding what to buy?

A shiny new system nobody uses is just a more expensive version of the old problem.

Fifth, know what you’re buying and who is responsible.

Whether you build it or buy it, somebody needs to own the outcome.

Who is evaluating accuracy, risk, compliance and vendor performance? Who understands what the technology can and can’t do? What happens when the model changes or produces a bad result?

Moving quickly doesn't eliminate the need for governance. 

And finally, measure again after implementation, and be willing to stop.

This might be the most important discipline of all.

We spend so much time building the business case to get something approved. How often do we go back 30, 60 or 90 days later and compare what actually happened against what we projected?

Did performance improve? Were the savings realized? Did quality suffer? Do we need to adjust, expand or maybe even discontinue the solution?

There should be no shame in an experiment that didn't deliver the expected results. There should be concern when nobody bothers to find out.

And here's another question worth asking: Are we using AI to fundamentally improve the way we operate, or just automate processes that weren't particularly efficient to begin with?

You can automate document collection, underwriting reviews and communication. But if you’ve built your process around unnecessary handoffs, duplicate work and disconnected systems, you might just be automating dysfunction.

Making a bad process faster doesn't necessarily make it a good process.

I came away from Chicago excited about what's possible. There are some genuinely creative people and companies working on solutions that could change how this industry operates.

But possibility alone isn't a business case. Move fast. Experiment. Challenge the status quo. Just bring the same discipline to evaluating the results that you brought to justifying the investment. Urgency is still not an excuse to abandon discipline.

Possibilities don’t pay the bills, execution does. Let's make sure we're getting our money's worth.

GSE repurchase activity increased in Q2'26 with 1,401 loans totaling $465.2 million in unpaid principal balance ("UPB") repurchased during the quarter, an increase of 19.5% from Q1'26.

Fannie Mae accounted for 58.2% of repurchase UPB in Q2’26 with Freddie Mac accounting for the remaining 41.8%. During the quarter, Fannie Mae issued 887 new repurchase demands, compared with 883 issued by Freddie Mac.

Among resolved repurchase demands—defined as repurchased loans plus withdrawn demands—83.2% of Fannie Mae demands resulted in repurchase, compared with 53.1% of Freddie Mac demands.

Loans issued in Q3’25 and Q4’25 accounted for 46.5% of total repurchase activity, making those two issuance vintages the largest concentration.

Rocket Mortgage, United Wholesale Mortgage, and Lakeview Loan Servicing were the three largest originators by reported repurchase share, accounting for 9.7%, 5.4%, and 2.4% of total reported repurchases, respectively. For comparison, those originators accounted for 13.9%, 5.5%, and 2.8% of new issuances in Q3’25 and Q4’25, respectively.

This report is brought to you by Milliman. Milliman provides strategic and quantitative consulting services across the mortgage market with expertise in origination, servicing, and capital investment.

Source: SEC; Note: includes MBS issuances since 2020 and excludes multifamily repurchase activity

Adam Wiener, the former Redfin exec who joined Lower in Jan. ‘25, has left the company. 

Recruiter & all-around good guy Connor Bartley is exiting, too. Bartley is moving on to a new company to run growth. 

Dan Snyder told me Lower’s doing well & would make its run during the downturn. Even after losing a big team to Atlantic Coast Mortgage, the RETR stats suggest Lower is growing in a down market. Lower just pulled a $300M+ team from Supreme Lending…

Mortgage tech gossip, questionable AI promises, the highs & lows of MBA's party circuit & a humanoid robot that I swear was stalking me. 🤖 The really juicy stuff is behind the paywall, so plz consider upgrading your sub today.