Fraud, Mortgages, & the Art of Judgment
Happy Wednesday, Fintech Listeners!
Only a few more Fintech Takes podcasts in 2025!
If you haven’t listened to enough this year, make sure to A.) listen to today’s pod (it’s excellent), and B.) Watch for Friday’s newsletter, which will feature the 12 most popular Fintech Takes podcasts of the year!
— Alex
P.S. — Do me a favor and also make some time to watch The Princess Bride or When Harry Met Sally or A Few Good Men or The American President or any of Rob Reiner’s other films.
3 BIG IDEAS FROM THE PODCAST
This week’s episode featured Simon Taylor and me doing what we do best: talking about startups we’re absolutely not giving investment advice on.
There was fraud detection built on federated learning. Agentic AI for mortgage apps. Document intelligence for small business lending. And one ambitious company building the infrastructure for SBA-style guarantees in India.
So yes, we talked a lot about startups. And broken workflows, too!
Tune in for the full conversation here
And read below for my three big ideas...
#1: The Memory Hole in First-Party Fraud DetectionCopy anchor linkCopied
Today’s fraud systems are built to block abuse, not to recognize trust.
That distinction is critically important in first-party fraud; when a real customer makes a false claim, like intentionally defaulting on a loan or requesting a chargeback they’re not entitled to.
First-party fraud doesn’t look like fraud. It looks like a loan default or a support ticket. It’s very difficult to tell, for a one-off transaction, if the person you’re dealing with (who is experiencing a problem) can be trusted.
Why, in the iterative, data-rich system that is the modern financial services ecosystem, is there no way for companies to find out which customers play fair?
Why is there no shared infrastructure for trust?
That’s what makes the network that Trudenty is building so interesting.
Instead of centralizing raw data, they use federated learning to let card issuers, PSPs, and merchants exchange signals without compromising privacy (Simon: “You don't get to see my data, but I do get to learn from yours.”) Everyone sees the patterns; no one has to cede control.
It’s a signal-sharing layer that lets good behavior travel, so when you dispute a charge, you’re already known quantity (and that context changes everything).
Trudenty is working with Worldline, JPMorgan Chase, and Mastercard. And their pitch is just as focused as their architecture: they sell one thing, and that one thing works.
The stat that stuck with us? 80% of chargebacks are actually fraud.
But solving first-party fraud doesn’t mean catching every lie, but recognizing patterns of honesty.
A customer says the product never arrived. The merchant says it did. The truth lives in gray areas. Historically, this has led to closed, vendor-specific consortia where fraud prevention becomes a blunt instrument.
The deeper tension is who gets to decide what fraud looks like. A PSP may see something as risky. A merchant might view the same behavior as normal.
The only way to make a trust network work is to let each participant access the signal at their chosen altitude: raw metadata, machine-learned features, or a simple red/yellow/green risk score.
#2: AI That Keeps the Loan Alive Copy anchor linkCopied
The mortgage industry keeps trying to reinvent the wheel.
Every few years, a new wave of tech promises to fix the point of sale (POS) experience. And while we have come a long way — mortgage POS systems (Blend, Encino, ICE, etc.) are all faster and far more digitized than they used to be — there are still so many edge cases where applications can (and do) fall out of the automated process. The moment something breaks, like a document upload that’s “slightly off center” or missing information, the workflow needs someone to catch it and move it forward.
That’s the front-end failure point. And it’s the exact seam where agentic AI can shine.
The startup Tidalwave is working on this layer. But what’s most compelling is the sequencing (as opposed to the tech). They’re building autonomous agents that operate like a 24/7 support staff to smooth out the rough spots (like connecting to data sources when something goes wrong, asking follow-up questions, and orchestrating the handoffs).
They’re strategically avoiding the regulatory risk that comes with touching the loan underwriting process (i.e., the loan origination system or LOS), while reinforcing the top of the funnel. Agentic AI probably shouldn’t make lending decisions. It just keeps the applications alive.
Tidalwave has raised $22M, with the largest homebuilder in the U.S. on the cap table. That tells you who understands the pain.
The real question, as Simon put it: can this AI outperform an entry-level loan officer who gets frustrated, doesn’t speak other languages, and needs to sleep?
If the answer is yes (and it should be yes!), AI’s big breakthrough isn’t as a brain, but as an always-on switchboard operator.
#3: You Call This a Credit Problem?Copy anchor linkCopied
Ask a lender why they don’t do more small business loans, and they’ll say the math doesn’t work.
The risk? Manageable. The returns? Acceptable. The problem, then, is operational: it takes the same manual effort to underwrite a $100K loan as it does a $5M one.
The credit risk isn’t the core issue; the unit economics of getting to the answer is.
Think about the revolving credit lines, equipment financing, and commercial mortgages. These loans require judgment, but also lots and lots and lots of paperwork. And there’s no universal layout or consistent structure for said paperwork; there’s no underwriting checklist that works every time.
So underwriters improvise. They structure documents by hand, patch together memos from scratch, and eventually stop doing that work on smaller loans because it isn’t worth the effort. Asking highly paid credit professionals to spend their time parsing tax returns and standardizing inputs, after all, is pretty insane.
The chokepoint in SMB credit isn’t demand or capital. It’s the bridge between input and judgment.
That’s the opportunity Kaaj is targeting (another company using agentic AI to support the human in the loop). The AI doesn’t make a decision; the underwriter still does. And faster, since the files are prepped for human review.
🎬 DIRECTOR'S COMMENTARY
Simplifying and speeding up the small business lending workflow using agentic AI is one of the hottest areas in early-stage fintech right now. And unlike other hot spots, such as compliance, there’s a strong link between an improved SMB lending process and more revenue, which makes the business case a much easier sell.
WHAT I'M LISTENING TO
#1: Trader Joe’s (Acquired) 🎧Copy anchor linkCopied
I love Trader Joe’s. I love the Acquired podcast. I loved this Acquired podcast episode about Trader Joe’s.
Sometimes the job of content curation is very simple.
#2: How Nubank Rewrote Brazil’s Banking Rules with Cristina Junqueira (This Week in Fintech) 🎧Copy anchor linkCopied
I saw this one when it was recorded live from Miami at Fintech NerdCon, and I can tell you that it’s very much worth a listen.
Thanks for the read! Let me know what you thought by replying back to this email.
— Alex
Join 2,444 other finance and fintech leaders in the Fintech Takes Network
The conversation doesn't have to stop here
Keep learning and connecting in the Fintech Takes Network
EVENTS | FEED | LIBRARY | DIRECTORY
