Happy Wednesday, Fintech Listeners!

I’ve got a great podcast for you today on one of my all-time favorite topics, so let’s get right to it!

— Alex

3 BIG IDEAS FROM THE PODCAST

This week on the podcast, David Snitkof (GM of SMB at Ocrolus) makes his Fintech Takes debut, although we’ve been chatting about small business lending for years, often in between basketball games. (Annoyingly, he’s also a fantastic basketball player.)

Small business lending is one of my all-time favorite problems in financial services because it’s so freaking hard to solve. Every fintech lender who's tried to crack it has been trying to answer one version of the same question: How do you deliver context at a price that the loan can support?

Nobody's fully solved it yet. Which is why I wanted David to come on the podcast and talk about it.

And read below for my three big ideas...

#1: Friction Is Part of the ModelCopy anchor linkCopied

In our conversation, David flagged the importance of negative selection.

Small business lending, as he puts it, is decision making under uncertainty and asymmetry: You make a little money on a good loan, you lose a lot of money on a bad loan. So naturally you want more context before you commit. But push too hard on a strong borrower with options — requesting months and months of bank data, an in-person interview, etc. — and you get Nah, I'm not going to give you all that data or sit through an in-person interview. I'm going to go to another lender.

The customer left standing at the end of that gauntlet might be, in his words, the one more desperate for financing than anyone else in the initial applicant pool.

Negative selection is obviously a risk in any type of lending, but small business owners amplify it. Nobody starts a carpentry/pizza making/septic system installation business because they love administrative back office work. They tolerate that work because they have to. And when they need the money, they’ll optimize for speed and convenience over pricing and service.

Ocrolus and OnDeck's own Small Business Cash Flow Trend Report backs up this instinct: 74.5% of surveyed borrowers skipped a traditional bank lender to apply to a non-bank lender first, and the hassle of paperwork was the #1 reason cited (at 46.2%).

There’s another detail I can't stop turning over from our conversation, though. Ocrolus still processes ~1.25 million pages of PDFs a day (a day!), and the reason isn't purely technological. David described borrowers uploading statements to a lender they didn’t trust enough yet for a digital bank connection, getting underwritten on paper initially, and then linking their actual bank account only once they knew the money was close to hitting. The PDF is clunky. But as the clunkier option, it may still be the one that the borrower prefers, at least at first.

Which means UX isn't something layered on top of underwriting after the fact. Change the process, and you change which borrowers finish it, and what they're willing to hand over along the way. The experience is already shaping the credit decision before the risk model ever gets involved.

#2: The Context / Margin ParadoxCopy anchor linkCopied

Here's a theory I've had for a while, and David more or less confirmed it: Small business lending is hard because it requires both context and operational efficiency.

I call it the Context / Margin Paradox.

At the top end of commercial lending, understanding context is expensive, but the economics work fine. A large commercial loan can support what David calls a fairly bespoke process: Experienced humans spending real time learning the company and the environment it operates in.

Consumer credit has thin margins, but the need to understand context isn’t as pressing as it is in commercial lending because (from a risk perspective) consumers are easier to understand than businesses. As such, it can be done in a highly efficient and automated manner.

Small businesses get stuck in the middle. They are much less homogenous than consumers (different industries, different sizes, different geographic locations, etc.) and they are constantly changing (pivoting into new industries, changing their names, going out of business, etc.) which makes them very difficult to lend to without a full understanding of their individual contexts. However, unlike larger commercial loans, there’s very little margin! Small businesses can’t afford to pay for the time it would take a human underwriter to properly understand them.

Lenders have responded to this paradox over the years by choosing different compromises.

Traditional lenders kept the manual underwriting and did their best to cost-engineer around it. Fintech lenders like Kabbage, OnDeck, and Bluevine took a page out of consumer lending’s book and tried to build a fully automated process, even if it meant excluding certain businesses or loan product categories.

Compromise was always necessary because the core trade-off — context or operational efficiency — was never resolved.

#3: The Silicon UnderwriterCopy anchor linkCopied

Over the last couple of decades, lenders have wrung a lot of value out of data … but only the data that fits inside a very specific box.

Traditional machine learning models want data that behaves, that has a standard structure and distributions that remain reasonably stable over time.

Consumer credit data is like this. Small business data is not.

Small business data is more symptomatic of the real world.

Take David’s example of an ice cream shop in Connecticut making $50,000 a month.

Is that good? Is that bad?

It depends on the context!

How much money does the shop make in the winter months? How stable is its revenue in the summer months, year over year? How does it stack up against other shops of the same size and same age in the same region?

These are the questions that human underwriters are good at asking. And if they’re sufficiently experienced at underwriting ice cream shops in the Northeast, they develop an intuition for gathering and understanding context that no ML-powered risk model or deterministic set of decisioning rules can replicate.

But, as we’ve already said, humans are expensive.

AI agents, David said, are very good at handling context, which may enable the type of personalized analysis that previously required those highly-paid humans. David’s concept of a silicon-based underwriter is, in theory, more observable and cost-effective than what he calls the “ultimate black box of the human mind.”

And beyond more efficient underwriting, it’s possible that AI agents will reshape the small business lending journey in even more profound ways.

Today, that journey is fairly linear. The borrower has to recognize the need for credit and initiate the process as the lender assembles enough context to decide what to offer. The loan application is the triggering event that makes all of this happen simultaneously.

David imagines something less episodic in the future, like a borrower-side agent that already understands the business well enough to determine what kind of financing fits and when it’d be useful. Lenders could have agents carrying their own contextual understanding into underwriting, and maybe eventually there’d be a panoply of agents working together, removing the need for a triggering event at all.

WHAT I'M LISTENING TO

#1: How the Largest Health Insurer in the US Came to Own a Bank (Bank Nerd Corner) 🎧Copy anchor linkCopied

Two of my Workweek colleagues talking about a really interesting company that sits at the cross section of their different beats.

This one earns an automatic recommendation from me.

#2: Katie Perry, Author of Ticker Shock (Fintech Business Weekly) 🎧Copy anchor linkCopied

I haven’t read Katie’s book yet, but this conversation got me excited to crack it open.

Thanks for the read! Let me know what you thought by replying back to this email. 

— Alex