The context margin paradox: why AI changes the math on SMB underwriting
TL;DR: On the Fintech Takes podcast, host Alex Johnson and Ocrolus SMB general manager David Snitkof discuss the context margin paradox: small business credit depends on deep context about each business, but small loans lack the margin to pay human underwriters to gather it. Snitkof covers four ways lenders have approached the problem, why so much of SMB lending still runs on PDFs and why cash flow data needs industry context. He argues that AI pre-trained on massive data could bring traditional underwriting techniques down market.
Small businesses have reported for years that they don’t get as much credit as they want, at the rates or terms they prefer. On a recent episode of Fintech Takes, host Alex Johnson workshopped a theory about why with Ocrolus SMB general manager David Snitkof. Johnson calls it the context margin paradox. Commercial lending is deeply contextual, and understanding context has historically required human loan officers and underwriters, who are expensive. Large commercial loans carry enough margin to absorb that cost. Small business loans don’t.
Why SMB underwriting needs context it can’t afford
Johnson sees two differences from consumer lending: companies aren’t homogeneous, and they aren’t static. Small businesses pivot, rename, change industries and close at a pace consumers never do.
Snitkof added the economics. “Lending is very much about decision making under uncertainty and asymmetry,” he said. A lender makes a little on a good loan and loses a lot on a bad one, and the range of business types magnifies the uncertainty. Expertise doesn’t transfer either. “The person who really knows how to underwrite a dentist is not the person who really knows how to underwrite an ice cream shop or a small manufacturer.”
Asking for more data has limits. A good applicant with options, Snitkof said, won’t hand over 100 months of bank data or sit through an in-person interview, and the lender risks selecting the most desperate borrowers.
Four ways lenders have tried to solve it
Snitkof’s view is that the fundamental lending model hasn’t changed. Lenders make money on interest and fees, lose it to credit and fraud losses, and spend it on operations, acquisition and capital. Within that model, he has seen three approaches and the start of a fourth.
- Cost-engineer traditional underwriting. Keep manual review and heavy documentation, and lower the cost with technology or outsourcing.
- Adopt a consumer-like model. Rely on consumer data and consumer-style machine learning, then make the margin work through high rates or scale.
- Automate with sophisticated analysis. The approach of most fintech lenders over the past 15 years: automation, cash flow underwriting, targeted marketing and dynamic customer management.
- Use AI agents that can take in a lot of context. Deliver personalized analysis that previously required highly paid experts.
Why so much of SMB lending still runs on PDFs
Johnson compared small business owners to new parents: exhausted, under pressure and optimizing for speed and survival over APR. Snitkof pointed to the quarterly small business survey Ocrolus runs with OnDeck, which has shown more owners going directly to non-bank lenders over time, citing speed and convenience.
Speed coexists with a lot of paper. Ocrolus processes about 1.25 million pages of PDFs a day, Snitkof said, on top of its digital data. Applications often arrive through brokers who collect bank statements and send them to funders. Commercial bank account coverage among digital aggregators lags consumer coverage. And many owners won’t connect accounts digitally until they know they’re approved, then connect right before funding.
Cash flow-based underwriting still needs industry context
Accounting data is easy to connect, but its quality drops as businesses get smaller. Some of the only small businesses that are great at bookkeeping, Snitkof noted, are bookkeepers. Across small businesses, he said, cash flow analysis of bank transactions “is just the best way to truly understand the nuanced operating dynamics and financial health of a small business.”
Reading that data still takes industry knowledge. Take an insurance credit: “For most businesses, that’s not revenue. It’s just a claim payment. However, if you’re a dentist or an auto body shop, that’s how you make your revenue.”
Snitkof also described how industry benchmarking adds peer context. An ice cream shop doing $50,000 a month means little alone. Against peers in the same industry and geography, it might rank in the 63rd percentile for revenue and the 41st for payroll.
What AI agents change for SMB lenders
Traditional machine learning needs large, standardized data sets. Small business data is fragmented. “AI that has been pre-trained on massive data now unlocks the power of small data,” Snitkof said.
He predicts traditional underwriting techniques will come down market, but instead of a “carbon-based underwriter,” lenders will have a “silicon-based underwriter,” which he described as in theory more observable, reliable and cost-effectively scalable. He expects agents for business owners that seek the right credit at the right time, and agents for lenders across prospecting, underwriting, fraud, servicing and renewals. Personalization at that level needs a new compliance framework, since compliance has relied on standardization.
Both agreed the industry is early. Johnson called small business lending the most persistently thorny problem in almost all of financial services. Listen to the full episode, and explore how Ocrolus supports small business funders.
Key takeaways
- The context margin paradox: commercial lending depends on context, gathering context has historically required expensive human underwriters and small business loans lack the margin to pay for it.
- Snitkof sees three established approaches (cost-engineered traditional underwriting, consumer-like models and fintech automation with cash flow underwriting) and an emerging fourth built on AI agents.
- Paper still dominates: Ocrolus processes about 1.25 million pages of PDFs a day, driven by broker channels, lagging commercial aggregator coverage and owners who connect digitally only right before funding.
- Cash flow data needs industry logic. An insurance credit is a claim payment for most businesses and revenue for a dentist or auto body shop.
- AI pre-trained on massive data “unlocks the power of small data,” which Snitkof expects to bring traditional underwriting techniques down market.
FAQs
What is the context margin paradox in small business lending?
The context margin paradox is Fintech Takes host Alex Johnson’s term for a structural problem in SMB lending. Commercial credit can only be underwritten safely with deep context about the business, and understanding that context has historically required expensive human loan officers and underwriters. Large commercial loans carry enough margin to cover that cost. Small business loans don’t.
Why is small business underwriting harder than consumer underwriting?
Consumers are relatively homogeneous and move through predictable life stages, and consumer credit bureaus have strong coverage. Small businesses differ by size, location and industry, and they pivot, rename, relocate and close far more often. That makes small business data much harder to standardize.
Why do small business lenders still rely on PDFs?
Many applications come through brokers, who collect bank statements and send them to funders. Digital aggregators cover commercial bank accounts less completely than consumer accounts, and many owners won’t connect digitally until they know they’re approved. Ocrolus processes about 1.25 million pages of PDFs a day as a result, in addition to digital data.
Why does cash flow underwriting need industry context?
The same transaction can mean different things in different industries. An insurance credit is a claim payment for most businesses but revenue for a dentist or auto body shop. Peer comparison adds more context, showing where a business ranks on revenue or payroll against similar businesses in the same industry and geography.
How will AI agents change small business lending?
AI models pre-trained on massive data sets can interpret the small, fragmented data that traditional machine learning struggles with. Ocrolus SMB general manager David Snitkof expects agents working for business owners to seek the right credit at the right time, and agents working for lenders across prospecting, underwriting, fraud, servicing and renewals.
He says the industry is still very early in this cycle.
