The first wave of fintech in small business lending brought real innovation. Not only could we apply for loans online for the first time, but the approval decisions were very quick, sometimes even instant. By the mid-2010s, companies like Kabbage, Square, and OnDeck had broken new ground, and small businesses had access to capital like never before.
Since then, though, we have had a decade of incremental progress. But now the tools seem to be in place for new breakthroughs. The convergence of artificial intelligence, embedded finance, and better data infrastructure is creating unprecedented opportunities to serve America’s 30+ million small businesses more effectively than ever before.
The numbers tell a compelling story of progress masked by persistent challenges. While fintechs now originate 10-15% of small business lending by dollar volume, they’re serving roughly twice as many businesses as traditional banks, according to Prashant Fuloria, CEO of Fundbox. “If you look at last year, there are roughly about a million small businesses that took a loan from a bank,” Fuloria explains. “The number for fintechs may be about 2x that… fintechs are averaging $10,000-20,000 average ticket size, serving a different kind of customer.”
Yet the scale of unaddressed need remains staggering. Rohit Arora, CEO of Biz2Credit and Biz2X, points to recent research showing “the small business lending gap just in the US is $750 billion. And that is growing every year right now.”
The Complexity Challenge
This gap exists because traditional underwriting approaches struggle with the complexity and heterogeneity of small businesses. David Snitkof, GM of Small Business at Ocrolus, frames it this way: “Small business lending has always been a decade behind consumer lending. In terms of technology and speed…every small business is a super unique entity with very distinct cash flow patterns. There are also many more types of financial products for small businesses with more complicated structures.”
Traditional machine learning approaches, which excel with large, clean, standardized datasets, struggle with the sparse, messy, heterogeneous data that characterizes small business lending. A million-dollar revenue florist operates fundamentally differently from a million-dollar software consultancy or manufacturing business.
The Evolution Towards Agentic AI
The transformation in what’s technologically possible becomes clear when you look at how the capabilities have evolved. Satish Palvai, CEO of Aion, has been building AI-powered lending systems since 2017, giving him a unique perspective on the progress. “Our early innovation centered on building advanced scoring systems powered by machine learning and natural language processing models trained on structured data. Both Microsoft and Google were trying to do certain things around that NLP toolset,” Palvai recalls.
Back then, Aion focused on small loans under $25,000 using basic cash flow underwriting. Today, the company handles million-dollar asset-based facilities. “We’re now harnessing large language models to digest unstructured data in real-time, a capability that’s transformed our speed-to-capital and opened entirely new opportunities for growth,” Palvai explains.
Max Eber, CPTO and Co-Founder of Taktile, explains that the capabilities of the AI models are advancing rapidly. “The frontier of what is possible in small business lending has changed dramatically in just the last 12 months.”
The new generation of AI systems, specifically agentic AI, can handle complex, nuanced decisions while automating routine work. Rohit Arora has been investing heavily in this transformation through a partnership with Columbia University. “I think the Holy Grail is putting agentic AI to work for small business lending.”
The Biz2Credit platform is implementing agentic AI across the entire lending lifecycle, from customer acquisition and application processing to underwriting, pricing, and portfolio monitoring. “We get close to 20, 25,000 applications a month right now. So as the models get trained, as they get better and smarter, we can then offer it to our partners also,” Arora notes. The system launches for Biz2X customers in late October.
Taktile has built an entire platform around this vision. Eber reports that their AI agents can “extract key data points from uploaded documents, turning unstructured information into structured data ready for your underwriting logic. Meanwhile, another agent goes beyond the submission and searches the entire open web, confirming the website is real, the business model matches the application, and surfacing any red flags you should know about.”
The improvements are dramatic. Eber says that customers are seeing approximately “5x improvement” in underwriting efficiency. “If you focus the humans on the 20% most difficult tasks, you can then automate the routine 80% of underwriting.”
The Power of Small Data
A critical innovation is how modern AI systems unlock what Snitkof calls “the power of small data.” Unlike traditional machine learning, which requires massive datasets, “the great thing about generative AI is if you do train a huge LLM, as the foundation labs have done, on massive, massive data sets, then a lot of that intelligence is coded into the base model. And now you can take that model and fine-tune it based on a relatively small number of observations.”
This matters enormously for small business lending. Instead of requiring millions of similar loan applications to train effective models, AI systems can now learn from hundreds or thousands of well-curated examples and apply that learning across diverse business types.
Fuloria at Fundbox has seen this firsthand. “GenAI is really good at making sense of descriptors that are typically unstructured,” he explains. “Our ability to categorize transactions appropriately into whatever bucket: is this a payroll transaction, is this a transaction that is paying a vendor, is it for buying supplies, is it paying off a loan? This has improved dramatically with the application of large language models.” The results: their AI-enhanced cash flow models now “slope risk three times better than FICO.”
For asset-based lenders like Aion, the challenges are even more complex. Palvai describes correlating master services agreements, purchase orders, and invoices, all highly unstructured documents. “Today we use AI and purpose-built tools to correlate all these sources seamlessly,” including comparing invoices against historical payment patterns in bank accounts and accounting systems.
Platform Approaches and New Data
Alex McLeod, CEO of Parlay, represents a platform approach focused on “aggregating and orchestrating data to make decisions with, creating net new data that does not exist, and then learning from how lenders structure deals inside the system.”
This “net new data” is the institutional knowledge of how experienced underwriters actually structure deals. “How did the lender ultimately structure the deal? What was that universe of small businesses that came to the bank? And which ones got through and why. We haven’t recorded any of that data at every single bank in the US,” McLeod explains. Her company reports “10 times more loan throughput, 95% less manual work” for platform customers.
Palvai’s team is taking this further by building vertical-specific intelligence. “We’re developing industry-specific AI agents that understand the unique characteristics of each vertical, asking a manufacturer different questions than we’d ask a CPG company or a software firm”.
Empowering Human Underwriters
While much focus is on automation, AI is also giving human underwriters dramatically more powerful tools. Palvai emphasizes this: “What matters most is giving underwriters the tools they need to be more efficient. Today they’re armed with substantially better data and information, and they’re able to provide answers much faster.”
This hybrid approach, AI handling routine analysis while humans make final decisions on complex cases, appears to be the prevailing model. McLeod cautions: “I don’t think just making everything agent-based is a solution when we don’t have enough of the data to automate the decisions that we’re trying to make…if we’re lending without the right data to power those models more quickly, then we risk exposing more small businesses to liability they can’t take on.”
The Headwinds
Despite enormous promise, significant challenges remain. Palvai identifies the key constraints: “The security, the fraud, the regulatory work, the compliance aspects, that hasn’t changed, right? I mean, with anything that’s improving, fraudsters are getting better, too.”
Privacy challenges are particularly thorny. Palvai outlines the dilemma of using AI systems that learn from aggregated data while protecting sensitive business information. “The privacy, the security, the fraudster, the compliance, I think every state in the country is looking at it, every company is looking at it. I think those are the things, if any, that could slow us down.”
Looking Forward
Arora expects his system to dramatically improve customer experience: “65% of our app-based applicants come after business hours or on weekends. Imagine agentic AI agents, you know, they work 24 by 7, if they can guide these customers to fill out the application, complete their documentation, and then agentic AI agents are working to do the initial piece of underwriting.”
Snitkof envisions a future where “you can use a combination of human and machine intelligence to actually find the right financial product at the right pricing for the right borrower at the right time from the right lender.”
Fuloria expects significant changes within three years: “I’d be disappointed if two things didn’t happen. One, on the technology front, I think that GenAI will be used to not only improve data collection and categorization, plus recommendations for humans-in-the-loop… but the technology and regulation will be at a place where GenAI will be deployed for end-to-end automated underwriting processes… And number two, I think that embedded lending will evolve beyond simply placing a generic product inside of a platform’s UI to delivering the most relevant product to the SMB based on context.”
The ultimate test will be whether these technological advances can finally close the persistent gap between small business demand for capital and the financial system’s ability to serve that demand profitably and safely. The early signs are promising, from companies scaling from $25,000 loans to million-dollar facilities, from simple cash flow analysis to complex asset-based lending, from manual underwriting to AI-assisted decision-making.
After years of incremental progress, the small business lending industry appears ready for its quantum leap forward. The question is no longer whether AI will transform small business lending, but how quickly lenders can adapt to harness its potential while managing legitimate concerns around security, privacy, and compliance.
