Biz2X, a business lending platform provider, is collaborating with Columbia University students and professors to develop agentic AI models to be used to make loans to small businesses and to help small businesses manage their finances. They're building the models on top of Google's Gemini large language model using cloud infrastructure from Amazon Web Services. Biz2X and Columbia plan to publish a research paper on this work.
During the past few years, Biz2X has hired several data science graduates from Columbia University's engineering and business schools.
"The idea was to take this partnership to the next level and institutionalize it, and also get Fortune 500 companies like Google, Amazon and others to join us," Rohit Arora, CEO of Biz2X, told American Banker.
Jeffrey Sachs, a Columbia University professor, said the technology and research of the combined team would help improve access to finance for small businesses, which "fosters innovation, job creation, and long-term resilience in both high-income and emerging economies."
For the past four months, graduate students and PhD students at the university have been spending time in Biz2X's office, using the company's data, systems and knowledge base.
"They're interacting with our data scientists, with our engineering team, with our underwriters," Arora said. "They're even listening in on conversations with end clients."
The students receive a stipend for their work. Some will be hired at Biz2X when the program concludes at the end of the school year, and others may get jobs at private credit companies, the intended users of some of these models.
"The Biz2X partnership with Columbia is a great way around the lender shortage that stymies both banks and their vendors," David O'Connell, founder of O'Connell Lending Insights, told American Banker. "New graduates, whether they're focused on technology or finance, are just not drawn to traditional lending, be it small business or commercial."
The talent shortage is particularly acute "when it comes to specialists in AI and data scientists, a prerequisite for strong AI outcomes," O'Connell said. "This partnership is now a de-facto talent pipeline for Biz2X and other small business lending vendors in New York City."
The first use case for agentic AI in this initiative is creating AI-driven workflows for small business loan origination, including customer experience, underwriting, closing, funding and portfolio monitoring, Arora said.
When an underwriter looks at a small business, he has to look at its cash flow, its records, its UCC liens and any existing credit facilities it might have, Arora said. An AI agent could go through the whole file and build an analysis around it, and perceive patterns in all the applications, in terms of which ones received funding, pricing, repayment and performance record. The underwriter could use this analysis to have an informed conversation with the borrower or make a decision.
"Today, automation helps us to create the ability to go in and say yes or no based on the rules engine," Arora said. "But this is taking it to the next level, where you're connecting the dots. You are looking at multiple data channels. You are benchmarking it against existing data, creating a synopsis out of it, drafting credit memos and then letting the underwriter work off that."
O'Connell agreed that small business lending could be a good use case for agentic AI.
"I don't care that AI is currently overhyped and enduring a backlash: it provides antidotes to many long-standing problems and productivity sink holes across business lending," he said. "It's particularly antidotal to small business lending, which requires dozens, if not hundreds of processes to be completed, the vast majority of which do not require extensive training or knowledge of business lending. Because small business loans are so small, you're losing money after just several hours of human credit analysis. Scale, speed and throughput from AI are absolutely the solution."
O'Connell also noted that the model developers can't just go "mad scientist" here. "Private debt might not be regulated and will likely remain so under the Trump administration," he said. "But they will have to be super sharp about disparate impact." Also, banks need to know every variable in their data models and their effect on the credit decision in order to ensure that a model isn't treating one class of customers unfairly relative to another, he said.
The second use case is to upgrade Biz2X's Virtual CFO platform, which according to Arora is used by small businesses to manage their cash flows.
"A great example is that if you're a retail business, you need to manage your inventory," he said. "You need to manage your rent cost. You need to manage your employee cost. If you're a B to B business, you need to balance your account receivables against your account payables, your travel costs and such." The agentic AI model will customize the Virtual CFO platform based on each business's needs and imperatives, he said.
O'Connell noted that principals of small businesses tend to be over-extended entrepreneurs who lack the time to be a great CFO or controller.
"So they make lots of mistakes like using payment rails that are too expensive and earning too little on their overnight balances," he said. "AI agents are great for popping up and saying things like 'Are you sure you want to use a wire for this payment? ACH is less expensive?' If a lender fields a sufficiently rich cohort of agents, it could do away with the 'business banker' entirely."
A goal of the work is to provide data about small businesses that private credit firms can use to understand which small businesses are worthy of funding.
"These guys are extremely sophisticated when it comes to understanding the rated instruments, but getting a rating in the SMB space is time consuming and expensive," Arora said. Agentic AI could perform the same analysis in real time, he said.
Another application of agentic AI is going to be portfolio monitoring, he said. For instance, small business loans are directly affected by Fed interest rate cuts which impact cash flow and overall performance, he said.
O'Connell expects to see more collaborations like the one between Biz2X and Columbia.
"The lender shortage is the biggest barrier to growth in business lending; hands down," he said. "Stealing them from rivals is too expensive. I've interviewed lots of lenders. The only truly fruitful best practice is partnerships with schools. You give kids first-hand exposure to business lending without the snark it receives in the broader culture. They learn about core things like risk assessment, deal making, pricing, and relationship nurturing. And they become a captive audience from which to recruit down the road."
