As banks test the many possible use cases for generative AI — deep-learning models that can analyze large volumes of data and generate statistically probable outputs when prompted — one area of interest is small-business lending. A generative-AI-based virtual assistant or copilot can help shepherd small-business owners through the time-consuming and sometimes perplexing loan application process and take care of much of the work loan officers typically do, such as data validation, qualification and loan approval.
One bank trying this out is
In the six months Ryan Hildebrand has been chief innovation officer at the community bank, he has been looking for ways to make it more efficient and innovative. In one example, he met the team of a startup called Cascading AI through an investor and thought the generative AI technology they had built could be applied to make small-business lending more efficient.
"I am a huge, huge believer in small business and banking," Hildebrand said in an interview. "But it's always been done in a way that just doesn't make sense for a bank."
Lending to small businesses is labor intensive, which is why some banks don't do it. Small-business loans are sometimes more complicated and usually less profitable than loans made to large commercial clients.
"Small businesses are less organized than larger ones, so leading them through the application process just takes a lot of time, a lot of handholding, writing emails back and forth to remind someone that you still need that 2021 tax return," said Lukas Haffer, CEO and founder of Cascading AI, in an interview.
Banks have been testing generative AI cautiously, well aware of the dangers of hallucination, wrong answers and plagiarism. In a not-yet-released American Banker survey of 127 bank executives, 61% of community bankers said they are still learning and collecting information about generative AI. Among all respondents, 80% were concerned that generative AI could create nonsensical or inaccurate information.
Bankwell Bank began piloting Cascading AI's Casca software to prequalify and make loans to potential small-business borrowers.
"We've just seen tremendous success with it, even in the first months of this, versus the prior four months," Hildebrand said. He declined to share specific numbers but said Casca helps bring leads that were "five or six times the quality" of what was coming in through organic marketing.
The bank recently launched a new Small Business Administration 7(a) lending program to issue more small loans to small businesses that were being offered high interest rates from online lenders. The SBA guarantees a portion of these loans, which enables the bank to offer a lower interest rate because it's not bearing all of the risk in case the loan defaults.
But the process of applying for these loans can be confusing and lengthy, Haffer said. "And when you don't have everything together, it takes a lot of work for the bankers," he said.
Small-business owners are always busy, and can rarely find time to talk to their banker. They sometimes have employees in the room with them, so they don't necessarily want to talk about their financial situation and why they want a loan, Hildebrand noted.
Small businesses can have random questions, too, such as, what is an EIN? (It's an employer identification number, akin to a business taxpayer identification number.)
In the pilot of Casca, a potential small-business borrower comes to Bankwell Bank's website and fills out a form, providing basic information to see if they prequalify for a loan. Several tasks need to happen to determine if the business is truly qualified for a loan and eligible for the SBA program.
Cascading AI's virtual assistant, which is named Sarah, works through those tasks, asking the customer follow-up questions like, "How many years has the company been in business?" and "Tell me why you need a loan."
A human lender would have to call the customer and flesh it out more. Casca can do this work at any time.
"We're able to do it through AI at 11:30 on a Friday night, which — go figure — is really the time that we see the most uptick," while loan officers are off duty, Hildebrand said.
The generative AI model analyzes the applications to determine which loans can be approved, but a human reviews each case.
"There's always a human in the loop," Hildebrand said. "We're not completely doing away with that. From a risk standpoint, there's all sorts of issues that obviously get brought up. But, there's always someone there reviewing all the conversations."
This helps alleviate the risk of
"At scale, I think that is the issue," Hildebrand said. "I just don't see us not having every single thing reviewed by a human."
Having the AI start the conversation is a big help, he said.
"The lender, in the past, would call them on Monday morning and still not get an answer," Hildebrand said. "Weeks would go by just to figure out if the borrower was going to be eligible or not. So the first part on the prequalification side is huge for us."
Sarah has been trained on SBA policies and the bank's lending policies.
"We're not fully ready to say, okay, Sarah, you're the true SBA lending expert at this point," Hildebrand said. "But I think the idea at some point is really being able to think through how to have a copilot for these things. I think that's a huge opportunity to save time."
Mike Abbott, global banking lead at Accenture, shared where he sees banks using the technology most effectively in the year ahead.
When customers get confused or realize they don't have some information they need, they tend to close out of an online loan application, Haffer said.
At Bankwell, "within a couple of minutes, they get a message from the infinitely friendly, infinitely kind, infinitely respectful Sarah, that has infinite time for them and says, Hey, I saw that you started an application for a $75,000 working capital loan on our website. I'd love to learn a little bit more about your company. Can you tell me what you're looking for funding for? And here's a link to go back into the application." Sarah will note where the potential borrower left off and offer to help with that field.
"The first time we sent the first message, about five weeks ago, everything changed," Haffer said. "The first person immediately responded to the message, said, Hey, I got confused. I didn't know what an EIN was. Sarah responded back, well, it is often called taxpayer identification number. You can find it in this letter." The borrower went back into the application and submitted the correct data and documents.
Sarah keeps borrowers informed about the status of their loan application and lets them know of any additional information or documents they need to provide.
"So this person who otherwise would've been lost and no one would've had the time to follow back up with them, now is getting this infinite amount of time and care and attention," Haffer said. "The secret behind all of it is people actually do like talking to someone and people actually respond to an email from Sarah."
Sarah identifies herself as an AI assistant. "But people don't care," Haffer said. "People just have someone reaching out, giving them the information they need, giving them the link to jump back into the application. And they do."
Haffer estimates Cascading AI automates about 90% of the loan process. Where a loan officer might normally spend a lot of the day chasing after applicants to upload the right documents, now they can just review the work and approve loans. The San Francisco startup raised $3.9 million this week, in a pre-seed funding round led by Peterson Ventures, with participation from Y Combinator, The Sarah Smith Fund and Clocktower Technology Ventures.
Behind the scenes, several large language models were used to create Casca. Haffer declined to share details about this.
"All of it has gone through the bank's compliance and security procedures to make sure that we're not sharing sensitive data with anyone that isn't supposed to be privy to that information," he said.
To provide explainability (and prove that Casca is not the kind of
"We have access to all the actions that an applicant took — when did they submit the application, when did they agree to the terms and conditions, etc.," Haffer said. "We have access to all the information about what loan officers and users of the system did, like which loan officer approved this loan, put it into a next stage, decided to collect these additional documents." There is a record of every prompt the model received and its responses.
Bankwell is Cascading AI's first client; Haffer said he is in talks with a few other banks.
"Our mission for the last four months was to build a system that actually brings AI into reality, is useful and fully compliant for one client and solves problems for them," Haffer said. Now the company is focused on scaling up the Bankwell installation and onboarding two more banks.
