Security Bank & Trust Co. President Andy Schornack entered 2026 trying to find a cost-effective solution that had been eluding him for years. He wanted to streamline the slow, manual credit workflow process used at the $1.3 billion community bank, based in Glencoe, Minnesota. That’s when his spouse told him about a new artificial intelligence tool from Anthropic, called Claude Cowork.
“My wife is in the software business and she’s like, ‘The last 60 days have changed the world in terms of how these things that were taking months to develop are now getting developed in a very short period of time, and you’ve got to check this out,” says Schornack, who also serves as CEO of the bank’s holding company, Flagship Financial Group.
Due largely to mergers and acquisitions, the bank has nearly doubled in asset size since early 2024. That means there are plenty of new small business and agricultural customers looking for loans. While that growth has been welcomed, the bank’s manual credit workflow process wasn’t built to handle the additional volume.
Chelsea Bottelberghe, Security Bank & Trust’s vice president and senior credit analyst, says the old process started with a lender sending her a spreadsheet with the loan term, collateral and any other necessary information. She then reentered all those details into her own spreadsheet before forwarding it on to the next available credit analyst.
“Everything the analyst did was all in [Microsoft Corp.] Word, and then we’d have embedded [Microsoft] Excel tables within the Word documents,” Bottelberghe says. “When they got to a point where they could send a draft to the lender, they would have to manually email them a draft, and then they would have to go back and forth with the lender. And once the credit presentation was good, it would be ready for the loan committee. Then they’d have to notify me. I would have to go back and update my spreadsheet. So, there’s a lot of back and forth of communication.”
An In-House Solution
Schornack first looked to outside vendors to help streamline that tedious process. That resulted in some sticker shock. “You’re talking six figures plus for the high-end, Taj Mahal product,” he says, adding that even lower-level models ran into the high-five figures.
That’s why Schornack was more than willing to listen to his wife’s suggestion. He got together examples of all the materials the bank uses to move a loan through its credit workflow process, and entered it into Claude Cowork. He then asked the program to develop an application that could live in Microsoft SharePoint and integrate with all of the lending data the bank stored there. But instead of building an app, Cowork instead provided him with instructions for building it himself in the low-code Microsoft Power Platform. “I’m like, ‘Yeah, this isn’t going to work for me. I don’t have hours,’” Schornack says.
Just when it seemed like he had hit a dead end, Anthropic in June released Claude Fable 5. The upgraded AI model essentially supercharged Claude Code, its agentic coding tool. Schornack gave Claude Code the same instructions he had given to Cowork. This time, he got what he wanted. The program wrote all the code necessary to build out the app. By late June, the first version of the app was finished, and the bank’s lending team started testing the new credit workflow system.
“Everything is entered right from the get-go – all the loan requests, the details, the borrowers, the collateral,” Bottelberghe says. “They then hit submit in the workflow and it routes automatically to credit when it is ready.”
Lenders can then monitor the progress the credit analyst is making instead of guessing where they are in the process. Once the analyst is ready to create a credit presentation, the app also provides a template for them to use.
Schornack also built a custom AI analyst into the app that can provide feedback on what is being entered. “It’s actually a really good training tool,” he says. “It asks you good questions that you probably should be asking before submitting a deal for approval anyway.”
Since SharePoint also leaves audit trails, the bank’s risk and compliance team can track all of the steps taken to build the app, and then monitor its use as well. Schornack also worked to ensure the bank’s nonpublic information wouldn’t be exposed to Claude. While Claude builds the code for the app, and the app lives in SharePoint, none of the data stored there is shared back to Claude.
Don’t Be Afraid To Play With AI
Security Bank & Trust Co. is now fine tuning the app through a process of trial and error.
“What I’ve been working on with Andy is running the financials and everything through there and making sure that everything is calculating correctly, how we want — that nothing is missing,” Bottelberghe says. “That’s been quite the undertaking.”
Each time a new problem is identified, Schornack instructs Claude Code to make the necessary adjustments. That means there have been many new versions of the app. The bank launched version 80 this week. But loans are now starting to make the full journey through the new workflow.
“From start to finish, all the way through to the loan admin, getting the loan documents and all the checklists and everything, it’s only been a couple that have run all the way through to the very end,” says Bottelberghe. “But we have been starting to use it more in the past couple weeks for credit presentations.”
Agricultural loans, which are far more complex than consumer or small business loans, will likely be last to fully move through the new system. But Schornack says he’s encouraged by where they are after just a few months of testing. He also is happy with the cost — which has only been his time, and the few hundred dollars the bank spends each month to run Claude Fable 5 and a suite of Microsoft 365 products.
“I think this has just proven my point that with a fraction of the cost now, you can build technology that any community bank can use,” Schornack says.
Chris Weidemann, chief AI officer and cofounder of Advisor Labs, works with small financial institutions to help them find simple AI solutions. He says too many C-suite executives are afraid to play with their AI models the way Schornack did. “They don’t know how to do it,” Weidemann says. “They feel like there’s too much in the way.”
That’s why Weidemann advises them to find AI enablers — employees who are interested in learning what AI can do, and then sharing their knowledge with the institution’s various subject matter experts. “I would bet most community banks and credit unions have 12-to-18 months of low-hanging fruit that they could just be picking off before they really need to get deep into building something custom or buying some big tool,” he says. “There are a lot of quick wins out there.”
Schornack says most won’t know what they can do until they try. “A lot of times people are like, ‘Well, I just don’t know how to even talk to it,’” he says. His advice is simple: “Just talk to it.”
Previous