At Grasshopper Bank, a digital bank based in New York City, Chief Technology Officer Pete Chapman started noticing that the bank's startup and tech company clients were using large language model chatbots like Anthropic's Claude, Google's Gemini and OpenAI's ChatGPT for almost everything.
"One of our beta clients uses Claude to help build her end-of-month financial reporting for her board every single month, and she's had to extract her bank account data and download it into an Excel spreadsheet, then plug it manually into that board packet," Chapman told American Banker.
This week, Grasshopper announced that it's letting such customers load their bank account data directly into Claude, using a feed built by its tech partner Narmi.
Now the bank's beta user, who previously relied on manual processes, feeds her bank data directly into Claude to create monthly financial reports.
"Ultimately, we're trying to listen to our clients," Chapman said. "We know our clients do things elsewhere, and we'd rather enable them in a safe and secure manner. This is just the next step in that evolution."
This is part of a broader AI adoption journey at Grasshopper that until now has focused on using AI to drive efficiencies in the company's back office.
"We spent the better part of a year and a half undergoing that effort, and we've made a lot of progress on that, but we always said to ourselves, at some point, we would like to take Gen AI and be able to deliver a better client experience," Chapman said.
No other U.S. bank has publicly announced that it's letting customers access their bank data through a large language model. While many banks have enthusiastically embraced generative AI for the sake of internal efficiency, most have hesitated to let customers interact with their bank that way – the risks of error, hallucination or bias are still considered too great.
Grasshopper is not using Anthropic's Claude as a customer service chatbot to answer questions about the bank or as a replacement for call-center representatives. It's giving business customers the option to connect their bank account information to the Claude chatbot and query and analyze their own financial data by asking questions, similar to the way they might bring their bank account data into QuickBooks or a personal financial management app.
Customers give their explicit consent before any data is shared with an AI model. They can use Claude to query any data they would normally see in their digital and mobile banking channels.
"This puts [users] in the driver's seat in terms of their data security and privacy," Chris Griffin, co-founder and co-president of Narmi, told American Banker. "They can select the providers that they feel comfortable with, and they're not required to use this by any means."
Chapman compares the Claude initiative to sharing data with fintechs in order to enable bank customers to use fintech apps.
"There's something larger happening in the industry as you think about ownership of data," Chapman said. "Some banks are building walls. Some banks are tearing down walls when it comes to ownership of that data. We've always thought here at Grasshopper that it's the client's data, and we're just stewards of that data. Strategically we want to differentiate ourselves from some other business banks by enabling our companies to access third-party accounting platforms or fintech apps in a safe and secure manner."
Grasshopper's initiative may be a sign of things to come.
"Over the next five years, I'm bullish on natural language interfaces as a powerful, differentiating customer engagement model," Alenka Grealish, lead emerging technology analyst at Celent, told American Banker. "The most advanced version of this model is tapping large or small language models to respond to customer queries about their financial data, essentially giving them a financial analyst. When I telescope out beyond five years, I envision banks 'leasing' virtual CFOs, comptrollers and accounts payable and receivable agents to the small to mid-sized business clients."
Grasshopper's decision to start with Claude wasn't due to any special capability of Claude itself. Rather, Grasshopper and Narmi decided they wanted to use the model context protocol standard (which was developed by Anthropic but published as open source code) to provide the data, and Claude is the first LLM bot to meet the standard.
The model context protocol is intended to enable developers to build secure, two-way connections between their data sources and AI-powered tools. But when Anthropic published it in 2024, it was missing security components such as authentication and authorization mechanisms, Griffin said.
Narmi set out to build these missing pieces. It repurposed the authentication and authorization mechanisms it already provides for its digital banking software, including two-factor authentication, to work with the model context protocol. It added Transport Layer Security version 1.2 to safeguard communication between the bank's servers and Claude. It provides an application firewall and ensures that Claude's data permissions are read only. It also provides a means of obtaining end-user consent.
"Generative AI has been making news headlines over the last few years, but one thing that struck us late last year was a study from Experian that found that two-thirds of Gen Z consumers were using artificial intelligence for personal financial management," Griffin said.
Over time, Chapman anticipates giving customers access to data through OpenAI's ChatGPT, Google's Gemini and other large language models, once they support model context protocol. Grasshopper's customers use all of these tools.
"We're a Google shop, and we are deeply into Gemini in our back office, and we're having conversations with Google about how they could support this a little bit more quickly," Chapman said. "Our hope is that all of the big chatbot providers end up standardizing around the MCP format."
Claude, like other foundation models, has been trained on almost everything on the internet.
"That's the beauty of this thing," Chapman said. "Our clients can ask Claude something about their finances that requires their data, but then Claude puts it in the context of potentially something larger, like maybe a looming rate cut or something broader that's happening in the economy."
The flip side of large language models consuming the entire internet is that there's a lot of crazy on the internet.
Grasshopper addresses this with an end-user agreement that spells out that what happens in Claude is not the bank's responsibility.
"We are provisioning this data, it's read-only, you have requested this," Chapman said. "But then what you go and do within Claude is your responsibility. That's one of the things we're going to learn a lot about in the coming days, weeks and months, as we listen to client feedback."
Customers have been beta testing the functionality. One customer wanted to understand how much money they would need in order to go on a specific vacation next spring. Claude obtained the travel cost information for that destination and was able to estimate how much money they would need to save. Then the customer asked Claude how much they actually have saved, how that has been trending, and what they need to change, if anything, in order to have enough money in time. Another use case for a small business owner is to ask Claude what revenue concentration risk looks like, and to get guidance on what to do about that.
"That's the power of this," Griffin said. "You can get really inventive in terms of what data sources, what information you're bringing into this context for you to get really hyper-personalized financial information catered to your own specific bank context."
Grasshopper is offering the service to its business clients first. (The bank does have consumer customers through a partnership with AAA.) It's been offering the Claude integration to clients in controlled beta settings and expects to make it generally available within the next few months.
