Where AI Might Help in Banking

I've spent the last two months on personal time exploring ChatGPT, digging into how large language models work, and running it through every banking scenario I can think of. The wonder phase is fading. What's replacing it is a more practical question. Where could this eventually help at a community bank, and where would it get a bank in trouble?

The exploration is purely on personal projects. There is no policy framework yet that I would consider sufficient for production use in banking, and until there is, this is a thinking exercise rather than a deployment plan. It is a useful thinking exercise, though, because the technology is moving faster than the operational and regulatory infrastructure, and being ready to engage thoughtfully when the time comes is part of the work.

Where I'd Start

Banking runs on text. Loan files, compliance documents, client correspondence, policy manuals, regulatory guidance. There is significant time spent across the industry handling all of it manually. Reading, summarizing, routing, responding. That is exactly the kind of work large language models are built for.

Client service is the first place my mind goes. Banks get the same questions constantly. Balance inquiries, rate questions, how to set up a wire, what documents are needed for a loan. A model that can handle those conversations naturally, instead of the clunky chatbot patterns that frustrate everyone, could free up staff to work on the interactions that actually require a human.

Compliance research is another one. Regulators publish guidance, and someone has to read it, figure out what applies, and translate it into internal procedures. That translation work is time-consuming and repetitive. Once policy permits, a model that can summarize regulatory documents and flag what is relevant to a community bank could save hours.

Then there is the internal productivity angle. Drafting emails, summarizing notes, writing first versions of documents. From personal exploration, the time savings on these kinds of tasks are real. The model handles the blank-page problem so that the work shifts to editing rather than generating from scratch.

Where I'd Wait

The places where AI looks most promising in banking are also the places where the cost of getting it wrong is highest.

Client-facing interactions without oversight worry me. ChatGPT hallucinates. It states wrong things with the same confident tone it uses when it is right. In a conversation about balance transfers or loan terms, a confident wrong answer could create real liability. Until there is a reliable way to constrain what the model says to verified information, a human stays between the AI and the client.

Regulatory reporting is another boundary I would not cross. Call reports, BSA filings, HMDA data. These have zero tolerance for errors. The consequences of filing incorrect regulatory data are serious enough that AI does not belong near the final output, at least not without multiple layers of human validation that would eliminate most of the time savings anyway.

Credit decisions fall into the same category. The fair lending implications of using a model that is not fully understood to approve or deny loans are obvious, and regulators are doing careful work in this space. Waiting for that guidance to develop is the right posture. Being the test case is not.

Deciding What's Worth Trying

When I think about which use cases would be worth piloting, three questions come up that help separate the candidates.

First, what is the cost of an error? If a model summarizes a policy document incorrectly and someone catches it during review, the cost is a few minutes of correction. If it gives a client wrong information about their account, the cost could be regulatory and reputational. The lower the error cost, the better the fit.

Second, is human oversight practical? For internal tools where an employee reviews the output before acting on it, oversight is natural. For real-time client interactions at scale, meaningful oversight is much harder. Use cases where someone is already in the loop are better starting points.

Third, does it give people their time back? The strongest applications are the ones where AI handles the volume and repetition so that people can spend more of their day on judgment, relationships, and the work that genuinely requires a person. That is where the leverage is. The goal is a better day for the staff and a better experience for the client, not technology for its own sake.

What's Next

The pattern that makes sense to me, once policy and governance allow, is to start small. Internal tools first, client-facing later. Low-stakes use cases first, learning what works before touching anything regulatory or client-facing. The technology is impressive, but impressive and production-ready are different things in a regulated institution. The operational and regulatory infrastructure has to mature before any of this becomes real, and that maturation is happening on the right pace for a regulated industry.

The potential is real though. If the approach is right and the policy work develops alongside the technology, AI could help small institutions punch well above their weight when the time comes.