AI Agents Are Here. Is Banking Ready?

Hey everybody. Welcome back to another episode of What’s Going On in Banking. I’m Ron Shevlin, chief research officer at Cornerstone Advisors, and I’m going to bring my co-host Stacey Bryant in quickly because my voice is already going.

Stacey, what’s going on?

I am fresh off New Orleans, or N’awlins.

I mentioned in the previous episode that I was heading there for the New Jersey Bankers Association annual conference. Somehow I had never been to New Orleans before, so of course I had the obligatory gumbo with extra hot sauce and some beignets.

I always bring gym clothes when I travel, and eight times out of 10 I ignore them.

New Orleans was definitely one of those trips. I decided to forget the diet and fitness goals and just jambalaya it up.

But on a more serious note, I kept hearing three major concerns from bankers at the conference.

The first was regulation, especially the fear of liability as banks try to introduce new technology and innovation.

The second was mergers and acquisitions. Who is actually going to survive and thrive as consolidation continues?

And third was the question we hear constantly: okay, we introduced AI, now what?

Every time I heard that, I asked the same follow-up question: what are you trying to solve?

That question applies to generative AI, agentic AI, or any other new technology. What is the use case? What is the actual business problem?

I did notice that bankers are becoming more curious and more willing to learn from each other. They want to know what other institutions are trying, what is working, and what has failed.

That is always one of my biggest takeaways when I’m on the road.

How are you doing, Ron?

Doing great.

And I have one completely unrelated question. Does anybody else find it strange that state banking and credit-union associations sometimes hold their annual conferences in completely different states?

If you are the New Jersey Bankers Association, shouldn’t you be supporting New Jersey and having the conference in Atlantic City or somewhere local?

I guess I just ruined my chances of getting invited if they decide to have the next one at The Breakers in Florida.

It’s okay. You can go to Atlantic City anytime.

I have been there. No comment.

We’re not going to badmouth anybody’s location today.

Everything is good here in Boston too.

People outside New England may not realize we do not really have four seasons.

We have winter, summer, fall, and then one weekend we call spring.

We just had spring. It was beautiful and in the 70s. Now we’re heading straight into 90-degree weather.

Let’s get into the banking topics.

You wanted to start with my recent Substack piece about Fiserv’s Agent OS.

Yes.

For everybody who is not following Ron’s Substack, get with the program.

The article was titled “The AI Agent Race in Banking: Fiserv Steps Up to the Starting Line with Agent OS.”

What grabbed me was the use of the term operating system.

An operating system sounds universal, neutral, and platform-agnostic.

But Fiserv’s Agent OS is designed to operate primarily across Fiserv’s own platform and infrastructure.

If your bank runs Fiserv core and payment technology, that can be very valuable because it extends the investment you already have.

If you do not run Fiserv, calling it an operating system starts to feel different.

I actually went to Perplexity and asked it to define an operating system, then asked what a community bank’s operating system would really consist of.

That helped me distinguish between a true operating system and an orchestration layer that coordinates agents inside a particular ecosystem.

Your article also pointed out that Fiserv is not the only company moving in this direction.

FIS has announced work with Anthropic. nCino has been doing agent-related work. Jack Henry will likely have its own approach. And there are specialty fintechs focused on narrower banking use cases.

That raises the strategic question we always come back to: who are you serving, what problem are you solving, and what does your future technology stack need to support?

I’ll stop there and let you react.

I have nothing to say.

Only kidding.

First, I should disclose that as we’re recording this, I have a LinkedIn message from a senior Fiserv executive saying, “Let’s talk about this.”

Thankfully, I do not think the article was particularly negative toward Fiserv.

I did challenge the “operating system” terminology, but overall I think the announcement is important and positive.

If you are a Fiserv client, you should absolutely be talking with Fiserv about this.

And we need to stay precise about definitions.

When I say AI agent, I mean a tool that can perceive information, make decisions, and take actions toward a goal with some degree of autonomy.

I am not talking about a virtual call-center agent or a chatbot that simply has a conversation with a customer.

An AI agent is performing work.

The reason the Fiserv and FIS announcements matter is that banks have been asking how third-party agents will fit into their existing technology stack.

I have spoken with a lot of bank and credit-union boards over the last year and a half, and one of the questions I keep raising is whether they are willing to integrate agents from outside vendors into their core systems, lending platforms, payment platforms, and other applications.

The fact that major providers such as Fiserv and FIS are now building orchestration and governance capabilities should make clients more comfortable that agents can be integrated in a controlled way.

That is important.

Where I pushed back was the marketing language.

When I look at what Fiserv described, it sounds more like a governance and orchestration layer than a traditional operating system.

That does not make it less valuable.

In fact, orchestration is exactly what the industry needs.

But banking already has a terminology problem around generative AI, agentic AI, assistants, agents, copilots, and everything else.

When a major provider uses a grand term such as Agent OS, the marketing can create even more confusion.

I am sure I’ll get grief for saying that.

One part of the Fiserv announcement I did like was the focus of its first agents.

Fiserv highlighted commercial-loan onboarding and anti-money-laundering work.

Those are strong early use cases because they involve a lot of manual information gathering and repetitive process work.

Agents can potentially reduce cycle time and improve productivity dramatically.

Let’s stay on commercial lending for a minute.

In the article, there was an example of daily operational reporting dropping from about 10 minutes to seconds.

That is a real improvement.

And when we think about why a financial institution would use agentic AI, it comes back to the same question: what are you trying to solve?

I am still giving presentations in May 2026 where one of my first slides breaks the broad term “AI” into categories.

People are still asking for that because the definitions matter.

From a community-bank perspective, I am actually optimistic about the competitive landscape.

Cornerstone research found that roughly a third of banks expect to rely on specialty fintech providers for some of these capabilities.

That means banks are not limited to whatever one core provider chooses to build.

Fiserv, FIS, nCino, and eventually Jack Henry may offer broad platforms, while specialty companies can address specific pain points.

The important thing is that institutions understand what they are buying and how it fits into the architecture.

Exactly.

And there is another data point from Cornerstone’s recent What’s Going On in Commercial Lending report that I think is instructive.

We asked bankers how much impact they expect AI to have across different parts of the commercial-lending process.

We broke it into lead sourcing, relationship management, application and proposal work, underwriting and approval, due diligence and closing, post-closing, and portfolio management.

Then we asked whether AI would have a strong impact, moderate impact, minimal impact, or whether they were simply not sure.

A strong impact meant more than a 20% reduction in cost or a significant reduction in cycle time.

Moderate meant roughly a 5% to 20% improvement.

Minimal meant less than 5%.

Across most subprocesses, about one in five respondents expected a strong impact.

Lead sourcing and relationship management were lower because bankers still see those areas as heavily human-driven.

Another 30% to 40% expected a moderate impact.

Then you had a meaningful group expecting only minimal impact, and about one in five in many categories simply said they were not sure.

The “not sure” group worries me most.

It is fine to be skeptical. It is fine to believe AI will have only a small effect.

You may be right or wrong, but at least you have formed an opinion based on some understanding of the technology and process.

If you are still saying “not sure” in 2026, that suggests an education gap.

That is a perfect segue into another article I read in The Financial Brand: “For Small Institutions on the AI Journey, Here’s How to Bring Your Skeptics Along.”

The biggest AI risk in many institutions is not the technology itself. It is the culture in the room.

I see this every time I’m on the road.

People talk about AI as if a robot is going to walk into the bank, fire everyone, and start approving loans.

The real divide is not simply believers versus laggards.

It is often between people who think AI is going to erase the value of their expertise and people who think AI can upgrade the tools they use.

One person says, “I spent 20 years learning these customers, these loans, and these quirks, and now a bot is going to do my job in 30 seconds?”

Another says, “You mean I can stop copying and pasting data into Excel and let AI handle the boring 80% so I can actually think for a living? Where do I sign?”

I think the strongest framing is a dual workforce.

Machines handle more volume and repetitive work. Humans retain judgment, empathy, and accountability where those things matter.

But there is another tension too.

One group fears becoming irrelevant if the institution moves too slowly. Another fears being blamed when the technology makes a mistake.

One side fears being left behind. The other fears a consent order.

That is where the cultural challenge sits.

I agree with the tension, but I worry about framing it as a binary choice.

I have a board presentation coming up where I’m using what I call an AI capability continuum.

On one end is human-directed work. On the other end is AI-directed work.

There are several stages in between.

At the most basic level, AI can be a tool. It helps you draft an email, summarize a document, or analyze information.

Then it can become an assistant. You give it more responsibility, but you are still directing most of the work.

Then it can become an operator. Under predefined conditions, it performs specific tasks automatically.

Then a collaborator. It performs substantial portions of a process and hands key decisions or exceptions back to a human.

And finally, at the far end, it can become an autonomous actor that executes a process with limited human involvement.

The FIS and Anthropic work in AML is a good example of the collaborator stage.

The agent may gather evidence and assemble information, but the human investigator still owns the final judgment.

There are also situations where I personally want full autonomy.

I hate calling a bank or vendor to cancel something or resolve a minor issue.

I would love to tell an agent, “Go cancel this subscription. Go resolve this payment problem. Let me know when it is done.”

So we should not reduce this to “AI replaces people” or “AI only assists people.”

There is a continuum, and different processes belong in different places on that continuum.

In three to five years, the lines will probably blur enough that the framework itself becomes less useful.

Today, though, it helps boards and executive teams decide where autonomy makes sense.

I think that also helps with the workforce conversation.

When someone starts as an intern, they often begin as an assistant. Then they become an operator. As they gain experience, they become a collaborator and eventually make more decisions independently.

There is a similar progression in how institutions can think about agents.

But none of that works without good data.

Quantitative and qualitative data across the institution become the raw material that lets the human workforce and the agentic workforce work together.

And again, when an institution is already struggling through a core conversion or major implementation, adding this conversation can feel overwhelming.

That does not make it optional.

Nobody said being a senior executive was supposed to be easy.

Executives are not paid the big bucks because they have 30 or 40 years of experience and can now sit back and coast.

They are paid to make hard decisions about where resources go and how work changes.

I think this has strong parallels to the arrival of personal computers 40 years ago.

Consultants used to tell companies to prioritize which departments should get PCs first.

Within a few years, the answer was everybody.

Then, 10 or 15 years later, business-process reengineering took off because organizations realized new technology allowed them to redesign the work itself.

We are back in a similar moment.

AI is going to affect virtually every job and department to some degree.

The management challenge is deciding where employees can use tools independently, where the organization needs standardization, and where entire processes should be redesigned.

I keep reducing that to two questions for boards.

How does AI change how work gets done?

And how does AI change who does the work?

That is the heart of it.

The opportunity is not only reducing cost. It is dramatically improving speed and potentially decision quality.

Humans are not perfect either. We are slow and we make bad decisions.

People love pointing to AI mistakes while forgetting how frequently humans make mistakes themselves.

That brings me to another story that involves AI and personal finance: OpenAI and Plaid.

What is happening there? And why am I apparently paying more money for another subscription?

OpenAI rolled out a preview of a new personal-finance experience inside ChatGPT for U.S. Pro subscribers.

This is not the $20-a-month plan. The Pro plan is closer to $100 a month.

The experience is powered by Plaid and lets users connect bank accounts, credit cards, investments, and other financial accounts from thousands of institutions.

The idea is that ChatGPT can give more context-aware financial guidance.

You could ask questions such as, “Why does cash feel tighter this month?” or “Help me build a plan to buy a house.”

Plaid’s CEO called it a step toward greater financial freedom for consumers.

Surprise, surprise, I am skeptical.

One useful piece of context is that OpenAI has made a couple of acquisitions in the personal-finance-management space over the last six months.

One was Roi, and another was Hiro Finance, which was founded by the same entrepreneur who started Digit.

I was a big fan of Digit because it did something, not simply advised you.

It analyzed cash flow, identified money that could safely be moved, and automatically transferred it into savings.

My biggest issue with the OpenAI-Plaid experience is that, at least initially, it is mostly advice.

Plaid and OpenAI have cited a huge number of people who ask ChatGPT financial questions every month, something around 200 million globally.

I do not dispute that people ask a lot of financial questions.

But how many of those are serious planning questions versus “How do I buy Bitcoin?” or “What meme stock should I buy?”

The second issue is price.

If you need the $100-a-month Pro plan to get this capability, that is roughly $1,200 a year.

Most consumers do not pay anything close to that for financial guidance.

And I am not convinced the advice will be dramatically better simply because ChatGPT can see a handful of accounts.

People’s financial lives are more complex than account balances and transactions.

You can ask someone about income, savings goals, debt, and expected changes without necessarily needing direct account access.

If OpenAI eventually recreates something like Digit and becomes the arms and legs of your financial life, moving money and executing decisions, then I get much more interested.

Advice alone does not justify the price to me.

Are you trying to put us out of business, Ron?

I have some questions too.

The 200 million figure has to be global, because there simply are not 200 million U.S. adults using ChatGPT for financial questions.

But I do think we should take the underlying behavior seriously.

I recently spoke with a small-business owner who is using ChatGPT for marketing, posters, community events, and other tasks.

You can see AI-generated content throughout the business.

That made me think about what banks are doing for their small-business customers.

If the small businesses themselves are increasingly relying on AI to run the company, should the bank be helping them understand how to use these tools responsibly?

If the customer’s business becomes stronger, the bank benefits too.

And when people ask ChatGPT questions about buying a house, starting to invest, or managing Bitcoin, I still count those as meaningful financial interactions.

So even if I would not personally pay $100 a month for the product, I understand why OpenAI wants to connect more deeply to financial behavior.

The Roi acquisition was especially interesting because OpenAI seemed to want the people and the behavioral knowledge more than a mature standalone product.

That suggests they are studying how people actually manage money, not only how to answer questions.

I think that is the right way to look at it strategically.

The personal-finance feature is probably one more reason to keep people inside the ChatGPT ecosystem.

Think about Amazon Prime. The value is not one feature. It is the accumulation of benefits that makes the subscription harder to cancel.

OpenAI needs the same thing. It needs people to keep finding new reasons to use ChatGPT instead of Claude, Gemini, Perplexity, or something else.

I still think this specific feature is overpriced if you look at it only as financial advice.

But if OpenAI eventually combines advice with execution, then the value proposition changes.

So what I hear is a two-part future.

First, the system understands the customer’s finances and provides guidance.

Second, it actually executes, moves the money, changes the subscription, reallocates savings, or completes whatever action was recommended.

At that point, we can quantify the value much more clearly.

Exactly.

And I think that is a good place to leave it for this episode.

Stacey, thanks as always.

And thanks to everybody listening. We hope you’ll join us for another episode of What’s Going On in Banking.

If you enjoyed today’s episode, follow us on Spotify, YouTube, Apple Podcasts, or wherever you listen.

Thanks for tuning in.