The banks agents will choose

There are approximately 4,200 banks in the United States. You’re likely familiar with the largest ones — JPMorgan Chase, Bank of America, Citi, and a handful of others. And for good reason. Just those three institutions alone hold a combined $11.5 trillion in assets.

Broadly, you can think about the banking system in a few categories: community banks with under $10 billion in assets, regional banks between $10 billion and $100 billion, large banks with over $100 billion, and a small number of institutions designated as G-SIBs, or Global Systemically Important Banks.

There are eight U.S. G-SIBs: JPMorgan Chase, Bank of America, Citi, Wells Fargo, Goldman Sachs, Morgan Stanley, BNY Mellon, and State Street. These are institutions whose size, interconnectedness, and role in the global economy make them critically important to financial stability. Their failure could have consequences far beyond their own customers or shareholders. They’re what we colloquially call too big to fail.

For decades, these institutions have built enormous competitive advantages. Trust, distribution, regulatory barriers, access to capital, and the arduous process of moving financial relationships have all contributed to their dominance. Importantly, the largest banks have benefited from a financial system built around a fundamental assumption: the customer is a human.

Humans are remarkably inefficient financial actors. We leave money sitting in accounts paying next to nothing in interest. We pay fees we don’t understand. We maintain banking relationships for decades without evaluating alternatives. We rarely optimize how our money moves, where it’s held, or what it’s earning. Banks have built extraordinarily valuable businesses around these behaviors.

For the first time in decades, there is a new customer profile. And the largest banks are responding by closing their doors to it.

The banks are blocking agents

Over the last week, Jamie Dimon at JPMorgan Chase, Brian Moynihan at Bank of America, Jane Fraser at Citi, and Rick Wurster at Charles Schwab have all been at the center of a growing conversation about restricting third-party AI agents from accessing customer accounts. The approaches vary.

JPMorgan and Schwab have explicitly said they’re blocking unauthorized agents. Citi is restricting third-party transactional access, while Bank of America is emphasizing controlled integrations and human authorization. The broader signal is increasingly clear: the largest financial institutions are neither prepared nor particularly incentivized to give autonomous software unrestricted access to their customers’ money.

On the surface, the argument is straightforward: consumer protection. And to be fair, a lot of it is legitimate.

We’re entering an era where agents are increasingly capable of taking actions across nearly every part of our personal and professional lives. They can browse the internet, access external systems, write and execute code, and make decisions without continuous human oversight.

But it’s still the Wild West. Security standards are immature. Permissioning models are inconsistent. Startups are racing to prove that their agents can do more, faster, often without having developed the infrastructure necessary to safely support that autonomy. Meanwhile, consumers are giving software unprecedented access to their lives, often without fully understanding the scope of the authority they’ve delegated.

Combine all of that with money movement, and it’s not difficult to imagine some truly catastrophic outcomes. We should take those risks seriously. But I don’t believe consumer protection is the whole story. There’s a much bigger, more important economic story unfolding underneath these announcements.

The largest banks aren’t just defending their customers. They’re defending the economics of their businesses.

The end of banking inertia

Banks are money-making machines. They use customer deposits to fund lending and other earning assets, generating interest revenue on the spread. There are, of course, other sources of revenue: fees, value-added services, and income from card transactions, among others. But the simplistic view is directionally correct.

What makes these businesses incredibly enduring is that customer relationships tend to be highly sticky and decades long. Banks fight to become your primary depository institution, and once they win that relationship, customers rarely leave.

This has been true for most of modern banking history. Financial institutions have competed to acquire and retain human customers. Branches, advertising, rewards programs, mobile applications, and relationship managers have all been built around the same objective: winning and retaining the customer relationship. Once established, that relationship has historically been remarkably difficult to displace.

Moving banks means changing direct deposits, updating bill payments, transferring balances, replacing cards, and learning a new interface. Even when the economics of switching are compelling, the inconvenience frequently outweighs the benefit. Agents change that calculation.

An agent doesn’t care that you’ve banked with the same institution for 20 years. It doesn’t care how beautiful the mobile application is, whether there’s a branch down the street, or whether a bank spent hundreds of millions of dollars advertising during the Super Bowl. It’s driven by the objective it’s been given and the tools available to accomplish it.

If an agent can earn another 200 basis points on idle cash, why wouldn’t it? If it can move money faster, cheaper, or more reliably through another financial institution, why wouldn’t it? If one bank exposes the APIs necessary to accomplish a task and another requires navigating a web interface built for humans, which institution do you think it will choose?

For the first time, banks will compete for customers that make decisions programmatically. And those customers will be extraordinarily good at comparing their options.

Optimal systems require programmatic access

This poses an entirely new challenge. If agents are highly efficient systems designed to maximize their reward functions, then they need software that lets them accomplish the objectives they’ve been given: open accounts, issue cards, move money, accept payments, and much more.

We’re already seeing this challenge play out in other industries. Take Instinct and restaurant reservations. What happens when marketplaces like Resy don’t provide authorized, programmatic access for agents? Systems designed around human interactions get hammered by agent-initiated requests. Traffic gets flagged. Accounts get blocked. Agents attempt to brute-force their way around these restrictions to fulfill their objectives.

The same tension will emerge in financial services and money movement. As agents become increasingly autonomous financial actors, they’ll either gravitate toward systems designed to support them or continue attempting to operate through infrastructure that wasn’t built for them.

The banking system wasn’t built for this

There’s a more fundamental infrastructure problem. Much of the banking system runs on technology designed long before autonomous financial actors were conceivable.

Across the United States, core providers like FIS, Fiserv, and Jack Henry underpin a substantial portion of financial institutions. The largest banks also operate enormous collections of proprietary and vendor-built systems accumulated over decades. These systems are responsible for the fundamental operations of banking: maintaining account ledgers, processing transactions, posting balances, calculating interest, and integrating with the infrastructure necessary to move money.

They’re extraordinarily important pieces of infrastructure. But much of the underlying architecture wasn’t designed for the world we’re entering.

Many legacy banking systems rely on a combination of real-time processing and batch-oriented operations. You can think of batch processing as queuing certain actions to be processed together at scheduled intervals. A transaction might affect an available balance immediately, while final posting, reconciliation, and other accounting processes happen later. Newer cores can support continuous posting, but even then, the surrounding systems and payment rails often operate on entirely different schedules.

That complexity compounds with the volume and frequency of financial activity at scale.

Maybe it’s manageable when you’re reconciling a few thousand transactions between parties, but what happens when those parties need to reconcile a few million events over the same period? How do you maintain an accurate view of liquidity, enforce transaction ordering, manage settlement risk, and provide immediate execution confirmation when the systems responsible for each function weren’t designed to operate together continuously?

When the number of agents becomes orders of magnitude larger than the number of humans, and those agents are capable of initiating transactions at machine speed, the limitations of fragmented, batch-dependent infrastructure become increasingly difficult to ignore.

The chain of command

Historically, banks have been built around people and businesses operating accounts, with software supporting those relationships. Agents invert that model. The actor initiating financial activity is increasingly software itself.

An agent needs to establish its identity, receive scoped financial authority, initiate transactions, manage balances, access payment rails, and prove what it did, why it did it, and who authorized it. All programmatically. Not through a mobile application. Not by logging into online banking with a customer’s credentials. Not by navigating interfaces designed for a human sitting at a computer. Through infrastructure built for software to act.

And not just occasionally. Continuously, at enormous scale, with the auditability, reliability, and security that operating inside the financial system requires. The challenge isn’t simply adding an API on top of an existing bank. It’s rethinking the architecture, permissioning, compliance, and operating model of the institution itself.

An agent-native bank rests on a fundamentally different assumption. Agents need their own verifiable identities, linked to the people or businesses they’re acting on behalf of, allowing financial institutions to distinguish authorized software from unknown actors.

Customers should be able to define precise controls and permissions governing an agent’s authority, with the bank enforcing those limits at the point of payment initiation. Agents should be able to take actions with real-time visibility into balances and transaction status, and access settlement rails suited to their needs. Every action should be documented alongside the full chain of authority: which agent took the action, on whose behalf, under which permissions, and for what purpose.

This is a profound change to the competitive structure of banking. For decades, banks have benefited from the friction associated with moving money. Agents are going to spend the next decade systematically eliminating that friction.

The end of banking distribution as we know it

For the better part of a century, banks have competed for customers through some combination of brand, physical distribution, trust, product offerings, and increasingly, software. JPMorgan Chase has nearly 5,000 branches across the United States. Bank of America spends billions of dollars annually on technology. The largest banks have spent decades and hundreds of billions of dollars building distribution networks designed to acquire customers and capture an increasing share of their financial lives.

All of that infrastructure has been built around one fundamental assumption: the customer is a human. But what happens when that customer is increasingly represented by an AI agent?

Agents don’t need branches. They don’t care about the mobile banking experience. They don’t develop brand loyalty in the traditional sense, and they have no reason to prefer an institution simply because it has historically dominated the market. They’re software, optimizing ruthlessly for a set of objectives on behalf of the people and businesses they represent.

In that world, banks aren’t simply competing to acquire a customer relationship. They’re competing to be selected by the software controlling that relationship.

That is a radically different competitive environment.

For the first time, the largest banks face the possibility that they won’t control the primary interface through which their customers interact with money. The agent becomes the interface, the decision-maker, and potentially the orchestrator of financial activity across dozens of underlying institutions. The banks become providers of infrastructure that those agents can choose to use.

And when software is responsible for choosing the infrastructure, decades of investment in consumer distribution become substantially less relevant. What matters is programmability, reliability, access to financial rails, breadth of capabilities, and economics.

Not every incumbent advantage erodes. Agents cannot independently provide the regulated infrastructure of banking: charters, balance sheets, insured deposits, custody, fraud and credit risk management, and trusted identity verification. If anything, as agents move more money, the importance of these capabilities will only grow.

What diminishes is the value banks derive from consumer habit and inattention. Their competitive advantage will increasingly depend on the quality of their offerings and how well they serve not only their customers, but the agents acting on their behalf.

This is where I believe the largest banks have the most to lose. Not because agents will eliminate the need for banks, but because agents could fundamentally change which institutions capture the economic value of financial activity.

The incumbent banks understand the importance of this transition. But understanding that something is coming and being structurally capable of responding to it are two very different things. And when the economics of your existing business are built around a world where humans make financial decisions, the incentive to embrace software that makes those decisions more efficient isn’t necessarily aligned with protecting your existing franchise.

That tension is going to become increasingly difficult to ignore.

Agents will accelerate banking consolidation

I believe agents are going to accelerate one of the largest consolidations of financial activity in the history of banking. Not necessarily through traditional mergers and acquisitions, although I expect plenty of those too. Through the concentration of deposits, payment volume, and eventually credit activity around a smaller number of financial institutions built to support programmatic financial actors.

Today, the United States has approximately 4,200 banks. In an agentic economy, the number of banks matters far less than the number of banks that agents actually choose to use.

Agents will increasingly route financial activity toward institutions that are the most programmable, provide the greatest degree of control, have the most reliable infrastructure, and offer the most rational economics. Unlike humans, agents won’t have the same inherent attachment to brand, convenience, or decades-old financial relationships. And as their capabilities improve, the cost of evaluating and switching between financial institutions will continue to fall.

That creates enormous pressure on both ends of the market.

Community and regional banks face a world where geography and local relationships may become less important to certain categories of financial activity, while the technological investment necessary to compete becomes increasingly significant. The G-SIBs face a different problem. Their scale, distribution, and entrenched customer relationships are enormous advantages today. But those same advantages can make it difficult to embrace technologies that compress margins, reduce switching costs, and challenge their control over customer relationships.

In both cases, the financial institutions that win the next decade won’t necessarily be the ones that won the last.

Consolidation vs. fragmentation

In theory, agents could make the banking landscape more fragmented. Intelligent software could spread deposits across thousands of banks in pursuit of the best possible yield.

But integration comes at a cost. Connecting to each additional institution requires building, monitoring, and maintaining another integration. As a result, developers will naturally gravitate toward platforms that offer the greatest breadth of payment rails, financial products, and capabilities through a single integration.

Settlement also favors scale. A payment between two accounts held at the same bank can often be posted immediately on its own ledger, without the fees or operational complexity of moving funds across institutions. Payments between banks, meanwhile, require interbank clearing and settlement, introducing additional infrastructure, costs, and constraints. The more agents and counterparties share a financial institution, the more efficiently money can move between them. Trust and compliance can also benefit from network effects, as common protocols and standards make it easier for additional participants to join.

Credit follows capital and data. An institution with real-time visibility into an agent’s cash flows has the potential to underwrite far more intelligently than one relying on historical statements. And extending credit on demand requires access to capital and the infrastructure to manage risk. This creates a flywheel: agents gravitate toward institutions where credit is readily available, bringing more financial activity and data. Better data can improve underwriting and risk management, enabling more competitive credit products that attract still more activity.

While deposits may remain distributed across many institutions, control over financial activity will likely concentrate within a much smaller number of platforms and institutions built to support agentic economic activity.

The next generation of banking

I don’t believe the answer is to force agents to behave like humans, or to retrofit every part of the existing banking system with increasingly elaborate ways to prevent autonomous software from taking action. The answer is to build financial institutions designed for agents from the beginning, allowing them to accomplish their objectives safely and efficiently under an account holder’s delegated authority.

Banks where software is treated as a first-class customer. Where access to accounts, payment rails, capital, lending, and financial infrastructure is fundamentally programmatic. Where permissions are explicit, transactions are auditable, and compliance is built directly into how money moves. Where agents can safely accomplish in milliseconds what takes humans hours, days, or weeks.

Banking built to serve software as its primary financial actor needs to be architected differently from the ground up, not merely retrofitted. This is what we’re building at Natural*.

The regulatory catch-up game

Regulation has always trailed adoption. Much of the framework governing electronic consumer payments today traces back to the Electronic Fund Transfer Act of 1978, written decades before autonomous software could initiate transactions on behalf of consumers. The rise of open banking involved years of back-and-forth between banks and fintechs before new regulations were drafted, with those rules still being contested today. At the rate agents are proliferating, agentic finance threatens to widen the gap between how money moves and the regulations designed to govern it.

In the absence of comprehensive agent-specific regulatory guidance, institutions and industry stakeholders are beginning to step in. Companies and industry groups are racing to develop protocols for verifying agent identity and delegated authority, while card networks are laying the groundwork for authenticating and managing disputes around agent-initiated payments.

The next few years will help establish the foundation for decades of banking regulation. The companies already running agents in production — those putting agents in people’s hands and those building the infrastructure that moves their money — will help establish the precedents governing agents as economic actors: liability, consumer protection, authorization, and identity verification.

What’s ahead of us

We’re very early. There will be real security failures, regulatory challenges, and plenty of reasons for institutions to move cautiously. Those concerns should influence how this infrastructure gets built, but they won’t change the fundamental direction of the market.

Agents are becoming economic actors. The financial system will have to adapt to them. That adaptation doesn’t require choosing between security and interoperability. Blanket restrictions on agents won’t solve the underlying challenges of either. The path forward is to build infrastructure designed for agents from the ground up, while engaging early with regulators and institutional leaders to ensure that new capabilities develop alongside the trust, safety, and regulatory frameworks necessary to support them.

For the last century, banks have competed to win the human customer. For the next century, they’ll have to compete to win the agent.