The Agent Becomes the Bank

In When Intelligence Becomes Cheap, I argued that the most important consequence of falling AI costs may not be better chatbots. It may be a migration of economic value. If intelligence becomes abundant and increasingly interchangeable, value moves toward the things intelligence is allowed to do: agency, identity, trust, payments, proprietary data, infrastructure and ultimately control over human intention.

That argument naturally leads to a more uncomfortable question.

What happens when the agent is no longer merely allowed to buy something for us, but is allowed to manage money itself?

The Uncomfortable Question

At first the answer seems straightforward. An AI agent can have an identity. It can have a wallet. It can authenticate itself, examine transactions, evaluate prices, make payments and maintain a complete digital history.

Why, then, should it still need a bank?

If almost every financial event is recorded digitally, why should creditworthiness depend on a traditional institution examining a static credit score? An AI could potentially understand a person’s or company’s financial condition much more precisely by looking continuously at income, expenses, assets, contracts, transaction history and cash flow. It could calculate risk, determine an appropriate interest rate, issue credit, monitor repayment and adjust the relationship continuously.

Follow that logic far enough and an intriguing conclusion appears.

If the AI controls the digital money, evaluates the borrower, decides whether to extend credit, manages repayment and executes the transactions, then at an operational level the AI has begun to perform what we traditionally call banking.

That was the central realization at the end of my original exploration: once software controls money, risk assessment, lending and transactions, the distinction between “AI used by a bank” and “AI functioning as the bank” becomes increasingly difficult to maintain.

But there is a complication, and it is precisely what makes the question interesting.

A bank is not simply a collection of tasks. It is also a legal institution with a peculiar relationship to the state, the monetary system and risk. Artificial intelligence can potentially automate almost everything a banker does. That does not mean it automatically acquires everything a bank is.

Understanding that distinction may tell us what banking becomes in the agentic economy.

What Is a Bank, Once You Remove the Building?

For most people, a bank appears to be an interface. It is an app, a branch, a credit card, an account number, a mortgage application, perhaps a relationship manager. Remove those interfaces and it is tempting to think that very little remains.

But the economically important parts of a bank lie underneath them.

A commercial bank performs several functions simultaneously:

• It stores claims that society accepts as money.

• It evaluates credit.

• It transforms short-term liabilities into longer-term assets.

• It absorbs risk on its balance sheet.

• It connects to payment and settlement systems.

• And, crucially, when it makes many types of loans, it can create new deposit money.

This last point is often described badly. Banks do not simply take a pile of existing deposits from savers and hand the same money to borrowers. Nor can they create unlimited wealth by typing numbers into a computer.

The Bank of England’s explanation is clearer: when a commercial bank makes a loan, it normally creates a corresponding deposit in the borrower’s account. New bank money has been created, while an equivalent debt has simultaneously been created on the other side of the balance sheet. As the loan is repaid, that deposit money can be destroyed again.

This ability is extraordinary. Most companies cannot do it. If a normal investment fund lends me €100,000, it must obtain that €100,000 from somewhere: investors, existing cash, borrowing or asset sales. One account falls while mine rises. The transaction reallocates existing money.

A deposit-taking commercial bank can create a new deposit as part of extending the loan. The Bank of England has explicitly described this capacity to create money through lending as a unique privilege associated with the banking licence.

But privilege comes with constraints. Banks must maintain capital against potential losses. They need liquidity. They settle with other institutions using central-bank money. They are supervised. Their deposits are expected to remain exchangeable at par with sovereign money. Depending on the jurisdiction, they may have access to central-bank facilities and deposit-insurance systems.

A bank that makes terrible loans cannot simply continue creating deposits indefinitely without consequences.

So when we ask whether AI can replace the bank, we are actually asking two different questions:

1. Can AI replace what bankers do?
   Increasingly, yes.

2. Can AI automatically inherit the institutional privileges that define what banks are?
   Not without the state.

And the future of banking may emerge precisely from the collision between those two answers.

Credit Scoring Was Built for an Information-Poor World

Consider lending.

A traditional lender has always faced an information problem. The borrower knows more about himself than the lender does. A business owner understands the health of her company better than a bank reviewing a loan application. A consumer knows whether his job feels secure, whether another large expense is approaching and whether he is already financially stretched.

Banks developed systems for compressing incomplete information into manageable proxies: income statements, collateral, credit reports, repayment histories, debt-to-income ratios, bureau scores and human underwriting.

These systems were necessary partly because gathering perfect information was impossible or prohibitively expensive.

But the economy is becoming increasingly digital.

A small company may already process nearly every transaction electronically. Its sales, payroll, inventory, tax payments, customer refunds, subscriptions and supplier invoices can exist as machine-readable records. A consumer may have years of transaction history revealing income stability, recurring obligations, savings behavior, spending volatility and existing debts.

An AI system does not necessarily have to ask, “What is your credit score?”

Given appropriate permission, it could ask something much richer:

• How stable is this person’s income?

• How volatile are expenditures?

• What proportion of income is already committed?

• Are cash reserves improving or deteriorating?

• Is the business acquiring customers?

• Are invoices being paid more slowly?

• Is revenue seasonal?

• What happened during previous shocks?

• How concentrated are suppliers and customers?

• What contractual income is already expected during the period of the proposed loan?

Instead of evaluating a borrower once every several years when a loan application appears, credit could theoretically become a continuously updated model.

That could alter lending profoundly.

A restaurant with temporarily weak cash flow but confirmed reservations, stable supplier relationships and years of transaction history might be understood more accurately than through a conventional score. A small online merchant experiencing a sudden surge in orders might obtain working capital based on live sales rather than waiting weeks for a bank officer to review historical accounts. Interest rates could potentially reflect changing risk more dynamically rather than categorizing millions of borrowers into relatively crude bands.

Some versions of this already exist in simpler form. Payment processors and commerce platforms have long used transaction histories to extend financing to merchants. What AI adds is the possibility of interpreting vastly more context, continuously and at much lower cost.

But this creates an important distinction.

Better credit intelligence does not necessarily require a bank. Credit creation may.

And that opens the financial system to new competitors.

If the Bank No Longer Decides the Loan, Who Does?

Imagine that an AI can assess credit better and more cheaply than a traditional lending department. It does not follow that lending disappears. Capital still has to come from somewhere, losses still have to be absorbed and somebody must ultimately own the risk.

There are several possibilities.

The first is the least revolutionary and perhaps the most likely: the banks themselves become AI-native. The bank survives, but much of what we currently recognize as banking employment and workflow disappears behind autonomous systems. The institution retains its charter, capital, balance sheet, access to settlement systems and regulatory obligations, while AI increasingly determines which customers qualify for credit, how loans are priced, how liquidity is managed, which transactions require investigation and how risk is distributed across the balance sheet.

In that world the bank has not been destroyed by AI. The bank has been hollowed out and rebuilt around it.

The second possibility is that increasingly sophisticated non-bank capital competes directly with banks. Private-credit funds, investment pools, corporations, fintech lenders and decentralized protocols do not need the legal power to create deposits in order to lend. They simply need capital. An AI can connect people and institutions seeking yield with borrowers seeking funding, evaluate both sides and construct the loan.

This model is fundamentally different from commercial bank credit. The money being lent must already exist somewhere. But if global capital becomes easier to discover, price and allocate automatically, that limitation may matter less for many forms of financing.

An AI agent representing a business could request €200,000 of working capital. Instead of sending an application to one bank, it could expose a verified financial profile to dozens or hundreds of competing capital providers. Other agents could analyze the risk, submit rates and terms, and negotiate automatically. The loan might ultimately be funded by a bank, a pension fund, a private-credit vehicle, a corporate treasury, a tokenized fund or some combination of them.

The borrower would no longer necessarily care who supplied the capital.

The agent would care about the conditions.

That resembles what happened to other parts of the internet. Consumers once navigated individual providers manually. Platforms aggregated them. Agents could now aggregate the aggregators themselves.

Credit could become another market that software continuously shops.

Amazon May Know Your Business Better Than Your Bank

There is another possibility, and it should concern traditional lenders even more.

The institution with the best understanding of a borrower’s risk may increasingly be the company through which that borrower conducts business.

Consider a merchant selling primarily through an online commerce platform. That platform may know its daily sales, customer-return rates, advertising performance, inventory turnover, seasonal patterns and payment history. A bank sees money entering and leaving an account. The commerce platform may see the economic activity that produced the money.

Who possesses the better credit model?

The same argument applies elsewhere. Payment processors see flows. Payroll companies see employment. Accounting systems see invoices and obligations. Logistics networks see inventory moving through warehouses. Marketplaces see actual demand. Enterprise software providers see contracts and sales pipelines.

Once AI can combine these signals, the informational advantage historically enjoyed by banks may weaken.

A platform could say to a merchant: your AI expects €310,000 of confirmed sales over the next four months. Inventory constraints are limiting growth. Based on your transaction history, return rate and customer concentration, here is €60,000 of financing at this price. Repayment will occur automatically as a percentage of future sales.

No traditional credit application is necessary.
No branch is necessary.
Perhaps no conventional credit score is necessary.

The lender might still ultimately be a regulated bank behind the scenes. Or the platform may fund the credit from its own balance sheet or a partner’s capital pool. To the customer, the distinction becomes invisible.

This is similar to the transformation I described in When Intelligence Becomes Cheap. The interface disappears, while the underlying infrastructure remains.

The same thing could happen to banking.

You may still be borrowing bank-created money.
You may simply stop experiencing the bank.

From Banking Interfaces to Financial Agents

The payments industry already gives us an indication of this transition.

Visa now describes AI agents that can book travel, reorder inventory, purchase computing resources and pay for services. Mastercard is building payment infrastructure specifically for a world in which agents transact continuously with other agents. In March 2026, Santander and Mastercard announced a live end-to-end payment executed by an AI agent within a regulated banking framework, with the agent operating under predefined permissions and limits.

By June, Mastercard had introduced Agent Pay for Machines, designed to support machine-speed payments across cards, bank accounts and stablecoins. Visa has similarly expanded infrastructure for agent-led commerce, including agent registries and scoring systems intended to establish identity and trust around autonomous transactions.

These developments matter because they reveal how incumbents understand the threat.

Visa and Mastercard are not assuming that humans will remain at the center of every payment interaction. They are redesigning themselves so that if software becomes the customer-facing decision-maker, the networks can remain underneath the transaction.

Banks will probably attempt the same thing.

Their application may disappear before their balance sheet does.

And this creates a useful way to think about technological disruption. Institutions often survive by surrendering the visible layer and defending the invisible one.

The question for banks is whether the invisible layer is enough.

The Bank Could Become a Balance Sheet Wrapped Around an AI

Imagine a financial institution twenty years from now.

There may still be a legal company with shareholders and directors. It may still be supervised by regulators. It may still maintain required capital and liquidity. It may still have accounts at a central bank and obligations to customers.

But almost everything between those legal boundaries could be software.

• An AI treasury system manages liquidity continuously.

• Another system evaluates loan applications.

• Another monitors existing borrowers.

• Another buys and sells assets to optimize the balance sheet.

• Another performs compliance monitoring.

• Another investigates suspicious transactions.

• Another manages currency exposure.

• Another negotiates wholesale funding.

• Customer agents communicate directly with the bank’s agents.

The traditional organizational chart begins to look strange.

Where is the bank?
Is it the employees?
The software?
The balance sheet?
The licence?
The capital?
The legal entity?
Or the regulated combination of all of them?

My answer is that the bank increasingly becomes a legally authorized balance sheet controlled through software.

That sounds less dramatic than saying AI destroys banking, but economically it may be far more important. The familiar institution survives while much of the activity that justified its organizational complexity disappears.

The bank becomes closer to a protocol.

Not an uncontrolled decentralized protocol, because regulators are unlikely to surrender authority over money creation and systemic risk. Rather, a supervised financial machine whose rules, capital requirements and permissible actions are established by law while most day-to-day decisions are computational.

A recent development in the United States illustrates how close parts of finance are already moving toward code-native institutional structures, although it is important not to exaggerate what has happened. In September 2026, the Office of the Comptroller of the Currency gave preliminary conditional approval to Catena Trust Bank, a proposed national trust bank designed around fiat currency, securities and digital assets, including payment stablecoins. This is not an autonomous AI commercial bank and it is not authorized as an ordinary deposit-taking lender. The OCC explicitly restricts its operations to those of a trust company and related activities. But its existence shows something important: the legal architecture of banking is already being adapted to institutions whose core business is digital assets, custody, clearing and programmable money rather than the traditional branch model.

The direction matters more than the label.

Finance is becoming increasingly software-native while remaining legally institutional.

Why the State Does Not Disappear

At this point it is tempting to take the argument one step too far.

If AI can assess risk, digital wallets can store value, stablecoins can move globally and smart contracts can execute payments automatically, why not remove banks—and perhaps eventually central banks—entirely?

Because money is not simply technology.

Money is also law, taxation, political authority, collective trust and ultimately a claim on a social system.

Commercial banks create deposit money, but that money works because society believes a euro in one regulated bank is equivalent to a euro somewhere else. Banks settle obligations through central-bank reserves. Governments collect taxes in sovereign currencies. Courts enforce contracts. Regulators decide what qualifies as capital. States determine which liabilities receive deposit protection and which institutions can access emergency liquidity.

The Bank of England describes central-bank reserves as the ultimate settlement asset and the anchor of the monetary system. Commercial-bank deposits function as money partly because the institutional architecture is designed to ensure that they remain convertible at par with central-bank money.

AI does not eliminate this architecture merely by making it more efficient.

Indeed, greater automation may make the state more important.

A financial system operating at machine speed can make mistakes at machine speed. If thousands of autonomous agents react to the same signal simultaneously, liquidity could move extraordinarily quickly. Credit could tighten automatically. Assets could be sold. Collateral could be liquidated. What once took committees, telephone calls and trading days could occur in seconds.

Efficiency and stability are not the same thing.

Human friction is often wasteful, but sometimes friction slows contagion.

An agentic financial system will therefore require something AI enthusiasts sometimes underestimate: constraints. Spending limits. Capital rules. Circuit breakers. Audit trails. Identity. Liability. Human override. Regulatory access.

The future bank may be much more autonomous precisely because its legal boundaries become more explicit.

The Real Threat to Banking Is Not Elimination

This leads to a more subtle conclusion.

The great threat to banks may not be that they disappear.

It may be that banking becomes dramatically less profitable as an activity.

Much of banking’s cost structure comes from information processing, administration, compliance, distribution and human decision-making. If AI reduces those costs substantially, competitors can operate on thinner margins. Credit decisions become faster. Switching becomes easier. Agents continuously search for better deposit rates, better loans and lower fees.

Today, customers are often remarkably passive. Money remains in accounts paying poor interest. Mortgages remain with existing institutions. Companies tolerate expensive payment services. Consumers fail to refinance debt because finding alternatives takes time and effort.

Agents do not become tired.

Imagine a financial agent continuously optimizing your balance sheet. It notices €40,000 sitting in a low-yield account and moves it to a better alternative within your approved risk parameters. It refinances a loan when the expected saving exceeds transaction costs. It renegotiates insurance. It changes payment providers. It adjusts investments. It searches globally for competing credit offers.

The economics of customer inertia begin to disappear.

That may be much more dangerous to incumbent financial institutions than a cryptocurrency replacing their currency.

Many mature industries earn substantial margins because humans do not continuously optimize.

AI potentially does.

Banks could therefore survive while losing some of the economic rents produced by inconvenience, information asymmetry and customer inattention.

This is the same pattern appearing throughout the agentic economy. The hotel survives. The booking interface becomes vulnerable. The manufacturer survives. The marketplace becomes vulnerable. Capital survives. The financial interface becomes vulnerable.

AI attacks friction before it attacks assets.

But Who Bears the Loss?

There is still one question that no amount of machine intelligence eliminates.

What happens when the loan goes bad?

An AI may predict default better than a human. It may price risk more accurately. It may monitor the borrower every second.

But prediction is not certainty.

A restaurant burns down. A company loses its largest customer. A property market collapses. A war begins. A previously stable correlation breaks. Thousands of borrowers who appeared unrelated suddenly default together.

Somebody must absorb that loss.

This is why capital remains fundamental.

The most sophisticated AI credit engine in the world does not eliminate financial risk. It decides who should own it and at what price.

A bank absorbs losses through its equity and reserves before they reach protected depositors. A private-credit fund passes them to investors. A decentralized lending pool distributes them according to protocol rules. A corporation lending from its own balance sheet absorbs them itself.

AI can optimize risk allocation.

It cannot abolish risk.

And this may ultimately define the institution that survives the agentic transformation. The future bank is not necessarily the organization best at predicting the future. It may be the legally recognized entity capable of standing behind the consequences when the prediction is wrong.

That is an important distinction.

Intelligence can be cheap.
Responsibility cannot.

When the Agent Becomes the Bank

The evolution may therefore be less dramatic visually than we imagine.

There may be no moment when we wake up and discover that banks have vanished.

Instead, functions migrate.

• Credit scoring moves into AI.

• Customer service moves into AI.

• Treasury management moves into AI.

• Fraud detection moves into AI.

• Payments become agentic.

• Contracts become machine-readable.

• Capital searches for borrowers automatically.

• Borrowers search for capital automatically.

• Bank accounts become APIs.

• Money becomes increasingly programmable.

And gradually the institution we call a bank changes character.

The human organization shrinks.
The software expands.

What remains around it is the balance sheet, capital, licence, legal identity and connection to the sovereign monetary system.

At that point the distinction becomes almost philosophical.

If an AI decides which loans to make, determines their price, creates or allocates the money, manages liquidity, monitors repayment, moves funds and continuously restructures the portfolio, is the AI merely working inside the bank?

Or has the bank become the legal shell through which the AI is allowed to participate in the economy?

I suspect the second description will increasingly become more accurate.

That does not mean autonomous machines will somehow escape human law and become independent financial sovereigns. The opposite may happen. The more financial autonomy software receives, the more precisely society will have to define the boundaries of that autonomy.

• Who owns the agent?

• Who supplies its capital?

• Who determines its objective?

• Who audits its decisions?

• Who absorbs its losses?

• Who can stop it?

• And who is liable when it is wrong?

Those questions will define the architecture of agentic finance far more than whether the underlying payment happens through a card, a bank deposit or a stablecoin.

From Cheap Intelligence to Autonomous Capital

This is where the argument connects back to the first essay.

When Intelligence Becomes Cheap began with a simple economic observation. If machine intelligence becomes less scarce, the value created by AI cannot remain concentrated forever in the price of intelligence itself. It moves toward the systems through which intelligence acts.

Banking may become one of the clearest demonstrations of that transition.

First AI answers financial questions. Then it analyzes accounts. Then it recommends transactions. Then it executes transactions. Then it evaluates credit. Then it allocates capital. Then it manages a portfolio. Then it manages a balance sheet.

There is no obvious line along that progression where we can confidently say: before this point the system was merely an assistant, and after this point it became the institution.

Eventually the software is performing almost every economically significant decision traditionally associated with the organization.

What remains is authority.

And perhaps that is the deeper pattern emerging across the agentic economy.

The future may not belong simply to whoever possesses the most intelligence. Intelligence may become abundant.

The scarce resource may be legitimate authority over what that intelligence can control.

In commerce, that means permission to transact.
In software, permission to act on corporate systems.
In marketplaces, permission to represent buyers and sellers.

And in finance, it means something even more consequential: permission to decide who receives capital, at what price, under what conditions, and who ultimately bears the loss.

Once an AI controls those decisions, asking whether it is “used by the bank” may become increasingly semantic.

Because when software controls the money, evaluates the risk, allocates the credit and executes the transaction, we may discover that we did not automate the bank.

We changed what a bank is.

Karim Spring