1. JPMorgan AI agents move from assisting staff to executing transactions
JPMorgan Chase, the largest bank in the United States, has built artificial intelligence agents that can carry out financial transactions with no human involved, following plans it set out in June 2026 for agents that work unattended for up to two hours. On October 13, 2026, Chief Executive Jamie Dimon said the bank blocks outside AI agents from customer accounts.
Taken together, the two positions describe a deliberate strategy. Inside its own walls, the bank is giving software more room to act. At the boundary where customers and third parties connect to it, the bank is closing the door on software it cannot supervise. Both decisions follow from the same concern: who answers for a transaction when no person pressed the button.
The shift is significant because of what it leaves behind. For several years, the most visible uses of AI in banking have been conversational assistants and tools that summarize documents. Those systems produce text for a person to read and act on. An agent that executes a transaction is a different category of software, since its output is an action with financial consequences and not a suggestion.
2. What do JPMorgan's AI agents do?
JPMorgan's agents are built to run complex financial workflows from start to finish with limited ongoing supervision. Some of them operate autonomously, and others are semi-autonomous, meaning a person remains involved at certain points in the process.
The distinction between an agent and earlier forms of automation is worth spelling out. Traditional automation follows a fixed script: when a defined condition is met, a defined step is performed. An AI agent is given an objective and works out the intermediate steps itself, calling on data and tools as it goes. That flexibility is what allows an agent to handle a workflow with many stages, and it is also what makes oversight harder, because the path the software takes is not laid out in advance.
In June 2026, the bank announced that it intended to deploy what it calls long-running agents. These are designed to keep working for as long as two hours without any human input. Duration matters in this context. An assistant that answers a question finishes in seconds, and a person checks the answer immediately. An agent that runs for two hours may take many separate decisions before anyone reviews the result, which raises the stakes for how the system is designed and monitored.
JPMorgan presented the rollout in terms of regulatory compliance and internal control. For a bank, that framing is not incidental. Every transaction a bank processes sits inside a framework of rules covering record keeping, approvals and accountability, and any system that acts on the bank's behalf has to fit within it.
3. A treasury example: catching a $2.3 million currency mismatch
A whitepaper the bank published in June offers a concrete picture of the kind of work these agents are meant to do. The example comes from corporate treasury, the function inside a company that manages cash, payments and exposure to moving exchange rates.
In the scenario, an agent processing a payment noticed that the currency of the transaction had changed. A change of that kind can easily go unnoticed in a busy payments operation, because the amount and the counterparty may look familiar even though the currency is different. Left alone, it can leave a company owing money in a currency it had not planned for, with no protection against exchange-rate movements.
The agent flagged the switch, and the whitepaper says this headed off a potential foreign exchange mismatch of $2.3 million. A mismatch arises when a company's obligations in one currency are not offset by holdings or hedges in the same currency, leaving it exposed if rates move against it.
The example illustrates the argument for agents in finance. Much of the value does not come from executing a routine payment faster. It comes from software that is watching every item, recognizes when something departs from the expected pattern and brings it to attention before it becomes a loss. It also shows the semi-autonomous model in practice: the agent detects the problem and surfaces it within the bank's control processes.
Treasury operations suit this kind of software for a practical reason. Large companies send and receive payments in many currencies every day, and the teams responsible for them review a small sample closely while relying on standard checks for the remainder. An agent does not need to sample. It can examine each payment against what was expected, which is how a single altered detail, such as the currency on one transaction, gets noticed at all.
4. Why is JPMorgan blocking outside AI agents?
JPMorgan blocks external AI agents from customer accounts because of liability. The bank does not want to be responsible for what happens when software it did not build, and does not control, moves money or takes other actions inside an account it holds.
By late 2026, the bank had put defensive measures in place that prevent such agents from gaining access to customer accounts. Dimon addressed the policy on the bank's earnings call on October 13, 2026, indicating that outside agents are blocked to avoid the liability risks attached to customer accounts.
He also drew a line on responsibility. Customers who choose to give an AI agent access to their accounts, Dimon said, take on the liability for that decision themselves.
The position reflects a real gap in how responsibility is assigned. When a customer makes a payment, or when the bank's own systems make an error, established rules and practices determine who bears the loss. A third-party agent sits outside that arrangement. If it misreads an instruction, is manipulated, or simply makes a poor decision, it is not obvious whether the customer, the company that built the agent or the bank should absorb the cost. JPMorgan's answer is to prevent the situation from arising on its systems and to tell customers plainly where it believes the responsibility lies if they proceed anyway.
5. Building agents and blocking them at the same time
There is no contradiction in a bank that develops agents and shuts out agents from elsewhere. The difference is control. An agent JPMorgan builds runs on infrastructure the bank manages, under permissions the bank sets, with activity the bank can log and review. If something goes wrong, the bank can trace what happened and is unambiguously accountable for it.
An external agent offers none of that visibility. The bank cannot inspect how it reaches decisions, cannot confirm what instructions it has been given and cannot be sure it is acting as the customer intended. For an institution whose business depends on being able to account for every movement of money, an actor of that kind inside a customer account is a risk it has chosen not to accept.
The same logic appears in the bank's commercial work. JPMorgan has partnered with Mirakl to support agentic commerce, the term for purchases that AI agents initiate or complete on behalf of buyers. The arrangement centers on payment processing and keeps control within JPMorgan's own systems. In other words, the bank is willing to take part in agent-driven transactions where the payment step runs through infrastructure it governs.
6. The economics behind the bank's AI effort
Money explains much of the momentum. JPMorgan's AI work is estimated to produce savings of about $2 billion a year. A figure of that size gives the bank a strong incentive to extend AI from tools that help individual employees to systems that can complete whole processes.
The bank's stated direction is to go past automation of single tasks toward agents that can link several workflows together and make decisions along the way. Each step in that direction increases the potential efficiency gain, because more of a process is handled without waiting for a person. Each step also increases the importance of the controls around the software, since a longer chain of automated decisions leaves more room for an early mistake to carry through.
That tension is why the bank's public description of the program leans so heavily on compliance and control. The savings come from removing people from routine steps, yet the bank's obligations to regulators and clients do not shrink when a person steps away. Whatever an agent does on the bank's behalf, the bank remains answerable for it, so the systems that record, limit and review an agent's actions are as much a part of the product as the agent itself.
7. What the two-track approach means for banking
The questions JPMorgan is answering are not unique to it. Any bank that deploys agents internally has to decide how long they may run unsupervised, which actions require human approval and how their activity is recorded for regulators. Any bank whose customers begin using AI assistants has to decide whether those assistants may touch accounts.
For customers, the practical consequence is clear from Dimon's remarks. Someone who hands account access to an AI agent that the bank does not control should not assume the bank will cover losses that result. Until rules or industry standards settle how liability is shared among customers, agent developers and financial institutions, that responsibility, in JPMorgan's view, rests with the person who granted the access.
For the bank, the approach amounts to a bet that autonomy is safe only where it is paired with control. Agents that JPMorgan has built and can monitor are being trusted with more, including transactions carried out with no human in the loop. Agents from anywhere else remain outside customer accounts.
