When the Numbers Are
Right and the Question Is Wrong JPMorgan’s card business had a good quarter. Debit and credit card sales volume rose 10 percent year over year to $539.5 billion. [1] The Card Services net charge-off rate fell to 2.91 percent, down from 3.15 percent a year earlier. [1] Average card loans climbed to $250.1 billion from $234.4 billion. [1] Every metric pointed in the direction a bank wants its metrics to point. And yet the most revealing statement from the earnings call had nothing to do with spending or credit quality. It was about who gets to touch the machines.
Jamie Dimon drew a line that most companies have not yet drawn. His bank uses AI agents internally — for software development, for operations, for tasks that used to require human hands. But unvetted outside agents, built by other companies, will not be allowed into JPMorgan’s systems. The CEO of JPMorganChase looked at the same technology everyone else is racing to adopt and said: we will use it, but we will not trust it.
That is the gap this quarter’s numbers expose. Not a gap between what AI promises and what it delivers — that story has been told. A gap between what AI appears to do and what it actually does when nobody is watching. JPMorgan’s card portfolio performed well. The AI deployment performed well enough that Dimon mentions it without apology. But the boundary he drew around it tells you something the performance data cannot: the system works, and the system cannot be trusted to work on its own terms.
The Productivity That Doesn’t Show
Up in the Productivity Jeremy Barnum, the chief financial officer, described how JPMorgan accounts for its AI investment. It is not a line item. It is distributed across the company — embedded in software engineering, in operations, in the thousand small processes that keep a bank running. The productivity gains are real, he said, but they do not appear as a single measurable number. They are absorbed into the work.
This is where the deception begins, and it is not the kind of deception that makes headlines. When a company spends billions on AI and cannot point to a line on the income statement that says “AI savings,” two things are true at once. The technology is working — otherwise the spending would have stopped. And the technology is not working in any way that can be independently verified. The gains are real enough to justify the investment and diffuse enough to resist auditing. Every manager who has ever claimed a productivity improvement knows this territory. AI did not invent it. AI just made it larger.
Barnum’s description of distributed AI investment is not evasive. It is accurate. Software engineering productivity improves when an AI agent writes boilerplate code, reviews a pull request, flags a bug. Operations productivity improves when an agent routes a customer query, reconciles a ledger, drafts a report. None of these improvements arrives with a receipt. They arrive as time not spent, errors not made, headcount not hired. The absence of a cost is the hardest thing to measure, and the easiest thing to claim.
Dimon’s internal agents are doing this work right now. They are not being tested in a lab. They are not being benchmarked against human performance in a controlled study. They are inside the systems of a bank whose average Consumer & Community Banking deposits rose 4 percent year over year to $1.097 trillion, and they are producing outcomes that the bank’s leadership describes as positive. [1]. The word “positive” is doing a lot of work in that sentence. It means the agents have not yet done something catastrophic enough to be mentioned on an earnings call. It does not mean anyone outside the bank fully understands what they are doing.
The Boundary That Reveals the Fear
Dimon’s refusal to admit unvetted outside agents is the most honest thing said on the call. He did not say JPMorgan’s agents are better. He did not say the bank has a proprietary advantage. He said it will not give unvetted outside agents access to its systems. The CEO of a bank that employs tens of thousands of people and manages trillions of dollars looked at the agentic AI ecosystem and decided it was not safe enough to connect to.
This is not technophobia. JPMorgan is not refusing AI. It is refusing the premise that AI systems built by others can be trusted to behave inside its walls. The distinction matters because it exposes a contradiction at the heart of the current AI moment. The same technology that is productive enough to deploy internally is dangerous enough to keep at arm’s length externally. The bank trusts its own agents because it built them, because it can see their code, because it can shut them down. It does not trust anyone else’s agents because it cannot see what they will do next.
The outside agents Dimon is describing are not hypothetical. They are the products being sold right now — autonomous systems that book travel, negotiate contracts, manage supply chains, interact with customers. They are marketed as labor-saving, cost-reducing, always-on. And JPMorganChase has looked at them and said: not here. Not connected to our systems. Not with access to our customers’ money.
That is a judgment about the state of the technology, delivered by someone who has every incentive to adopt it faster. If JPMorgan could safely use outside agents, it would. The productivity gains would be substantial. The cost savings would be real. Dimon is not leaving money on the table out of caution. He is leaving it there because the alternative — an unvetted agent with access to a bank’s internal systems — is a risk that no earnings call can justify.
The Consumer Who Doesn’t Know They’re the Test
The card numbers tell a story of health. Spending up 10 percent. Credit losses down. Deposits growing. The labor market holding steady, as Barnum noted, keeping household credit in good shape. The consumer, in the language of the earnings call, remains healthy. But the consumer is also the subject of an experiment they did not consent to. Every internal AI agent that touches a customer query, a credit decision, a fraud flag, or a collections process is making a judgment about a person’s financial life. The judgment is faster than a human’s. The judgment is cheaper. The judgment is not always right.
Dimon cited rogue agents, data misuse, and liability transfers among the concerns surrounding autonomous agents. [1]. These are not abstract risks. They are the specific ways an AI system can harm a person while appearing to help them. A rogue agent that approves a loan it should deny. A data misuse that leaks a customer’s spending patterns. A liability transfer that leaves the bank and the customer arguing over who is responsible for a decision made by a machine. The protection Dimon describes is not a feature. It is a recognition that the technology is not yet safe enough to operate without a human hand on the switch. [1].
The deception here is structural. A customer interacts with an AI agent, receives a response, and assumes the response reflects the bank’s best judgment. What the customer cannot see is the boundary Dimon drew — the line between what the agent is allowed to do and what it is not. The customer experiences a seamless interaction. The bank experiences a controlled deployment. The gap between those two experiences is where trust lives or dies. And right now, the bank is managing that gap by keeping the most dangerous parts of the technology outside the walls.
The Regulation That Arrives After the Damage
There is a pattern in how societies respond to powerful new tools. The tool arrives. The early adopters deploy it. The harms accumulate. The regulation follows. By the time the rules are written, the technology has already reshaped the landscape it operates in. The rules describe a world that no longer exists. The harms they were meant to prevent have already occurred.
Agentic AI is in the early-adopter phase. JPMorgan is deploying it internally. Other companies are deploying it externally. The harms are not yet visible on a balance sheet. The rogue agent that misroutes a payment, the data misuse that exposes a customer, the liability transfer that leaves someone holding a loss — these are not yet line items. They are possibilities that Dimon named because he can see them coming. The regulation that will eventually govern these systems is not yet written. The technology is already inside the bank.
This is the point at which regulation falls behind. Not because regulators are slow, but because the technology moves through the economy faster than any rulemaking process can follow. By the time a law defines what an “unvetted outside agent” is, the definition will be obsolete. By the time a regulator establishes a testing standard, the systems being tested will have evolved past it. The gap between the technology and the rules is not a failure of governance. It is a feature of how technology diffuses through a market economy. The rules come after the harm because the harm is what makes the rules necessary. [1].
JPMorgan’s internal deployment is happening in that gap. The bank is using AI agents in ways that no regulator has approved, no auditor has verified, no law has defined. Dimon’s boundary — no unvetted outside agents — is a private regulation, written by a CEO, enforced by a security team. It is not a public standard. It is a judgment call made by someone who understands the risk because he can see the systems from the inside. The customers whose data flows through those systems have no equivalent visibility. They have a bank statement and a mobile app and the assumption that the institution knows what it is doing.
The Question That Reassesses Everything
Dimon’s position on outside agents is a snapshot, not a permanent stance. It is a snapshot of a moment when the technology is powerful enough to use and dangerous enough to contain. The question he leaves open is the one that matters: what would have to change for the boundary to move? What would an outside agent have to prove — about its behavior, its transparency, its liability — before a bank like JPMorgan would let it inside? And if that proof is possible, who verifies it? Who writes the standard? Who enforces it?
The card numbers will keep coming. Spending will rise or fall. Charge-offs will improve or worsen. Deposits will grow or shrink. The quarterly ritual will continue, and the AI deployment will continue alongside it, mostly invisible, mostly unmeasured, mostly trusted because it has not yet failed in a way that shows up in the data. The boundary Dimon drew will hold until it does not. And when it moves, the reason will not be a change in the technology’s capabilities. It will be a change in the bank’s ability to verify what the technology is actually doing.
That is the open question. Not whether AI works — it does, at least well enough to deploy inside a bank. Not whether AI is dangerous — it is, at least dangerous enough to keep outside the walls. The question is whether the gap between those two facts can ever be closed. Whether a system can be productive enough to trust and transparent enough to verify at the same time. Dimon is not betting on it yet. He is using the technology and refusing to trust it. That is not a contradiction. It is the most honest position available in a moment when the tools are ahead of the rules, the productivity is ahead of the proof, and the only thing standing between a customer and a rogue agent is a CEO’s judgment about what he will not let through the door.