The market briefing arrives before the coffee. Every morning, bankers in JPMorgan Chase’s private bank open dashboards assembled overnight: AI agents that spent the dark hours screening earnings releases, rate decisions and client-adjacent news, and left a prioritized summary on each desk. The agents ran for one to two hours without human help, up from two or three minutes a year ago.
The jump is the detail JPMorgan’s chief analytics officer shared with employees this week, and it is the number the rest of Wall Street is now studying. One to two hours of continuous autonomous work crosses what the bank calls the enterprise bar: the point at which an agent is reliable enough to be trusted with real workflows rather than demo scripts.
The officer said the bank expects 2026 to be the year agentic AI moves from experiments into core operations, and the private bank is the proof of concept. Agents screen information overnight so that bankers start their mornings with intelligence instead of raw feeds, and the results are measurable: total sales up about 20%, and client coverage expected to expand 50% as agents handle the screening and ranking work that once consumed junior staff.
Chief Executive Jamie Dimon has been blunt about the trade. Some employees will be replaced by AI, he has said, and the bank is building retraining programs for the roles that survive. The message, delivered repeatedly over the past year, has become a recruiting document as much as a warning: JPMorgan intends to do the automating, not be automated.
The scale of the effort is visible in the budget. JPMorgan spends close to $20 billion a year on technology, a figure that rivals the annual revenue of many banks, and a growing slice of it is going to the models, data pipelines and evaluation systems that keep agents on task.
How the agents work says a lot about why they improved. The bank uses a mix of its own models and vendor systems, wrapped in guardrails: every agent action is logged, high-value actions require human approval, and outputs are checked against sources. The overnight screening agents were deliberately given narrow jobs, and the reliability gains came from scope discipline as much as from better models.
The technical change is a matter of engineering rather than a single breakthrough. A year ago, agents lost context after minutes and drifted off task; the bank added retrieval systems, memory layers and automated evaluation loops that check an agent’s work as it goes. The result is that a task that used to end in a mess now ends in a summary a banker can act on.
The human side of the equation is changing faster than the technology. Bankers in the private bank now spend their mornings reading agent-produced briefings and their days calling clients, a mix that puts a premium on relationship skills and devalues the spreadsheet work that used to occupy the first hours of every day. The entry-level analyst role is being redefined around reviewing and escalating machine output rather than producing it.
Rivals are running similar programs. Goldman Sachs and Morgan Stanley have described comparable experiments with agents that summarize research and draft client communications, and executives at all three firms have said the constraint is trust, not capability. The difference at JPMorgan is scale: with $20 billion a year and a chief executive who treats AI as a strategic weapon, the bank can afford to fail at experiments other firms cannot justify.
The overnight workflow is the most visible change. The private bank’s agents now monitor client portfolios, flag news that touches a holding, and draft talking points before markets open, and the coverage expansion to 50% comes from agents maintaining relationships with smaller clients that would not justify a dedicated banker. The bank says the economics of the expansion only work because the agents do the monitoring.
The risks are the same ones every bank faces. An agent that misreads a filing and flags the wrong stock is a compliance problem; an agent that acts on its own in a payment system is a catastrophe. JPMorgan’s answer is layered approval: agents recommend, humans approve, and the audit trail records every step. The overnight agents are deliberately constrained to information work, where a bad summary costs time rather than money.
Dimon’s bet is that the bank that automates its own middle office wins the next decade, and the numbers he is citing suggest the bet is working: sales up a fifth, coverage up by half, and the headcount question deferred into retraining programs that the bank says are already moving staff into new roles. The market opens, the summary is ready, and the banker, for now, is still the one who calls the client.
