How JPMorgan Is Rolling Out AI Agents to a 318,000-Person Bank Without Breaking It
JPMorgan put an AI assistant in front of 250,000 employees, attributed $2 billion of annual value to AI, and only then started deploying multi-step agents. The sequencing is deliberate, and it makes this the most documented agentic AI programme in banking.
The starting point
JPMorgan Chase is the largest bank in the United States: 318,512 employees at the end of 2025, a technology budget of $19.8 billion (around 10% of revenue), and more than 65,000 technologists. It has been ranked first in the Evident AI Index, the independent benchmark of AI capability in banking, for four consecutive years.
The bank's AI story did not start with ChatGPT. Its COiN contract intelligence platform went live in 2016, automating commercial loan agreement reviews previously consuming 360,000 hours of lawyer and loan officer time each year. By April 2024, Jamie Dimon's annual letter reported more than 2,000 AI and machine learning specialists and over 400 use cases in production across marketing, fraud, and risk.
Then came the part most companies got wrong. In February 2023, JPMorgan restricted employee access to ChatGPT. Not because something went wrong, but because a bank handling client data under heavy regulation had no way to control where those conversations went. Rather than ban generative AI and move on, the bank spent the next eighteen months building its own front door to it.
What they built
LLM Suite: a gateway, not a chatbot
LLM Suite is a proprietary portal sitting between employees and the frontier model providers. Models from OpenAI and Anthropic are swapped in and out behind it, tested for safety before deployment, with proprietary data kept out of public models. The platform is updated every eight weeks, each cycle connecting more internal data and applications.
The rollout was staged and measured. Around 50,000 employees had access by July 2024, positioned internally as "a research analyst that can offer information, solutions, and advice on a topic". By May 2025 there were 200,000 active users. By late 2025, roughly 250,000 employees had access, everyone except branch and call centre staff, with about half using it daily.
Alongside the general-purpose assistant sit specialist tools: Connect Coach for advisors, SpectrumGPT for portfolio managers (turning analysts from "content creators into content editors", in the words of the bank's data science lead), and an internal coding assistant used by more than 40,000 engineers.
The first production agents
The agentic layer came after, and on top of, the assistant layer. Ask David, a multi-agent investment research system in the Private Bank, uses a supervisor agent orchestrating specialist agents for structured data queries, document retrieval, and proprietary analytics, with a human in the loop for decisions. The bank's engineers presented it publicly at LangChain's Interrupt conference in May 2025, and were candid about its limits and the need for human oversight in high-stakes decisions.
In September 2025, chief analytics officer Derek Waldron laid out the destination: "Every employee will have their own personalized AI assistant; every process is powered by AI agents, and every client experience has an AI concierge." By June 2026 the bank confirmed agents were evolving from single tasks taking two to three minutes towards "digital employees" managing workflows across multiple systems for one to two hours autonomously.
Governance built for agents, not retrofitted
The detail here is the part worth studying. Agents at JPMorgan get their own identity and scoped credentials, set per function. HR agents get narrower data access than the humans they support. Software engineering agents get more latitude because a validation layer catches their errors. Chief data and analytics officer Teresa Heitsenrether puts the design problem plainly: "you have to permission them and make sure you have the right guardrails around what they can and can't do, and how many autonomous decisions they can make." Her sharpest observation from production experience: "When it's an agent, they don't work well with ambiguity."
In February 2026 the bank reorganised its Commercial and Investment Bank around this agenda, appointing a COO to lead AI strategy and giving each major business its own chief data and analytics officer, explicitly "preparing our infrastructure for more advanced AI and the expanded use of AI agents".
The results
| Metric | Detail | Source |
|---|---|---|
| Adoption | ~250,000 employees with LLM Suite access, ~50% daily use | CNBC (Sept 2025) |
| Time savings | 1-2 hours per week average; several hours for heavy users; 2-4 hours per day for some analysts | American Banker, Quartz |
| Attributed value | ~$2 billion annually, fraud prevention the largest component; 30-40% annual growth in AI-attributed value since 2020 | CIO Dive, American Banker |
| Engineering productivity | 10-20% efficiency gain across tens of thousands of engineers | Reuters (Mar 2025) |
| Use cases | ~1,000 AI applications running, 50-60 described as significant | Banking Dive (May 2026) |
| Workforce composition | Over 2025: operations staff -4%, support staff -2%, client-facing roles +4%, headcount flat overall | CNBC (Feb 2026) |
| Operations efficiency | Accounts handled per operations employee +6%; per-unit fraud handling cost -11% | CNBC (Feb 2026) |
| Wealth management | AI credited with contributing to ~20% revenue growth in the wealth division | CNBC (June 2026) |
One demonstration from the September 2025 briefing captures the direction: LLM Suite generated a credible five-page investment banking pitch deck in around 30 seconds. Work of hours for a junior banker, gone in the time it takes to pour a coffee.
What makes this case interesting
The sequencing is the strategy. Assistants first, agents second, customer-facing AI last. Two years of 250,000 people using an assistant taught the bank where the model fails, which processes have clean data, and which staff lean in. Only then did it point autonomy at multi-step work. Compare Klarna, which went agent-first in customer service and had to walk it back. The boring order of operations turns out to be the fast one.
Agent governance is treated as infrastructure, not policy. Scoped credentials per agent, tiered data access by function, validation layers where the cost of error is recoverable. Heitsenrether admits the full machine-to-machine governance model "doesn't exist yet". For me, this honesty is more useful than any vendor's assurance: the biggest AI programme in banking says agent governance is still being invented.
The workforce impact is documented, not hidden. Operations down 4%, client-facing up 4%, total headcount flat. The bank's CFO for consumer banking put a 10% operations headcount reduction on record as a conservative projection. Dimon says it directly: "we have displaced people from AI, and we offer them other jobs." Most companies publish nothing on this. JPMorgan publishes the composition shift each year.
Value attribution comes with receipts and caveats. The $2 billion figure is the bank's own, and its executives say formal ROI "remains difficult to quantify". A company this sophisticated, four years running the top of the Evident Index, still finds AI value measurement hard. Budget for the same difficulty in your own business case.
The challenges
The bank's own leaders are the best source on friction. Waldron concedes there is "a value gap between what the technology is capable of and the ability to fully capture that within an enterprise", and says realising the vision "will take years" because work happens in thousands of applications needing connection into an AI ecosystem. Global CIO Lori Beer calls validation and controls "the harder part", including teaching regulators how the models are built and governed. On agents specifically: "You don't want them to go outside the bounds of the specific tasks that they can do, because they don't have the same thinking a human does."
Hallucination remains a live constraint in a regulated environment, which is why customer-facing deployment is deliberately last in the sequence and why the assistant-era tools kept a human between the model and every consequential output. And the redeployment story has tension in it: retraining and internal moves are the stated policy, but the operations projection of a 10% headcount reduction sits on the record alongside it.
Beer's summary of the whole programme is the line I would put in front of any board: "The change management and how you think about the ways of working is ultimately the hardest part here."
Lessons for your programme
Earn the right to deploy agents with an assistant phase. JPMorgan ran two years of supervised, human-in-the-loop AI at scale before granting autonomy, and used the assistant data to pick where agents go first. Map your own sequence in Section 04: Opportunity Identification before anyone builds an agent.
Design agent permissions like employee permissions. Scoped credentials, function-level data access, validation layers where errors are cheap. If your governance framework has nothing to say about non-human identities, it is not ready for agents. Section 08: Risk and Governance covers the classification and control work this requires.
Publish the workforce numbers internally before the rumour mill writes them for you. The composition shift (-4% operations, +4% client-facing) is a more honest and more useful message than either "no jobs will change" or silence. Section 15: Designing for Transformation includes the workforce impact modelling to get ahead of this.
Treat value attribution as a discipline, not a press release. JPMorgan tracks AI-attributed value annually, publishes the number, and admits its limits. Section 13: Business Case Production sets out how to build benefit tracking with the same honesty.
Sources
- JPMorgan Chase is going all in on AI: Here's how its private bank is using it (CNBC, 2025)
- JPMorgan plans more powerful AI agents this year (CNBC, 2026)
- JPMorgan CEO Jamie Dimon says AI is reshaping the workforce (CNBC, 2026)
- How JPMorganChase democratized employee access to gen AI (American Banker, 2025)
- The Most Powerful Women in Banking: Teresa Heitsenrether (American Banker, 2025)
- JPMorgan CIO on reshaping the bank with a $19.8 billion budget (Fortune via Yahoo Finance, 2026)
- Jamie Dimon's 2023 letter to shareholders (JPMorganChase, 2024)
- JPMorgan engineers' efficiency jumps as much as 20% from using coding assistant (Reuters via US News, 2025)
- JPMorgan to resist headcount growth as AI investment rises (Banking Dive, 2025)
- Ask David: JPMorgan's multi-agent investment research system, presented at LangChain Interrupt (LangChain Interrupt conference, 2025)