How JPMorgan put generative AI in front of ~200,000 employees

A trillion-dollar bank with one rule it wouldn't break: data stays inside

JPMorgan Chase couldn't simply tell employees to go use a public chatbot. As a heavily regulated bank handling some of the most sensitive data in the economy, sending internal information to an outside tool was a non-starter. But sitting out the generative-AI moment wasn't an option either, not for a firm that spends in the neighborhood of $17 billion a year on technology.

So instead of buying a single product, the bank built a layer of its own. “LLM Suite” is an internal front door that routes employees' requests to multiple third-party frontier models while keeping the data inside JPMorgan's environment. It reframed the question from “which vendor do we pick?” to “how do we give everyone safe access to the best models available?”

That decision — own the access layer, stay model-agnostic, keep the data in-house — is quietly becoming the default pattern for regulated enterprises that want frontier AI without surrendering control of their information.

From a few thousand seats to two hundred thousand

The rollout was deliberate rather than dramatic. LLM Suite went live in the summer of 2024; by that August, reports put usage around 60,000 employees, with plans to reach roughly 140,000. By the middle of 2025, about 200,000 employees had access, using it for drafting, summarizing dense documents, and generating ideas.

Crucially, it was optional-first and woven into existing workflows rather than dropped on people as a mandate, with the firm's Chief Data & Analytics Officer, Teresa Heitsenrether, owning the effort. Adoption was treated as something to be earned across the workforce, not assumed the moment a seat was assigned.

AI has the potential to really deliver amazing scale and efficiency as well as client benefit.

Teresa Heitsenrether, Chief Data & Analytics Officer, JPMorgan Chase — Bloomberg, 2025

Two hundred thousand seats is the start, not the finish

Handing out access is the easy part; getting people to actually use it is the real work. By public accounts, roughly half of the employees with LLM Suite use it on any given day — a strong number for an enterprise rollout, and a reminder that even at a sophisticated firm, “everyone has it” and “everyone uses it” are two different things.

Reported productivity gains land in the range of one to two hours saved per user per week. Modest per person, enormous across 200,000 people — but only realized to the extent that the habit sticks. That gap between assigned access and genuine, measured adoption is exactly where most enterprise AI programs quietly stall.

Coding assistants, fraud detection, and the move toward agents

The assistant is only the most visible piece. JPMorgan has pushed AI into software engineering with coding copilots, and deeper into fraud detection, risk, and marketing — the unglamorous, high-volume work where small efficiency gains scale into real money. Leadership has publicly framed the firm-wide value of AI at up to $2 billion, up from an earlier $1.5 billion estimate.

The next phase, by the bank's own signaling, is agentic: tools that don't just draft an answer but take a step — carrying out multi-stage tasks rather than waiting for a prompt each time.

AI as default infrastructure, not a one-off pilot

JPMorgan's trajectory is a tell for the whole enterprise market: AI is moving from scattered experiments to governed, firm-wide infrastructure — a single internal door to many frontier models, owned by a senior leader, measured by real usage rather than license counts.

The lesson underneath the headline numbers is the one Fautons keeps returning to: access is not adoption. The win wasn't buying 200,000 seats. It was the discipline of rolling them out, tracking who actually uses them, and treating the behavior — not the invoice — as the thing to manage.