JPMorgan LLM Suite: What the Bank Built and What It Changes

JPMorgan blocked ChatGPT in early 2023. Chief Risk Officer Ashley Bacon: "The first thing we did was to block it, while working in parallel to get an internally developed version into the hands of every employee to unleash its power." That internal version, LLM Suite, is a bank-controlled window onto outside AI models that now reaches roughly 250,000 employees.

For a candidate, the more useful fact is what it changes: JPMorgan's chief analytics officer describes junior staff moving from makers to checkers, verifying a draft LLM Suite already produced rather than building one from scratch.

LLM Suite at a glance

Summer 2024

Firmwide launch

after a 2023-24 alpha

~250,000

Employees with access

by September 2025, per JPMorgan

OpenAI + Anthropic

Models named by JPMorgan

as stated in September 2025

$19.8bn

2026 technology budget

the whole firm, not only AI

What LLM Suite means for the job

Makers become checkers

The clearest account came from Derek Waldron, JPMorgan's chief analytics officer, in September 2025: "Without a doubt, AI technology will have changes on the construction of the workforce. That is certain, but I think it's unclear as to exactly what those changes will look like." He described the shift underneath that uncertainty as workers moving from makers to checkers, with senior staff still managing client relationships while AI systems handle more of the work underneath them.

Before LLM Suite

• Draft the first version of a memo from a blank page
• Track down comparable filings and transcripts by hand
• Spend the night formatting a pitch deck

The maker to checker shift

• Start from a draft LLM Suite produces in seconds
• Verify the numbers and citations it already found
• Spend the night checking output instead of typing it

What JPMorgan has said about jobs

JPMorgan's chief executive, Jamie Dimon, has been the most direct voice at the bank on where this leads, across a run of earnings calls, conferences and interviews.

On that same July 2026 call, Dimon put a number on how much of the bank's AI activity actually matters: "I think there's almost 1,000 use cases today, though I would say that the really important ones are 50, probably 50 across risk, fraud, marketing, hedging, prospecting, note-taking, idea generation, document reading. And it's kind of just starting."

His own statement on jobs is narrower than the number that gets repeated most: discrete, unnamed areas of the bank cut jobs by 30 to 40 percent, and most of those people were offered other roles inside the firm. A separate, much larger number follows LLM Suite around, the claim that JPMorgan's AI plans could eliminate two-thirds of junior investment banking jobs; a companion guide on this site traces where that number actually comes from.

Why JPMorgan built its own ChatGPT

That shift did not happen by accident; it follows directly from why JPMorgan built the platform in the first place. Blocking ChatGPT solved one problem and created another. Employees wanted the productivity a chatbot offered; the bank could not let them run client and firm data through a tool it did not control.

JPMorgan appointed Teresa Heitsenrether chief data and analytics officer in June 2023 to lead the effort, and she put the reasoning in a single sentence: "Since data is our key differentiator, we had to ensure that our data was not being used to train external models. We also had to design a system to comply with still-evolving regulatory requirements across jurisdictions."

Two things had to be true before employees could get anything like ChatGPT back:

• Client and firm data could never be used to train an outside company's model.
• Whatever the bank built had to satisfy regulatory requirements that were still being written in real time, across every jurisdiction it operates in.

The route in was a March 2023 contract with Microsoft, which gave the bank access to large language models through Azure's cloud rather than through the open ChatGPT product. Chief Analytics Officer Derek Waldron called it a turning point: "This was a big deal: it was the first time using confidential data in a scaled way on shared computing infrastructure. We needed a really trusted partner."

Inside weeks of getting that access, a central intake portal had logged 1,000 ideas from employees for what they might do with it. Most of those ideas turned out to be variations on the same handful of tasks: summarizing a document, drafting a first pass at an email, or looking something up faster than a manual search would allow.

JPMorgan grouped the recurring patterns and built one tool to serve all of them instead of a separate app for each. That tool became LLM Suite.

That logic held as the platform scaled from an alpha to a firmwide rollout. JPMorgan does not break out a separate AI line inside its technology budget, but the whole number climbed every year LLM Suite existed: $17 billion in 2024, $18 billion in 2025, and a planned $19.8 billion in 2026.

2024

$17bn

2025

$18bn

2026 (planned)

$19.8bn

What is JPMorgan's LLM Suite?

Heitsenrether has given the clearest description of it herself: "It looks like a ChatGPT window, but behind that, we link to our data and retain the option to swap in and out new models based on their strengths. Think of it as a fund of funds model that we designed from the start to access the best tool for each task."

LLM Suite is not a large language model JPMorgan trained. It is a bank-controlled front door onto other companies' models, with the bank's own data, logging and permissions wrapped around it. That design choice was deliberate from the start. Waldron has said the bank built an "aggregation strategy to abstract ourselves from the underlying providers" precisely because nobody could tell in 2023 which AI lab would end up ahead.

Heitsenrether has made the same point about flexibility: "Ultimately, we'd like to be able to move pretty fluidly across models depending on the use cases." When JPMorgan told Decrypt in 2024 that LLM Suite "is not an independent large language model like ChatGPT," it described the product instead as an interface onto outside tools the bank does not own. That was not a hedge; it was the whole design.

A few features follow directly from that choice, and they explain why the platform grew the way it did:

• Every prompt and response is logged, because a regulated bank cannot let staff use an AI tool without a compliance record.
• The interface sits inside software employees already use, rather than shipping as a separate app.
• JPMorgan's own data scientists can build plug-ins on top of it and roll them out to everyone at once, so the platform grew from a chat window into what the bank calls a landing pad for internally built tools.

How many JPMorgan employees use LLM Suite?

LLM Suite did not appear firmwide overnight, and the different numbers JPMorgan has given since 2024 measure different things: access, onboarding, and daily use are three separate counts that get flattened into one headline every time a new figure appears. Here is the ladder as the bank has stated it, oldest first.

The same adoption curve every time

Waldron has separately described the adoption pattern itself, independent of any one date. Out of every group given access, JPMorgan saw the same split regardless of how large the group was: about 30% became active users straight away, another 10% became heavy users, and the rest were slow to start. Once a platform reached full scale, that 30% active-user share tended to rise to roughly 50%.

Which AI models does LLM Suite use?

JPMorgan has changed what it says about the models behind LLM Suite as the platform matured, which is consistent with a bank that built the whole thing to be model-agnostic rather than locked to one provider.

That same 2026 posting for the "LLM Suite Engineering" team also named agent-to-agent communication, the Model Context Protocol, and agentic orchestrators among the skills it wanted, which points at where the platform is headed next: from a chat window toward systems that can carry out multi-step tasks on their own rather than answering one prompt at a time.

What LLM Suite actually does

Strip away the branding and LLM Suite does what most workplace AI assistants do, just with a bank's data and controls wrapped around it. The Harvard Business School case on the rollout describes "typical users" as people searching, generating, and summarizing text, or asking about firm policies, market products, and legal and regulatory specifics.

Credentialed staff can point it at internal research, earnings transcripts, news, and planning documents. Over time the bank added PowerPoint generation and Excel analysis modules on top of the same base assistant, so the same tool that answers a policy question can also build the slide that presents it. Put those documented pieces together and one illustration looks like this:

1. Ask LLM Suite to summarize an earnings call transcript.
2. Carry the key figures into an Excel model built on the bank's own templates, checking each number against the transcript rather than retyping it.
3. Let the same tool draft the slide that presents the result, using the PowerPoint module built on top of the base assistant.

The clearest illustration JPMorgan has actually confirmed came from a demonstration Waldron gave CNBC in 2025, watched by an outside journalist for the first time: the platform built a credible-looking investment banking presentation in about 30 seconds, the kind of task that used to keep a team of junior bankers at their desks late into the night.

Three things distinguish LLM Suite from a plain chatbot once it is running at this scale:

1. It works across roles rather than needing a separate tool built for every job. Waldron has said it can "service" traders, wealth managers and risk officers alike from one platform.
2. JPMorgan's own staff build plug-ins on top of it and ship them to everyone at once, so new capability spreads without a separate rollout each time.
3. The bank has already started layering agentic workflows on top: multi-step tasks the system carries out rather than a single question it answers.

Connect Coach, IndexGPT and the rest of the family

LLM Suite is the firmwide platform, but it is not the only AI system JPMorgan has built, and mixing them up is an easy mistake to make. Reuters has referred to the Private Bank's advisor tool as JPMorgan's "so-called Coach AI tool," but the product's actual name, used by JPMorgan's own executives, is Connect Coach.

How Connect Coach has grown

Connect Coach's growth mirrors LLM Suite's, on a smaller scale. It launched to 3,000 advisors in October 2024 and, per the 2025 annual report, now serves 12,000 users across the Private Bank and Wealth Management with 25 specialized agents built into it. Advisers using it find information up to 95% faster, according to the executive who oversaw the rollout.

Its hyperlinked citations, letting an advisor trace a suggestion back to its source, put the same checking habit to work on a different desk.

What a candidate should know

Three things separate a candidate who has actually thought about LLM Suite from one repeating a headline:

1. Know the family, not just the headline tool. IndexGPT is a client product sold through Markets, not an employee tool, and mixing it up with LLM Suite in an answer is the fastest way to look like the research stopped at the name.
2. "Checker" work has a specific shape. It means tracing a number back to its source and applying house judgment, the same discipline Connect Coach makes visible when it shows an advisor which page a suggestion came from.
3. Only JPMorgan staff can actually use LLM Suite. It runs inside the bank's own systems, so the only verified way to talk about it in an interview is by citing what the bank has said publicly, not by claiming hands-on experience.

For engineers: the skills LLM Suite hires for

Candidates aiming for a technical role rather than an analyst seat get a more concrete map. JPMorgan's own engineering postings for the "LLM Suite Engineering" team, part of its Corporate Technology organization, name a specific set of skills rather than a generic AI background:

• Building GenAI services on Azure OpenAI models and AWS Bedrock
• Agent-to-agent communication between AI systems
• The Model Context Protocol, a standard for connecting a model to outside tools and data
• Designing agentic orchestrators that coordinate multi-step, autonomous workflows
• Building AI and machine learning solutions on public cloud infrastructure at bank scale

That list reads less like a chatbot company and more like an infrastructure team, which is closer to what LLM Suite Engineering actually is: a group building the plumbing that lets a regulated bank route millions of employee prompts to outside models without the data ever leaving its own control.

The bottom line

JPMorgan's answer to ChatGPT was never really about the chatbot. It was about who controls the data, which models sit behind the interface, and who inside the bank gets to use it. LLM Suite settled all three in the bank's own favor: a wrapper it controls, model-agnostic by design, now reaching nearly the whole workforce. Understanding that decision, rather than the chat window, is what makes someone sound informed about AI at JPMorgan.