How JPMorgan Chase built AI at scale.

The short version.

JPMorgan Chase, the largest bank in the United States, has the strongest public record of AI capability of any bank. It has ranked first on the Evident AI Index in every edition, the fifth in October 2026.²⁶ It built its data and cloud foundations first, put a chief data and analytics officer on its Operating Committee in 2023, and gave most of its staff one controlled generative AI platform, LLM Suite, which it rolled out to more than 200,000 colleagues.¹² It has also published specific operating results: know-your-customer unit costs down nearly 40%, software engineers up to 10-20% more productive, and transaction screening at double the volume with half the manual checks.²³

Its headline value figure is a different kind of evidence. It moved from a $1 billion target in 2023 to "about $2 billion of benefit" for "$2 billion of expense" in 2025, given each time in an interview or at a conference.¹⁶⁵¹⁹ No method has been published, and the figure has not appeared in an annual report, a 10-K or an earnings call.²⁸ The executives who give it say it is hard to measure, and say so in public.⁵⁸

The bank paid about $450 million in 2024 and $200 million in 2022 in penalties for failures in trade-surveillance data and message recordkeeping.^10111213 None was an AI failure, but each concerned the data controls AI depends on. The case is useful to a board because it shows both halves of an AI programme at scale: what a well-run build looks like, and what is still missing when the value has to be shown to someone outside the bank.

Status as of 12 October 2026: third-quarter 2026 results are due in mid-October 2026. No method for the headline value figure has been published, and no external audit or academic test of it was found. Every time-sensitive claim on this page was checked on that date.

The reference case for AI in banking.

Boards and regulators use JPMorgan as the benchmark for AI in financial services. Its technology budget was about $18 billion in 2025 and is expected to be $19.8 billion in 2026.⁴⁷ When another bank's board asks what good looks like, this is usually the bank it means. That makes it worth reading closely: which parts of the record are measured, which are claims, and which parts a smaller organisation can copy.

The foundations came first.

The bank's AI work predates generative AI. COiN, a contract-review system that went live in 2016, was reported in 2017 to do in seconds a review of about 12,000 commercial loan agreements a year that had taken 360,000 hours of work.²⁹³⁰ The bank also moved its applications to modern infrastructure. By May 2025 about 65% of its applications ran a large part of their workloads on public or private cloud, 80% including virtual servers, and the finance chief said the bank was "past the point of peak modernization spend".⁴

It gave AI senior owners. In June 2023 Teresa Heitsenrether became the first Chief Data & Analytics Officer, and the 2023 letter says the role sits on the Operating Committee.¹⁷¹ The same letter counts "more than 2,000 AI/machine learning (ML) experts and data scientists".¹ And it set controls early. In February 2023, before it had its own platform, the bank restricted staff use of ChatGPT under its normal rules for third-party software, "not because of a particular issue".³¹

What the bank built.

LLM Suite is the centre of the generative AI programme. It gives staff access to models from more than one provider inside the bank's controlled environment. The bank launched it in 2024 and rolled it out to more than 200,000 colleagues, starting with Asset & Wealth Management.²³⁶ By September 2025 about 250,000 staff had access, about half used it near daily, and the models came from OpenAI and Anthropic, refreshed every eight weeks.²² It won American Banker's 2025 Innovation of the Year Grand Prize.²¹

On top of it sit tools for specific jobs: Coach AI for wealth advisers, GPT-Insights for investment research, an Employee Assistant agent, and AI coding assistants used by over 90% of engineers in the corporate and investment bank.³⁸²³ In February 2026 the bank said it had doubled the number of AI use cases in production in a year, and in June 2026 that it would deploy agents that run on their own for hours.⁷²³

What the bank did well.

1. Platform before use cases. Cloud migration and data work came before generative AI, so LLM Suite had somewhere governed to run.⁴²
2. One controlled platform for everyone. Staff got a single model-agnostic tool inside the bank's controls, which reached more than 200,000 colleagues in about eight months on an opt-in roll-out.²²¹
3. Senior ownership. The chief data and analytics officer and the chief information officer both sit on the Operating Committee.¹⁷
4. Talent at scale. More than 2,000 AI and machine learning specialists, and prompt training added to onboarding.¹⁴⁰
5. Specific published results. KYC unit costs, engineering productivity, screening volumes and fraud cost are named and quantified, which few banks do.²³⁴
6. Plain speaking about measurement. The head of consumer banking said much of the efficiency was "good old-fashioned process automation", and the finance chief declined to make teams prove AI savings because it could push people to use AI where it is not efficient.⁴⁸
7. Early controls and supplier pressure. ChatGPT was restricted in 2023, AI models fall under model risk management, and the chief information security officer wrote publicly to suppliers about the risks AI agents add.³¹⁹¹⁴

The fraud result is the one with most weight for a board. The head of consumer banking said in May 2025 that the attack rate had grown 12% a year while the cost of fraud in basis points stayed flat, and that "AI has been a very significant part of the solution".⁴ Like the other results, it is the bank's own figure and is not audited. Unlike most AI claims, it names the measure.

Where the record falls short.

1. A headline number without a method. The value figure has moved four times and lives in interviews and conference remarks. The annual reports give only growth rates, for consumer banking, with no base: up 35% in 2024 and nearly 60% in 2025.⁴³²⁸
2. Data-control failures in a regulated function. Regulators found billions of trading events on dozens of venues went unsurveilled for years. Not an AI failure, but a data-governance one.¹⁰¹¹
3. Workforce messages that are easy to misread. Executives have given a 10% operations reduction, "not fewer headcount" and 30% to 40% cuts in discrete areas, each with a different scope.⁴³²⁴
4. Pressure on staff to adopt. Business Insider reported that coding-assistant usage feeds engineers' performance goals; one developer described "a lot of anxiety".³⁹
5. Concentration and rising running costs. LLM Suite relies on two model providers, and no exit plan has been disclosed. The finance chief expects model usage costs to accelerate in the second half of 2026.²²²⁴²⁵

How to read the $2 billion figure.

The figure has a history. A $1 billion target for 2023 was raised to $1.5 billion during the year.¹⁶ At Investor Day in May 2024 the bank's president, Daniel Pinto, said: "difficult to measure, but roughly the value that we assign to our artificial intelligence use cases is around between $1 billion to $1.5 billion".⁵ By September 2024 he put it near $2 billion.¹⁸ In October 2025 Jamie Dimon told Bloomberg TV: "We have shown that for $2 billion of expense, we have about $2 billion of benefit."¹⁹²⁰

Three points matter for a reader. The $2 billion of expense is AI spend inside a technology budget of about $18 billion, not the budget itself.⁴ "Benefit" mixes revenue, cost avoidance and risk and fraud, so on Dimon's own framing benefit roughly matches spend; it is not $2 billion a year of savings.⁷ And the closest thing to a method came in February 2026, when an analyst asked for a scorecard. Dimon said:

"In AI, there are NPVs, some are revenue enhancements, some are cost avoidance, some are risk and fraud ... we can measure it, but we don't give it credit ... 150,000 people use it every week. They think they're saving four hours a [week]. That's not in an NPV."⁷

That is a defensible position: the bank counts project returns and refuses to count self-reported time saved. The chief financial officer, Jeremy Barnum, went further on the October 2025 call. Making teams prove AI savings "turns out to be a very hard thing to do, hard to prove", he said, and might leave "people scrambling around to use AI in ways that are actually not efficient". The bank relies instead on expense discipline and slower headcount growth.⁸ What the record does not contain is the breakdown that would let an outsider check the total.

What the bank has said about jobs.

Read closely, the bank's statements on jobs fit together. Read as headlines, they did not. The head of consumer banking, Marianne Lake, said in May 2025 that operations headcount, excluding Home Lending, should "trend down by about 10% over the next five years or so, even as the business grows by another more than 25%", and that she "would take the over".⁴ In the 2025 annual report the chief operating officer wrote that fewer jobs "will likely be the outcome in certain jobs" but "the goal is not fewer headcount", and Dimon wrote that AI "will definitely eliminate some jobs" and that the bank would have "definitive plans" to redeploy affected staff.³

The bank had more than 320,000 employees at the 2025 annual report.³ No firm-wide figure for AI-related reductions has been published. The press has reported the operations figure as "at least 10%" and the July 2026 remark as AI cutting 30% to 40% of jobs at the bank.²² Neither matches the scope the speakers gave.

The controls record.

The penalties below are not about AI. They concern the completeness and retention of data in regulated functions, which is what any AI system in those functions would draw on. The OCC found the bank surveilled trading "with gaps in trading venue coverage and without adequate data controls"; the Federal Reserve dated the gaps back to 2014.¹⁰¹¹

The bank's own filings treat AI as a risk to manage as well as a tool. The 10-K brings AI and machine learning models under model risk management and reports a higher volume and complexity of cyber attacks, including those using AI.⁹ Its chief information security officer, Patrick Opet, wrote to suppliers in April 2025 that software-as-a-service "is quietly enabling cyber attackers" and that AI agents amplify those risks.¹⁴ As of September 2025 generative AI did not yet talk directly to customers.²² That is prudent, and it also means there is little public evidence on customer outcomes.

DBS and JPMorgan count AI value differently.

DBS, the subject of an earlier case study, is the other bank most often cited for AI value. The two figures are measured differently and should not be ranked against each other.

DBS publishes its figure where shareholders read it and names its method. JPMorgan publishes more operating detail by function and states its figure outside its formal reports. A board can learn from both: name the method, put the number where it will be read with the accounts, and show the named results underneath it.

What is often misreported.

• JPMorgan is not saving $2 billion a year with AI. Dimon described about $2 billion of benefit for about $2 billion of expense.¹⁹⁴²
• The $2 billion investment is AI expense, not the technology budget of about $18 billion.⁴
• The operations figure is about 10% over five years or so, for consumer-bank operations excluding Home Lending.⁴
• The 30% to 40% cuts were in "discrete areas", with most people offered other roles.²⁴
• The 2024 fines were for trade-surveillance data, not AI.¹⁰
• First on Evident means most capable and most transparent on Evident's measures, not most value delivered.²⁶
• COiN's 360,000 hours is a bank-sourced figure for annual review work. No headcount effect was reported.³⁰

What boards can take from it.

1. Build the platform before the use cases. JPMorgan's scale came from cloud, data and one governed platform first. A board should ask what its AI runs on before it asks how many use cases are live.
2. Give AI an owner at the top table. A data and analytics chief on the executive committee made AI a line responsibility.
3. Publish named results with their measure. KYC unit cost and fraud cost against attack rate are claims a reader can understand and challenge. Prefer them to a single total.
4. If you give a total, give the method. State what counts as benefit, over what period, against what baseline, and put the figure where it will be read with the accounts.
5. Treat data controls as AI controls. The surveillance penalties came from incomplete data. Any AI built on that data inherits the gap.
6. Give every workforce statement its scope. Function, period and what happens to the people affected. Unscoped numbers become headlines.
7. Watch the running cost and the supplier list. Model usage costs rise with adoption, and two providers is a concentration. Ask for both in the board pack.

For government and regulators. The bank's best results came where a measure already existed: fraud cost, unit cost, screening volumes. Supervisors asking banks about AI value can ask for those measures, and for the method behind any total that appears in investor materials.

A checklist for an AI programme at scale.

To test a value claim of your own against these points, record it in the AI value ledger, and place a planned agent deployment on the agent risk assessment matrix.

Dates to re-check.

Full timeline.

Common questions about the case.