Ask most bank executives about artificial intelligence in 2024 and they would describe a chatbot. Staff typed in a question, the software produced a paragraph, and somebody in the technology division counted it as adoption.
The systems now being installed across the largest American banks work differently. They are handed an objective rather than a question. They break it into steps, open whichever internal systems those steps require, pull the data, carry out the task, and report back when it is finished. The industry calls this agentic AI. In practice, it means software that is given a job rather than asked for an opinion.
Whether this is a genuine break with what came before, or the same technology with more permissions attached, is a fair question. What is not in doubt is that the banks are spending as though it is the former.
What JPMorgan has built
JPMorgan Chase has been unusually open about its programme, which makes it the easiest place to see the shift.
Its platform is called LLM Suite. It started in 2023 as a secure internal version of ChatGPT for drafting emails and summarising documents. Around 250,000 employees now have access to it, and roughly half use it on any given working day. The platform is model agnostic, meaning it routes work to systems built by outside AI firms, including OpenAI and Anthropic, rather than depending on a single supplier. It is updated roughly every eight weeks as the bank wires it into more of its own databases and applications.
That wiring is the slow part of the project and the part that determines whether any of it works. A capable AI model that cannot reach a bank’s payment records, client files or risk systems produces confident text, and nothing else. The value only appears once the model can see the underlying data and act on what it finds.
Derek Waldron, the bank’s chief analytics officer, has described the goal to CNBC as a ‘fully AI-connected enterprise’, with an assistant for every employee, agents running back-office processes and AI shaping client interactions. In a demonstration for the broadcaster, the system assembled an investment banking pitch deck, including news, earnings figures and peer comparisons, in about thirty seconds. That is work which previously occupied a team of junior bankers for most of a night.
The bank has been automating document work for longer than the current wave suggests. A tool called COiN, short for Contract Intelligence, was reported by Bloomberg in 2017 to have taken over the review of commercial loan agreements that previously consumed around 360,000 hours a year of work by lawyers and loan officers. That was conventional machine learning applied to one narrow, high-volume document type rather than a large language model. What has changed is the range of work now in scope. The bank is training systems to draft confidential merger memoranda, which bankers then check.
The return on investment is still an open question
JPMorgan spends around $2 billion a year specifically on AI. That sits within a technology budget the bank raised to about $19.8 billion for 2026, an increase of roughly 10% on the previous year. Jamie Dimon has said the AI investment has already paid for itself.
The claim is harder to pin down than it first appears, and Dimon has said so himself. Pressed by analysts at the same investor event, he replied that ‘the hardest thing to measure has always been tech projects’, adding that time saved is often too vague to quantify. Both statements were made in the same session. He believes the money is coming back, and he cannot demonstrate it line by line.
Chief financial officer Jeremy Barnum was blunter about the cost side, telling investors that technology remains a major driver of the bank’s expense growth.
Paying for itself would in any case mean breaking even, roughly two billion dollars of benefit set against two billion dollars of cost. That is a modest outcome for a technology being described as transformational. It looks impressive mainly by comparison, because a large majority of companies attempting generative AI projects cannot show any measurable financial effect at all.
There is a further difficulty. The savings that are easiest to evidence come from the least exciting work, meaning contract review, code generation and summarisation. Those are real hours removed from real cost centres. The claims attached to client-facing AI, personalisation, better service and deeper relationships, are the ones the bank is least able to price. Whether the second category ever produces returns of the kind the spending assumes is not yet established, by JPMorgan or anyone else.
What the bank has done with its accounting is more revealing than the figures. It has moved AI spending out of the discretionary innovation category, the line that is cut first when results disappoint, and placed it alongside data centres, payment systems and core risk controls. That reclassification commits the bank to the spending regardless of what the next few years of returns look like.
Morgan Stanley is letting outside agents in
Morgan Stanley is doing something structurally different, and it has attracted less attention than it probably deserves.
The bank runs two platforms, ShareWorks and Equity Edge, which administer employee share schemes for around 3,400 corporate clients. The work involves grants, vesting schedules, option exercises, tax handling and compliance reporting. It matters commercially because administering a company’s share plan puts Morgan Stanley in front of employees before those employees become wealthy. Executives have credited the workplace strategy with gathering some $1.2 trillion in assets.
The bank is now opening those platforms so that clients’ own AI agents can connect directly, bypassing the interfaces built for human administrators. Mark Mitchell, chief product officer of Morgan Stanley at Work, told CNBC that in future corporate clients ‘will not be logging into ShareWorks or Equity Edge’ at all, and will instead run agentic tools inside their own companies that interact with the bank’s platforms directly. He has argued that the firms which survive will be the ones holding proprietary data and business logic, which he describes as the foundation of what Morgan Stanley sells. A small number of clients have early access, with a wider rollout planned across the client base.
The technology behind this is the Model Context Protocol, or MCP. It is worth being precise here, because it is often described as an investment in AI infrastructure. MCP is not a Morgan Stanley product, and the bank has not funded its development. It is an open standard, created by Anthropic in late 2024, and handed to the Linux Foundation at the end of 2025, where it is now governed by an industry body with backing from Google, Microsoft, Amazon and OpenAI.
Think of it as a USB-C port for AI systems. Before it existed, every connection between an AI model and a piece of business software needed its own custom integration, which is a large part of why so many corporate AI projects stalled between the demonstration and the deployment. MCP standardises the socket. Build the connection once, and any compatible system can use it.
The decision to adopt a shared standard rather than build a proprietary one tells you what Morgan Stanley expects. It is planning for clients who arrive already running their own AI, and it would rather be the system those agents connect to than the one they work around. Whether corporate clients actually take this up at scale is untested. The rollout is still ahead of the bank, not behind it.
Headcount is already moving
The commercial argument being made to clients is about hiring. Fast-growing technology and biotechnology companies want to run increasingly complex share schemes without adding administrators. Internally, Mitchell has said Morgan Stanley sees agentic AI as a way to scale customer support, plan administration, and the wealth management funnel without adding thousands of employees.
Across the sector, the numbers have started to reflect that. In the first quarter of 2026, six of the largest American banks shed around 15,000 jobs while posting roughly $47 billion in combined profit, up sharply on the previous year.
The rhetoric has shifted just as fast. Bank of America chief executive Brian Moynihan, asked on television what he would tell his employees about AI replacing human work, said they did not have to worry. Within months. his bank was crediting AI for reductions achieved through attrition.
Dimon has been more direct, saying AI will eliminate jobs and that ‘people should stop sticking their heads in the sand’, while pointing to attrition, retraining and redeployment as alternatives to mass redundancy.
Citigroup chief executive Jane Fraser has told staff that some roles will change, some will emerge, and others will no longer be required, as part of a turnaround that includes cutting 20,000 posts. Among posts cut were employees from Citi’s own internal programme dedicated to persuading colleagues to adopt AI.
How much of this is genuinely attributable to AI is contested. Banks reduce headcount in tight years and have done so for decades, and AI is a convenient explanation for decisions that might have been taken anyway. The clearer signal is which roles are exposed.
Work that involves moving structured information between systems is the easiest to redesign around agents, and that covers much of the back office along with a growing share of the analyst work that has traditionally been the way into the industry.
JPMorgan’s consumer banking head has told investors that operations staff will fall by at least 10% over five years.
Trading, and the regulators arriving early
The claim that machines will soon trade rather than advise turns out to be the part regulators are least worried about today, and most worried about tomorrow.
In a speech titled Agents of change, delivered at the European Central Bank’s Sintra forum on June 30, Bank of England deputy governor Sarah Breeden said the evidence suggests trading firms currently use autonomous AI mostly for lower-risk operational work, such as research, while warning that this could change quickly. Her more immediate concern was not markets at all. It was cyber security, which she described as her most proximate financial stability worry, driven by a steep change in what agentic systems can do when hunting software vulnerabilities. She cited the heads of the Five Eyes cyber agencies, who said in June that the relevant timeline is months rather than years.
On trading, her argument was about correlation. If agents respond similarly to the same prompts or triggers, they could amplify volatility under stress, particularly if their objectives drift from what their designers or public policy intended. The financial stability question, she said, is no longer only whether individual firms use models well, but whether the system can observe and contain how those models behave together.
That is why the kill switch idea has been raised, and it is worth reporting accurately. Breeden did not propose one. She set out a research agenda, including work with the Bank for International Settlements innovation hub and the Bundesbank on which features of agent design drive herding, and asked whether guardrails analogous to circuit breakers or market-wide kill switches might eventually be needed.
Her broader conclusion was blunt. “Our frameworks were not built to contemplate autonomous agents,” she said, adding that relying on a human approving every agent action is unlikely to be realistic.
Markets absorb shocks because participants disagree. If several large institutions run agents built on the same small set of underlying models, trained on similar data and pointed at similar objectives, that disagreement could thin out at the moment it is most needed. The 2010 flash crash removed around a trillion dollars of value in five minutes, and the algorithms involved were considerably simpler and less autonomous than what is being deployed now.
The Bank’s Financial Policy Committee published its updated assessment on July 7. Rapid advances in frontier AI, the committee concluded, have increased financial stability risks connected to cyber and operational resilience. That is a firmer position than it held in April, and the Treasury Committee questioned the Governor on it a week later.
Internationally, the machinery is already turning. The Financial Stability Board (FSB), chaired by Bank of England governor Andrew Bailey, published a consultation in June setting out twelve sound practices for responsible AI adoption, covering governance, the stages of AI development and deployment, and cyber and third-party risk. Comments closed on July 22, and a final report is due in October as a G20 deliverable.
Michelle Bowman, who chairs the FSB committee behind the work and serves as vice-chair for supervision at the US Federal Reserve, said the report reflected collaboration on an accelerated timeframe to keep pace with AI. The board was careful to add that the practices are not an international standard and were not designed for the frontier model risks that have emerged most recently.
What remains unresolved
The argument about whether banks will use AI is settled. The open question is what happens when agents stop being tools inside a single institution and start operating across the boundaries between them.
Morgan Stanley opening its systems to external agents is the first real test of that. Once a client’s AI can act inside a bank’s infrastructure, the line between the two organisations becomes a question of permissions rather than architecture. Nobody has yet had to supervise that arrangement under stress.