Major Wall Street banks are moving aggressively into agentic AI – those systems that complete tasks with little or no human prompting. AI agents are being deployed inside wealth management, trading, and internal operations, making banking a revealing test case. Banks operate in one of the most heavily regulated corners of the economy, so they have to answer the hard governance questions earlier than most industries will. Questions such as: For whom is the agent allowed to act? What systems can it access? What data can it see? Which actions require approval? What logs must be kept? Who is responsible when the agent is wrong?
Those questions reach well beyond finance.
Reuters reported this week that in a bid to win the productivity race, banks including Morgan Stanley, UBS, BNY, Goldman Sachs, JPMorgan, and Citi are testing or deploying AI assistants across wealth management, client onboarding, trading, treasury, and internal operations. Reuters also cited a KPMG survey finding that 51% of banks are piloting AI agents.
That matters because banks are rarely the first place one looks for reckless experimentation. They handle money, personal data, investment advice, trading activity, compliance obligations, and customer trust. If banks are beginning to put AI agents into real workflows, the technology is moving beyond demos and into the cautious middle of white-collar work.
Getting paid twice
There's another reason Wall Street is worth watching: The banks are using AI inside their own businesses while also making money from the AI economy around them.
A rush by technology companies to fund AI infrastructure is boosting dealmaking and financing activity for Wall Street, generating fees from capital raising and loans. Goldman Sachs CEO David Solomon described the AI infrastructure buildout as a "multi-year investment cycle," while Reuters noted major AI-related financing and capital-markets activity, including SK Hynix's $26.5 billion ADR offering and SpaceX's $86 billion IPO.
The Wall Street Journal reported that Wall Street traders are having what could become their best year ever, with major banks benefiting from heavy trading in AI-related shares, short-dated options, ETFs, and broader bullish market activity.
AI is showing up on both sides of Wall Street's business. Banks are helping finance, trade, advise, and profit from the companies and investors driving the AI boom. At the same time, they're trying to use AI to reduce friction inside their own operations. Wall Street is playing several roles at once: intermediary, investor, beneficiary, customer, and user of the technology reshaping the economy.
The phrase "AI agent" can be slippery. In the simplest terms, it means an AI system that can take steps toward a goal, instead of merely responding to a single prompt. A chatbot may answer a question, but an agent may gather information, update a system, prepare a memo, flag a compliance issue, schedule a follow-up, or route a task to the right person.
It's been busy inside the banks
That difference is why the banking examples are worth watching. According to Reuters, Morgan Stanley is preparing to test digital assistants that can interact with clients at any time of day, while helping advisors with reminders and recommendations. UBS says AI is helping advisors spend more time with clients and less time on routine tasks.
Additionally, BNY has treated digital employees as teammates – complete with login IDs, nicknames, assigned tasks, and human managers. Citi is preparing to roll out an AI-enabled virtual wealth management assistant, while Goldman Sachs and JPMorgan are developing agentic AI tools for functions such as trading, client vetting, onboarding, and treasury.
What happens to the humans?
The workforce implications are becoming harder to avoid. According to Bloomberg, JPMorgan CEO Jamie Dimon said the bank would likely hire more AI specialists and fewer traditional bankers in some categories as AI adoption accelerates. Bloomberg also reported that Goldman Sachs President John Waldron described the bank as a "human assembly line" facing automation, with AI helping the bank scale without requiring much more hiring.
Human bankers, advisors, traders, and operations staff are still central to the work, at least for now. The most interesting part of the Reuters report may be the emphasis on human oversight – banks are trying to identify which roles are hybrid, which tasks can be handled by agentic systems, and which work should remain human-led.
That's likely to be the pattern in many industries: a redistribution of tasks instead of a clean replacement of workers.
The productivity pitch is obvious. Banks have enormous amounts of routine, repetitive, documentation-heavy work. They also have large internal knowledge bases, complex customer histories, compliance requirements, and high-value employees whose time is expensive. A useful AI agent that saves even a few hours per week across thousands of workers could be worth a great deal.
Where it can go wrong
The risks are equally obvious. An AI agent that drafts an email is one thing. An AI agent that moves money, suggests trades, touches customer accounts, summarizes compliance information, or acts on behalf of an advisor is something else. Errors, bias, hallucinations, unauthorized access, privacy failures, and unclear accountability all matter more when the system is operating inside financial infrastructure.
The caution flag isn’t merely theoretical. In April, researchers from Handshake AI, McGill University, and other institutions introduced BankerToolBench, an open-source evaluation suite that tested frontier AI agents on end-to-end investment-banking workflows developed with input from more than 500 bankers. The researchers found that even the best-performing model they tested — OpenAI’s GPT-5.4 — scored just 58.1% on average, and bankers rated none of the tested models’ outputs as client-ready.
The result is a useful reality check. These were not toy tasks or simple chatbot answers. The benchmark asked AI agents to navigate data rooms, use market-data and SEC-filing tools, and produce multi-file deliverables such as Excel financial models, PowerPoint pitch decks, and PDF or Word reports.
In other words, it tested the kind of messy, multi-step work that AI agents are supposed to make easier. That is why the banks’ cautious approach matters. Agentic AI may be ready to assist with serious work, but BankerToolBench suggests it is not yet ready to own that work.
Colorado draws the line
Colorado is already answering a version of "who's responsible when the agent is wrong" in law, and not just in theory. The state's original AI Act, Senate Bill 24-205, was signed in 2024 and had its effective date pushed back once, to June 30, 2026.
Then, just a month before that deadline, lawmakers repealed and replaced it with a narrower bill, SB 26-189, which pushes the effective date out again, to January 1, 2027. Attorney Tyler Thompson, who walked through the rewrite in an interview with Colorado AI News, described the result as a "path forward" rather than a retreat – a law that's narrower than its predecessor but on firmer footing.
The detail that matters most for banking: The original law had a murky exemption for financial institutions already subject to similar regulatory oversight. SB 26-189 eliminates that exemption. For banks and credit unions operating in Colorado, the law's scope now lands squarely on lending and pricing decisions — precisely the areas where AI pilots are being pushed into production right now.
Once the law takes effect, deployers will have to notify consumers when a covered AI system materially influences a decision about them, explain adverse outcomes within 30 days, and offer a path to meaningful human review. Enforcement runs through the Colorado Attorney General, not private lawsuits, with a 60-day cure period before penalties attach for most violations. The AG is also still expected to issue rules clarifying exactly how the disclosure and review requirements apply – including, potentially, sector-specific guidance for banks – before the law takes effect.
This is an indication that banks aren't the only ones setting these boundaries. Colorado's lawmakers are now doing it, too, through law rather than internal policy.
What's here today, and what's ahead
For now, the safer use cases are likely to be assistant-like: summarizing information, preparing drafts, surfacing reminders, checking documents, helping onboard clients, and reducing the administrative burden on human workers.
In the big picture, the boundary will keep moving. The more capable agents become, the more businesses will be tempted to let them act directly – and the more directly they act, the more governance matters: permissions, audits, identity, accountability, and the right to have a human in the loop when decisions matter.
That's the harder question this story raises: whether an organization can trust the AI agent enough to let it act – and whether customers, regulators, and employees will be able to understand afterward exactly what happened. Hospitals, law firms, universities, government agencies, manufacturers, and small businesses will all face versions of this as AI agents move from pilot to daily use.
The banks appear to be answering that question one workflow at a time. That may be less flashy than the usual AI hype, but it's probably closer to how AI agents will actually enter the workplace: carefully, unevenly, and with humans still nearby – at least at first.
