Major banks spent the first half of 2026 quietly redefining what it means to have a coworker.
At BNY, some AI agents are treated less like software and more like staff: assigned to specific tasks, given their own login credentials and nicknames, and paired with a human manager responsible for training them and reviewing their work. That’s according to a Reuters report on how large banks are rolling out agentic AI across daily operations. One such assistant, nicknamed Payment Pete, was described by chief executive Robin Vince as functioning inside the bank’s systems with the same kind of oversight a new employee would get: defined tasks, quality checks, and someone accountable for how it performs.
Banks Are Racing to Define What an Agent Employee Actually Does
The push isn’t confined to one institution. 51% of banks reported piloting AI agents as of a June survey cited by Reuters, and the functions involved span wealth management, client vetting, trading, and corporate treasury. Goldman Sachs has partnered with Anthropic to build agents for trading and transaction accounting. JPMorgan is treating corporate treasury as one of the areas most ready for agentic transformation. Citi is preparing to roll out a virtual wealth management team member of its own.
The raw number of banks experimenting is less interesting than how narrowly each one is scoping the work it hands over. Morgan Stanley’s wealth management arm has built agents that help financial advisors prepare for client conversations and plans to test versions that interact with clients directly later this summer, but the bank has been explicit that agents won’t get autonomy over actual portfolio decisions. At UBS, agents comb through meeting notes, account records, and email threads to flag when a client needs to act, an annuity maturing, a rebalance overdue, then execute the transaction once an advisor signs off. The bank reports that arrangement has freed advisors to spend 70% of their time talking to clients instead of doing the paperwork around those conversations.
Oversight Is the Feature, Not the Afterthought
That pattern, agents handed a specific, well-bounded job with a human still holding the final decision, is what separates this wave of enterprise deployment from the more speculative promise of a fully autonomous digital coworker. None of the banks cited in the reporting are handing agents the authority to move client money or rebalance a portfolio unsupervised. The judgment stays with a person. What moves to the agent is the surrounding work: gathering context, flagging what changed, drafting the next action, and doing it continuously rather than whenever a human gets around to checking.
That’s a meaningfully different design goal than building an assistant meant to answer anything. It’s closer to what enterprise software vendors are increasingly describing as an always-on intelligence layer: a system that sits underneath a specific, bounded set of operational data and keeps watch on it around the clock. It surfaces what a human needs to see and act on. It doesn’t try to replace the judgment call itself.
Trust Has to Be Earned Function by Function
The slow rollout isn’t a capability problem. Banks operate in a regulated, client-facing environment where trust gets earned one function at a time, not granted wholesale because a demo looked good. A digital employee with a login and a manager is, in effect, an admission that the technology is treated as capable enough to do real work, but not yet trusted enough to do it without a name attached and someone checking the output.
That’s likely to be the template other industries follow as agentic AI moves out of pilot programs and into daily operations: narrow scope, clear ownership, and a human kept firmly in the loop until the track record earns something closer to full autonomy.
