Momentum is building around agentic AI, as banks like JPMorganChase,
For managers used to leading teams of junior workers (whose tasks these agents are typically taking on), these digital workers will present new management challenges.
Agentic AI takes a large language model like OpenAI's ChatGPT or Anthropic's Claude and allows it to do things, with minimal human intervention. Agentic refers to acting as an autonomous agent, capable of performing tasks on its own.
Not everyone is sold on the idea. Gartner analysts recently predicted that 40% of agentic AI projects will be shut down by 2027, due to escalating costs, unclear business value or inadequate risk controls.
"Most agentic AI projects right now are early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied," said Anushree Verma, senior director analyst at Gartner. "This can blind organizations to the real cost and complexity of deploying AI agents at scale, stalling projects from moving into production. They need to cut through the hype to make careful, strategic decisions about where and how they apply this emerging technology."
But many banking industry experts believe agentic AI is gaining traction.
"We are still in the early days," said Sumeet Chabria, CEO and founder of ThoughtLinks and former chief operating officer of global tech and ops at Bank of America. "Almost all banks are talking about agentic AI right now, discussing it in architectural teams within the company, and a few of them have already taken the first batch of AI agents into implementation."
The fact that AI agents can look at thousands of records a day, working 24/7 without resting or sleeping, and flag errors in the morning, "is a wonderful thing," Chabria said.
"AI agents offer significant operational benefits," said Andy Winskill, AI strategist and consultant. "They're tireless, consistent yet adaptable and scalable. A manager can easily scale a digital workforce to increase productivity dramatically."
Agentic AI use cases in banking
One bank that Chabria works with is considering using AI agents to do "four eyes checks," where two people review and approve a transaction or action and "six eyes checks," where three people approve it.
"These are processes where you're making a high-value payment, or you're inputting a standard settlement instruction, or you're looking at sanction screening, and you need to have two people look at that document and OK it," Chabria said. These agents give feedback to human workers on their completed tasks.
French bank BNP Paribas is launching a U.K. fintech incubator that will host generative AI startups that have developed applications for banks.
Another place Chabria has seen agentic AI deployed in banks is on trading floors.
"On trading floors, traders are under a fair amount of pressure to make sure that they capture the data properly, not of the trade itself, but of the client, the counterparty, etc.," Chabria said. "If you don't capture this properly, it becomes a huge issue for regulatory reporting in the back end. So agentic AI systems are being built that can verify what the traders captured are correct or not, and flag it back to the trader. It's basically mimicking a middle office person who does repetitive, low-value activities."
At the banks Chabria works with, executives are debating questions like: Should these agents be assistants to human beings, or co-workers to humans? Will they collaborate with human beings and do things autonomously, or will groups of agents support a team that is doing investment banking or know-your-customer work?
"A lot of banks are thinking about the framing of this, because on one side, certain banks are only talking about it in terms of not replacing employees, being assistants to employees," Chabria said. "But on the other hand, some banks are talking about the concept of digital workers, which are autonomous. They have some level of intelligence. They can sense, perceive the world. They can carry out tasks. They are adaptive. They can change their behavior or their inner workings to improve the underlying activity. And they could do a handoff."
At banks Chabria works with, AI agents are autonomous, yet boxed in.
"I have not seen an agentic AI system so far in banking where agents are let loose in the bank, and have been told to just do whatever they want," Chabria said. "They have a job scope, a set of clear roles and responsibilities. They have inputs, they have outputs, and they have been put under guardrails. All of them have a level of very direct supervision from a human manager. None of them are running without management or supervision right now. And the role of the manager is to very frequently certify that the agent is performing as expected."
Job scope, roles and responsibilities and handoffs or handshakes to human beings all need to be done with strong documentation compliance policies and procedures to make sure that agents have embedded controls, which is what risk management needs as well, Chabria said.
Casetllum.ai has a family of AI agents that handle alert resolution for sanctions, anti-money, laundering, know your customer and adverse media screening. These agents don't file suspicious activity reports, but they do the research and investigative work that leads up to filing a report, including writing a first draft — in other words, the work of a junior analyst.
That means, in theory anyway, that a bank can grow in assets under management and transaction size and not have to hire a lot of compliance officers, according to Peter Piatetsky, co-founder and CEO.
In this scenario, "all the humans get promoted," Piatetsky said.
"Our approach here is to treat an AI agent in a similar way that a manager would treat a human employee," he said. "The AI agents are being trained specifically to assist with tasks. We have our AI agents review their own work. So part of the training is to make sure that it follows the procedures that you have been trained on, in the same way that a human would."
Challenges of managing digital workers
AI agents lack emotional intelligence and intuition, meaning they can't navigate complex human interactions or ethical dilemmas without human support, Winskill said. Human managers have to oversee AI decisions to anticipate and mitigate the risks an AI agent cannot perceive, he said.
"Managing an AI agent fundamentally differs from managing humans," Winskill said. "Humans require emotional intelligence, motivation, and career support. AI agents, by contrast, are driven by technical parameters and clearly defined objectives. Managing AI means moving from task supervision to goal-driven collaboration; you set outcomes and let the AI autonomously identify the best path."
The AI and cybersecurity-focused startups completed the vetted proof-of-concept program to prepare for their next level of growth.
One challenge of managing AI agents is handling access entitlements, Chabria said. Banks already have to make sure that all their human employees have access to the data and systems they need and no more. Now, AI agents have access to a lot of data and systems, and that has to be mapped and managed. "Otherwise, you've got to revoke access," Chabria said.
Another hurdle, according to Chabria, is that processes often need to be redesigned before they can be handed over to an AI agent.
He recently analyzed 5,000 banking processes to determine what parts of those processes could be moved to agentic AI.
"What I found is that in most cases, the process has to be completely redesigned," Chabria said. "You cannot just take a process that was run in a traditional way and just split it halfway and give half to an agent. You've got to redesign and redefine it."
In most cases, a process that brought an AI agent and a human reviewer together (human in the loop) worked best.
"My view is that, if done well, it should elevate everyone," Chabria said. "Human beings don't do the repetitive, routine work that they do today, but they do more insightful, creative work."
Best practices for managing AI agents
Instead of giving an annual performance review, as a manager would do for a human team member, managers of AI agents need to give feedback more quickly, intraday or instantaneously, Chabria said.
Another best practice is to have AI agents managed not only by the manager of a team, but also by the people who created and deployed them, Chabria said.
"If the agent performance is slightly off, then the people who know the mechanics of the agent can fix it," he said. The best AI agents are created with feedback loops and are given clear performance metrics in terms of the outputs they generate.
Banks also have to choose the right people who can collaborate with AI agents and train those people to work with them. "You can't just say, hey, you did this work before, and now this is an AI agent, work with it," Chabria said.
Performance reviews for AI agents have to address model drift, the gradual degradation of AI accuracy, Winskill said.
"That means AI managers need regular alignment reviews," he said. "Managers must consistently sample AI outputs, detect any divergence from intended performance, and promptly recalibrate or retrain the model. Unlike quarterly human performance reviews, AI reviews are continuous, requiring near-daily monitoring and adjustment."
"Managing AI agents isn't easier, it's a different kind of challenge," Winskill said. "You're exchanging the complexity of human management for technical oversight, continuous monitoring, and performance validation."
What happens to managers whose people are swapped out for AI agents?
If a manager used to oversee 300 people and through deployment of AI agents, now that person is managing 30 people and 200 AI agents, what happens to that person's level of influence and career trajectory?
"Over time, I think we should reward managers that adopt responsible technology," Chabria said. "It's inevitable. This is happening. It's just a matter of time. We can debate whether it's one year, five years, seven years, but this is happening. We're going to start using it on the iPhone, and then people will say, 'Why can't we do this at work?'"
But companies need to go about this in a sensitive manner, he said. "Do you reward an AI agent, do you give it a bonus? If so, do you really give it to the manager of the bot? Or to the developers of the agentic AI system? You can't just introduce these things at scale without thinking through this, the whole framework of human capital management versus agentic AI capital has to be put in place."
Having fewer human direct reports doesn't diminish managerial influence; it redefines it, according to Winskill.
"Managers now derive power from their ability to orchestrate outcomes across a hybrid human-AI workforce," he said. "Influence shifts from headcount to strategic oversight."
