AI agents are gradually moving beyond chatbots and gaining the ability to carry out financial operations: transferring money between accounts, paying for purchases, and trading assets. This could simplify working with banks and investment services, but it also creates new risks. What happens if an agent sends money to the wrong place, makes a loss-making trade, or independently performs an action the user did not authorize? The financial industry still has no clear answers to questions of liability, reliability, and regulation for such systems.

These issues were discussed by representatives of the financial sector at an Axios Live event held on October 20 in Las Vegas. Participants included Travis Hill, chairman of the U.S. Federal Deposit Insurance Corporation (FDIC), and Yoshi Yokokawa, co-founder and CEO of fintech company Alpaca, Axios reports.

When AI gets access to money

AI agents differ from ordinary chatbots in that they can not only answer questions and offer recommendations, but also perform actions in external systems. In finance, this means the ability to delegate part of everyday operations to software — from paying for goods to managing accounts and investment transactions.

This approach could simplify interaction with banks. For example, an agent could help allocate funds across accounts or select a suitable financial product. However, along with convenience, the client hands over some authority to the system and therefore must understand what decisions it is allowed to make on its own.

Travis Hill noted that modern technologies are already making it easier to transfer money between banks. But as AI agents develop, a more complex question arises: how far depositors are willing to go in delegating financial decisions to automated systems.

The issue is not only whether a person trusts AI recommendations. It is also whether they are ready to give software the ability to manage money independently — especially if transactions happen quickly and without confirmation of each individual action.

Who will pay if an agent loses money

One of the main unresolved questions is how responsibility should be allocated for the consequences of AI actions.

“Who is responsible when something happens, when agents go out of control and start engaging in criminal activity or simply make a bad trade and lose money?” Yokokawa asked while discussing restrictions on fully automated trading on his company’s platform.

The possible scenarios differ significantly. An agent may make a mistake while carrying out a correctly worded instruction, misinterpret the user’s instructions, or perform an action it was not authorized to take at all. In addition, its behavior may depend on software errors, incorrect data, or vulnerabilities in the system.

In each case, the question arises of who should be liable for the damage: the user, the model developer, the financial platform provider, or the company that integrated the agent into a banking service. The answer will depend on the circumstances, contractual terms, and applicable laws, but there is still no universal solution for all such situations.

Yokokawa also cautioned against excessive expectations regarding automated trading. He called the belief that AI can consistently generate large profits dangerous. Even an advanced model cannot guarantee successful investments, and automation can accelerate not only useful actions but also mistakes.

Regulation is lagging behind technological development

Financial regulators have to strike a balance between two tasks: not hindering the development of useful technologies while at the same time preventing new systems from creating uncontrolled risks.

Hill explained that the FDIC’s current guidance on model risk management does not apply to generative AI and agentic systems in the form the agency is considering them today. The regulator needs to assess how to apply existing approaches to technologies that are changing rapidly and are capable of independently carrying out sequences of actions.

According to Hill, it is important to both monitor the industry’s development and approach future rules thoughtfully. Regulation introduced too hastily risks failing to take into account the specifics of new tools, but the absence of clear requirements also leaves financial organizations without a unified system for assessing risks.

The problem is especially noticeable where an agent is given access to real accounts and payments. In such a case, it is not enough to check how well the model answers questions: it is necessary to understand what operations it can perform, how its authority is limited, and how an undesirable action can be stopped.

Making a payment is easy; doing it safely is much harder

Zak Anderson of J.P. Morgan, who is responsible for data and analytics in the Payments and Global Banking division, also noted the potential of agentic commerce. In his view, AI can simplify shopping and making payments, but this model has a significant limitation.

“Right now, it is fairly easy to get a bot or agent to make a payment. The problem is that doing this reliably and securely is genuinely difficult,” Anderson emphasized.

This is precisely the key obstacle to the mass adoption of financial agents. The technical ability to make a transfer does not yet mean that the system correctly understood the user’s intentions, selected the right recipient, and avoided making a mistake. The more operations an agent performs independently, the more important access control, confirmation of risky actions, and the ability to reconstruct the sequence of decisions become.

For banks, this means that deploying AI is not simply a matter of integrating a new model. Mechanisms will be needed to limit the agent’s authority, make it possible to track its actions, and protect the client from the consequences of errors or abuse.

In the near term, AI agents may become convenient assistants in managing personal finances and shopping. But turning them into fully autonomous participants in the financial market depends not so much on their ability to carry out transactions as on whether banks, developers, and regulators can ensure the safety of such operations and define responsibility for their consequences.