Who is accountable when an AI agent walks into a bank?

Some of the world’s largest banks have started drawing a line in the AI sandbox. As reported in the Financial Times on October 16, Citigroup, Bank of America, and JPMorgan Chase are restricting external AI agents from accessing customer accounts and executing transactions. The concerns range from unauthorized movement of funds and liability, to the potential disruption of traditional deposit relationships, to “the dent” paying higher interest rates could put in banks’ profits.

Agentic models have advanced enough to reason, plan, converse and, increasingly, act on customers’ behalf. So, it’s not surprising that we’ve reached the moment in history where personal AI agents will be managing more of our financial lives—making decisions and initiating transactions, many of which have always required direct human involvement.

Still, we shouldn’t confuse the ability to act with the authority to do so.

It’s one thing for a customer to ask an AI agent to find a better savings rate, but it’s a different—and potentially more dangerous—step entirely for that agent to independently move the customer’s money between institutions.

Such “advancements” in agentic capabilities have the potential to set the banking industry back far more than they would move it forward. This is why I believe that financial institutions, frontier labs, and technology companies like Backbase that bridge the gap between the two, must come together to envision the next stage of agent-enhanced banking that benefits all parties involved.

Banks have spent decades building the systems, controls, and regulatory frameworks required to safeguard people’s money. They have obligations that extend far beyond providing a convenient customer experience, and are responsible for protecting assets, verifying identities, enforcing permissions, preventing fraud, and ensuring that every consequential action can be explained and accounted for.

As I wrote in my open letter to the leaders of the frontier labs, those responsibilities don’t disappear because an AI agent enters the picture. And they can’t simply be delegated to whichever technology company happens to provide the most intelligent model.

I understand why banks are drawing boundaries. By all means, banking leaders should set up appropriate guardrails to ensure that agents don’t overstep bounds that can’t be retraced and recovered, if necessary. But let’s not mistake those guardrails for a long-term strategy that acknowledges the imminent participation of agents in customers’ financial lives.

Agentic AI is already being integrated into banking. We've crossed the Rubicon, and there's no turning back now. Nor would retreat be beneficial for customers or banks alike.

At the same time, we can’t build the future of banking by allowing intelligence to operate without the bank’s authority.

BBVA’s own announcement makes the ambition clear: enable AI agents to act on behalf of customers, while preserving security, privacy, and customer control.

This sets an important precedent; we need open standards for interoperability, clear models for delegated authority, and systems that make autonomous actions governable and auditable.

Yet, having worked with financial institutions for more than two decades, I know that a protocol is not enough. Because knowing who an agent represents is not the same as knowing what that agent is authorized to do. And granting permission is not the same as ensuring that every action remains within the bank’s policies, risk controls, and regulatory obligations.

Banks require a practical, structured framework to implement and scale these capabilities safely within their existing software lifecycles. And yet, we know that this challenge is bigger than any one company, platform, or protocol.

All banks should be able to embrace the best intelligence available, including external agents, without surrendering responsibility for the relationships and transactions entrusted to them. This requires concerted collaboration between banks, AI companies, technology providers, and regulators.