Understanding the banking industry’s reaction to AI agents

• Key insight: The introduction of AI agents capable of transacting on behalf of their customers has many banks placing restrictions on their use. But the reason for banks’ caution is not well understood, and is sometimes misrepresented.
• Forward look: it seems likely that if customers want to use AI agents to act on their behalf, the security and legal liability issues will be resolved through private negotiations, industry standard setting, regulatory actions, legislation or litigation. And eventually there could be an effect on deposit betas. But we are a long way from there.

The potential effect of agentic AI on banking has become a topic across media and this quarter’s bank earnings calls. Much commentary, however, misunderstands what is at stake and what is motivating banks to take some of their recent actions.

Processing Content

Agentic AI in banking includes two major use cases. One involves an agent accessing a customer’s account to make payments or otherwise transact. A second involves an agent managing the customer’s account — either maximizing yield by regularly sweeping funds from a deposit account to a higher-yielding product or reducing the customer’s risk by moving uninsured deposits out of a struggling bank.

These use cases raise four important questions that should not be conflated.

Question 1: Will customer use of AI agents increase cyber and fraud risks? As we have seen online fraud explode in volume, defensive work has focused on confirming customer identity; increasingly, that verification will be biometric. Agentic AI, however, can obfuscate identity and bypass a bank’s fraud controls designed to verify that there is a human in the loop, and that the transaction is not the product of a rogue or hallucinating agent. Banks are still developing tools to manage this risk.

 
Question 2: Who is responsible in the event an agent makes a mistake — for example, sends money to the wrong place or orders the wrong item? Certainly, Muse’s terms of service disclaim practically all legal liability to the user, and anyone hoping to contact Meta or Muse will likely find no one available to help with the problem. Unhappy customers will inevitably call the bank that processed the payment. Thus, for banks (and one hopes, policymakers) to be comfortable with agentic transactions, liability and responsibility will have to be established through private sector negotiations, regulatory actions, legislation or litigation.

Question 3: Will the use of agents to move deposits to higher-yielding products increase banks’ deposit betas — how much they pay for deposits in comparison to market rates? Of course, depositors wishing to maintain deposit insurance protection already have an incentive to increase the yield they earn on their cash by moving excess cash to savings accounts or CDs. For those seeking even higher yields, the concept of sweeping excess cash to a money market fund is ages old, and corporate customers negotiate fee waivers and other allowances in exchange for holding deposits.

That said, doing so takes time and attention. But agentic AI can manage this task for depositors, reducing the frictional costs of yield seeking, and is therefore likely to foster a “just in time” deposit balance world.

Read more:

In such a world, consumers and businesses will earn more on their money, but banks will face a business challenge, as the low-cost funding of a deposit franchise is a principal benefit of the charter. Over time, banks that offer a wide range of services will be better able to absorb higher deposit costs — whether by earning more on a varied set of assets, offering the money market fund or other account to which deposits are swept or adjusting their fees. Banks might hope to offer the agent that provides these services, leveraging the trusted relationships they have built and maintain with their customers.
Question 4: Will the ability of agents to move funds based on preprogrammed instructions lead to bank runs? This issue has received a lot of media and market attention, springing from a short note published by Apollo’s chief economist, Torsten Slok, last month, entitled “Is an Agentic Bank Run Coming?” That said, the note itself was about long-term yield optimization over time, as already discussed. So, on its own terms, not a run, but more of a leisurely stroll, notwithstanding the title.

In theory, agentic runs could be possible — for example, a depositor instructing an agent to move a deposit if a bank’s parent company receives a credit ratings downgrade. But that seems highly unlikely. For a business, that means massively disrupting its payroll and vendor processing, among other consequences, and would not be done lightly. On the retail side, deposits under $250,000 are insured, so many depositors simply may not have an incentive to prebake deposit transfers based on some preset criterion. So, it seems likely that there would continue to be a “human in the run” even if banks did not require customer authentication of the transfer.

Currently, many banks are blocking Muse and similar agents from accessing customer accounts or moving money without customer authentication by the bank. Their statements cite the security and liability issues above. In many media reports, though, their motivation is alleged to be the last two issues — fear of having to pay more for deposits or fear of a run. The former concern is longer term, though, and the latter is dubious.

It is worth noting that banks have already dealt with a similar issue in the form of data aggregators, which several years ago began using consumer credentials to “scrape” data from banks’ systems to enable third-party fintechs to offer financial products to banks’ customers and, not incidentally, harvest and aggregate the data for sale. Significant security concerns initially prompted some banks to block screen scraping. Over time, though, major banks and aggregators negotiated agreements requiring data to be gathered through a secure API and setting various other terms of the banks’ consumer-permissioned data sharing.

Of course, banks did so because their customers wanted products offered by third-party fintechs, which required use of customers’ bank data, and banks like to keep their customers happy. Thus, it seems likely that if customers want to use AI agents to act on their behalf, the security and legal liability issues will be resolved through private negotiations, industry standard setting, regulatory actions, legislation or litigation. And eventually there could be an effect on deposit betas. But we are a long way from there.