In September 2026, Meta introduced Muse, its personal Artificial Intelligence (AI) Agent. The company describes Muse as an AI agent that can perform tasks across browsers and applications, subject to permissions for sensitive actions such as “agentic shopping” on their behalf. Similarly, Robinhood offers agentic brokerage trading. These technical innovations make a new question tangible: Could agents that search for and automatically transfer money to bank accounts with higher deposit rates ultimately trigger a bank run?
Consider a household that keeps part of its savings in a highly liquid checking or savings accountto pay bills and cover unexpected expenses. These funds could earn a higher return elsewhere, but comparing offers, opening accounts, and transferring money between accounts requires time and attention. Personal AI agents could increasingly take over these activities by identifying suitable accounts and executing transfers on behalf of their users, effectively reducing efforts for households. This automation may improve how households manage their savings. However, it could also make deposits more mobile and change how banks compete for funding.
How simple money transfers could become a bank run
A potentially pivotal mechanism is the synchronization of agents’ decisions, i.e., many agents transferring money in response to similar information or decision rules. For instance, agents may rely on the same bank rankings, withdrawal thresholds, or risk indicators. A single platform’s update of conditions, e.g., interest rates, could consequently initiate a large number of transfers at approximately the same time, even when the underlying AI models differ.
In a 2024 report, the Financial Stability Board warned that shared data, models and concentrated providers can contribute to correlated behavior in financial markets. A bank would face substantial pressure if a sufficiently large volume of delegated deposits could leave faster than the institution could obtain the necessary liquidity due to synchronized decisions across agentic AI models. That is because in case of a customer transfer of deposits to other banks, the originating bank must settle these payments using central bank reserves, thereby reducing its available liquidity. If these reserves are insufficient, the bank must obtain additional funding, for example, through interbank borrowing or asset sales.
This feedback mechanism is central to the famous Diamond and Dybvig’s theory of bank runs: deposit withdrawals that make sense from the individual point of view can destabilize banks, undermine confidence in the bank’s ability to cash out customers and result in a spiral of deposit outflows. The collapse of Silicon Valley Bank in March 2023 serves as an illustration. On March 9 alone, the bank lost more than US$40 billion in deposits, according to the Federal Reserve. Even though SVB differed from most retail depositor banks in its concentrated depositor base, the episode illustrates how quickly liquidity pressure can develop. With the broad use of AI agents by households in search of the highest-yield deposits, a piece of new information could trigger runs on otherwise healthy banks.
Why deposit runs may still be unlikely
However, three arguments shed doubt on a bank run caused by agentic AI.
First, while AI agents could reduce the attention, time, and effort required to compare accounts and transfer money, households would first have to trust them with access to their bank data and funds. Empirical evidence suggests that people can be reluctant to delegate consequential tasks to AI, even when they know that it outperforms human alternatives. Adoption, therefore, does not follow directly from technical capability. Moreover, identity checks and payment authorizations that require human involvement could continue to impose behavioral frictions even after adoption. These frictions could slow transfers and limit the speed at which an agent-driven withdrawal of deposits unfolds.
Second, the economics of AI agents may also limit adoption. For households, the relevant gain from account-switching is the additional interest income net of the cost of using an agent to identify and act on better offers. Managing funds across accounts requires repeated planning, reasoning, and action, all of which consume tokens and thus incur computational costs. These costs could substantially reduce, or even outweigh, the additional interest income, particularly for households with modest balances or when interest rate differences are small.
Current pricing schemes may obscure this trade-off. Some agents are available for a flat fee or free of charge, subject to usage limits, so users do not directly bear the cost of each additional task. Over time, however, providers may pass more of these costs on to consumers or banks, for example through pay-per-use pricing. Declining costs per token would mitigate this pressure, but would not necessarily reduce total costs if more intensive agent use increases token consumption. If households ultimately bear these costs, continuous account-switching may be worthwhile only for those with sufficiently large balances, or those whose current accounts offer significantly lower interest rates than alternative options.
Third, adoption need not imply synchronized behavior. Households differ in their assessments of risk, liquidity needs, and sensitivity to interest rates. Agents that learn and retain these preferences could initiate transfers at different times and with different frequencies, limiting the extent of coordinated withdrawals.
Taken together, these considerations suggest that automated account-switching could become attractive for some households without becoming universal or generating synchronized transfers. They therefore qualify the argument that widespread access to AI agents would necessarily increase bank-run risk, although they do not eliminate that possibility entirely.
Regulatory Safeguards and Outlook
Even if the additional bank-run risk proves limited, the use of agents to execute bank transfers warrants regulatory safeguards. The EU AI Act already establishes a risk-based framework for artificial intelligence, including transparency obligations, requirements for human oversight, and rules for high-risk systems. Although many consumer-facing financial agents may not fall within the Act’s strictest categories, providers and deployers may still face obligations concerning transparency, risk management, data governance, record-keeping, and users’ ability to understand when they are interacting with AI.
Detailed regulatory proposals require a better understanding of households’ adoption of AI agents and potential financial stability risks. However, several guiding principles can already be set out. Consumers should give informed consent, be able to set clear limits on autonomous actions, and have an immediate means of stopping an agent or revoking its permissions. Large transfers could require explicit human confirmation. An AI literacy requirement could also help users understand the scope and limits of delegation, although its benefits would need to be weighed against the barriers it creates.
Finally, responsibility and liability for erroneous or unauthorized transfers must be clear. Personalization complicates this issue because an agent’s actions may reflect both preferences supplied by the user and choices made by the provider in designing and operating the system. A regulatory framework should clarify the respective responsibilities of users, agent providers, and banks, including who bears losses when an agent acts outside its mandate or fails to execute it correctly.
Blog entries represent the authors’ personal opinion and do not necessarily reflect the views of the Leibniz Institute for Financial Research SAFE or its staff.
Authors:
Kevin Bauer is Coordinator Digital Finance and Bridge Professor at SAFE and Professor for Game-Theoretic and Causal AI in Business and Economics at Goethe University Frankfurt.
Alexander Hillert is Co-Director of the SAFE Department “Financial Intermediation” and Professor for Finance and Data Science.
