Key Takeaways

  • Apollo says automated financial agents could shift household cash away from low-yield bank deposits at unprecedented speed.
  • The scenario is a conditional systemic-risk warning, not evidence that an AI-driven bank run is underway.
  • Banks face intertwined liquidity, authorization, operational-resilience, and technology-concentration risks as agentic AI adoption grows.

Apollo has raised a fresh concern for the banking industry: AI agents could make deposit flight faster, broader, and considerably more automated. Rather than waiting for customers to compare rates and manually transfer money, an agent could continuously scan available products and move cash toward accounts offering higher returns.

The warning describes a potential scenario, not an ongoing bank run. Still, the mechanics deserve attention. Household deposits are a relatively stable and inexpensive funding source for banks. If AI agents begin reallocating that money routinely, institutions may have to pay more to retain deposits, rely more heavily on wholesale funding, or accept greater volatility in their balance sheets.

A consumer-authorized agent might make a perfectly rational individual decision by identifying that a savings account is paying below-market interest and transferring funds elsewhere. At scale, however, thousands or millions of agents making similar decisions at roughly the same time could amplify liquidity pressure. What happens when software, rather than human inertia, determines where deposits sit each morning?

Meta Muse has been cited as an example of the type of agent that could eventually automate cash reallocation. Large incumbent institutions such as JPMorgan Chase and Morgan Stanley illustrate the banks that may encounter pressure across both consumer deposits and treasury management. Corporate cash could become more mobile too, particularly if finance departments authorize agents to optimize short-term returns within predefined risk limits.

Adoption is already moving beyond experiments. The Cambridge Centre for Alternative Finance found that 52% of surveyed financial-services firms are adopting agentic AI. At the same time, 46% identified operational resilience as a leading industry risk. Those figures capture the tension: institutions want the productivity and service benefits of autonomous systems, but they also recognize that automation can create new dependencies and faster failure paths.

A separate KPMG survey found that 40% of banking leaders are deploying AI agents. Banks are also directing 69% of their AI budgets toward risk and compliance, suggesting that governance is becoming a central part of deployment rather than an afterthought. That said, conventional approval workflows may struggle when agents can initiate transfers, interact with several providers, and revise their actions as market conditions change.

Authorization is one weak point. Research cited by the Cloud Security Alliance found that 62% of financial-services organizations have deployed AI agents, while 85% expect agents eventually to initiate financial transactions. Some 65% said new authorization models will be required. Banks may need granular transaction limits, explicit customer consent, real-time anomaly detection, revocation controls, and clear records showing why an agent took a particular action.

There is also a concentration issue lurking behind the customer-facing applications. A Bank of England estimate cited by Deloitte indicates that the three largest cloud providers supply approximately 75% of cloud services, 45% of AI models, and 30% of data services used by UK financial firms (source). If many banks and agents depend on the same infrastructure, an outage or flawed model update could produce correlated behavior across institutions.

Existing governance approaches offer a starting point. The NIST AI Risk Management Framework 1.0, published in 2023, organizes oversight around governing, mapping, measuring, and managing AI risk. The Financial Stability Board’s 2017 AI and machine-learning principles also provide context for liquidity, third-party dependencies, and operational resilience. Banks could adapt those approaches to stress-test rapid automated outflows and impose friction when activity departs from expected patterns.

AI agents may ultimately improve competition by helping customers find better returns and avoid leaving money in poor-value products. But efficiency for depositors can translate into funding instability for banks. Apollo’s warning puts that trade-off on the agenda before autonomous financial transactions become routine, giving institutions and regulators a limited window to design controls without stripping away the consumer benefits.