Citi and Bank of America Restrict AI Agent Access to Customer Accounts

Recently, some media reports said that Citigroup and Bank of America have followed JPMorgan Chase & Co. in restricting AI agents' access to customer accounts. Previously, several U.S. banks had expressed concerns to regulators that these intelligent agents might move idle cash in pursuit of higher returns, thereby causing bank deposit outflows.

These concerns are not unfounded. On September 8 this year, U.S. tech giant Meta officially launched its personal AI agent Muse. Relying on an ecosystem of external connectors, it has the ability to independently initiate payments, compare prices, and move funds, so as soon as it was released, it touched a sensitive nerve in the financial industry.

The market quickly signaled concern about this new variable. On September 22, the S&P 500 Financials Index fell nearly 2%, JPMorgan and Wells Fargo both dropped more than 3%, and Charles Schwab fell more than 6%. Bloomberg's interpretation got to the core: investors worry that personal AI agents represented by Muse may break the "consumer inertia" on which the financial industry has long relied—even if higher-yield, lower-cost financial options exist in the market, users will still choose to stay with their original institutions due to usage habits, information gaps, and switching costs.

At the same time, Meta's intention to extend into financial services is becoming clearer. On October 12, it launched two new models, Muse Wonder and Muse Wonder Full. According to the official introduction, the models can complete specialized tasks such as "understanding personal financial situations," "handling insurance claims," and "interpreting lease terms."

Muse's growth momentum is quite rapid. Public information shows that just over a month after launch, its global downloads have approached 9 million. Although Muse at this stage is still a long way from being an "automated financial manager" capable of fully managing all of a user's financial assets, as its capabilities continue to penetrate scenarios such as personal finance, market concerns about AI agents disrupting bank deposits, wealth management, and the payments landscape are continuing to build.

How long has it been since you reviewed your auto-renewing subscriptions? A membership service costing a few dozen yuan a month and long left unused may keep charging you as you forget about it; tens of thousands of yuan in funds often sit for a long time in low-yield accounts because you cannot be bothered to compare interest rates... In the past, users had to investigate, compare, and act on such financial matters themselves; now, personal AI agents are trying to take over.

Canceling subscriptions is one of Muse's most easily perceived functions. In a media test, the tester handed credit card statements to Muse for verification, and it identified an Adobe auto-subscription the tester had long forgotten and helped complete the cancellation, saving nearly $25 a month.

Although the amount is small, forgotten auto-renewals are widely hidden in various subscription services such as software memberships, fitness services, and communications plans. Before this, most AI in the consumer internet played a marketing role: predicting demand through algorithms, pushing products, and ultimately increasing conversions. Muse, by contrast, represents a new paradigm—AI stands on the consumer's side, reviewing existing spending and proactively comparing and uncovering more cost-effective alternatives.

Goldman Sachs analysts have already turned their attention to companies that profit from "consumer inertia." In their judgment, AI agents can greatly reduce the cost for users to compare prices and switch service providers, and the customer retention logic of telecommunications, insurance, travel, and even financial services may therefore come under pressure.

Industry discussion has also extended from small-scale scenarios such as canceling subscriptions to the more critical migration of deposits. Compared with canceling services, moving funds across accounts is more complicated, but the underlying logic is the same: in the past, the cost of finding alternatives and completing the switch was always borne by consumers themselves. Some market participants believe that U.S. consumers lose money every year due to "inaction" such as failing to refinance, missing bill payments, and leaving funds in low-yield accounts for long periods, and this hidden cost is called the "Laziness Tax" in the industry.

Another media report, citing data from the Federal Reserve Bank of St. Louis, said that about $7.12 trillion is held in U.S. personal and business checking accounts, a considerable portion of which earns only very low interest or none at all. Even if users know that higher-yield savings accounts and cash management tools exist in the market, a series of operational hurdles such as product comparison, opening a new account, identity verification, and fund transfers still leave large amounts of money sitting in existing accounts for long periods.

Personal AI agents, however, are precisely expected to greatly reduce such switching costs. Bank of America analyst Ebrahim Poonawala suggested that if AI can smoothly move funds from low-yield accounts to higher-yield products, it could put substantial pressure on banks' net interest margins. "For banks, deposits are both customers' assets and the core source of funding on the liability side. If customer funds frequently migrate in pursuit of yield, it will become increasingly difficult for banks to maintain low-cost deposits."

Apollo Global Management chief economist Torsten Slok, meanwhile, raised a more extreme scenario: an "agentic bank run." By his calculations, some fintech companies can currently offer deposit yields of 3.3% to 5%, while the average U.S. checking account rate is about 0.1%. If large numbers of households use AI agents to automatically seek out higher-yield accounts, banks could lose some of the low-cost deposits they use for lending—and that would become a problem for the entire financial system.

"If banks are forced to raise deposit rates to prevent retail customers from moving their money elsewhere for higher rates, that could deal a major blow to bank profits," Slok warned.

This claim quickly sparked controversy in U.S. fintech and banking circles. Some industry insiders believe that an "agentic bank run" is still a fairly extreme scenario for now. Whether consumers are willing to hand money-management authority to AI, whether the balances in ordinary household accounts are worth migrating frequently, and whether banks can retain customers through service and comprehensive financial relationships could all limit the actual scale of fund flows.

The worry over deposit migration: authorization sets the first threshold

Although there is no evidence at this stage that Muse's rollout is triggering deposit migration, vigilance in the U.S. banking industry has already risen. Several banks have expressed concerns to regulators: such AI agents may automatically seek higher-yield targets for users and transfer idle funds, thereby causing deposit outflows.

On October 13, JPMorgan said on its third-quarter earnings call that it would bar AI agents from accessing customer accounts, citing the risk of thousands of "malicious agents." "These methods are not safe. JPMorgan is doing as much as it can to block such access while studying more appropriate ways to implement it."

Soon afterward, media reports said Citigroup has explicitly prohibited third-party AI agents from directly logging into customer accounts or executing transactions on its digital banking platform, while still allowing some customer-authorized read-only data access. Bank of America said it is restricting AI agents' access to consumer deposits; the restriction will remain in effect until the bank establishes agreements determining ultimate liability for AI-authorized payments.

The banks' differentiated controls conceal a clear permission boundary: for a personal AI agent, identifying redundant spending and comparing yields across accounts requires only retrieving billing and account information; initiating payments, transfers and similar actions on the user's behalf, however, means obtaining a higher level of operational permission.

According to Meta's disclosures, Muse is equipped with a dedicated cloud-based virtual computing environment and browser, and can access web pages, call external services and execute tasks within a defined scope. U.S. fintech company Plaid announced on September 8 that it provides Muse with user-authorized financial account connectivity, supporting the reading and parsing of account data according to user instructions.

However, having the ability to read and analyze account data does not mean Muse has permission to directly transfer funds or allocate assets. Meta has likewise set operational red lines for Muse: according to public information, for sensitive operations such as purchases and sending emails, Muse first asks the user to confirm.

Beyond permissions, consumers' willingness to trust is another hurdle for the large-scale rollout of financial agents. A Visa Agentic Commerce survey released in September this year showed that only 23% of surveyed U.S. consumers trust generative AI to complete payments directly on their behalf; another 61% said they trust Visa to handle agent transactions. The survey was conducted in May this year and covered 2,065 U.S. adults.

Visa Group President Oliver Jenkyn said consumers have already shown interest in using AI for shopping and payments, but trust remains an important prerequisite for agentic commerce to scale.

Building trust, moreover, goes far beyond the single threshold of whether users are willing to hand over transaction permissions. Even with user authorization, banks holding account permissions still need to carefully define boundaries: in what ways external agents may retrieve account information, whether they can trigger fund operations, and the entire set of access rules and risk prevention and control systems all need to be clarified and strictly controlled by banks one by one—this is precisely the core reason why the aforementioned banks introduced restrictive measures.

Yet blocking unsafe external agent access cannot eliminate customers' genuine demand for convenient money management. Faced with this contradiction, JPMorgan chose to offer a solution through a self-developed tool. It is reported that the bank is advancing Smart Cash, a cash management AI tool intended to use cash flow forecasting to enable intelligent fund allocation and open a channel between checking accounts and higher-yield brokerage products. JPMorgan confirmed at the aforementioned earnings call that this tool has entered a targeted small-scale pilot.

Although Smart Cash is not yet an AI wealth-management agent capable of fully autonomous decision-making, it clearly sends a signal: banks are proactively responding to customer needs and reducing the manual operating costs in fund allocation.

On one hand constraining risky external agent access to accounts, and on the other relying on self-developed products to lower the operating costs of customer money management, JPMorgan's two sets of actions reflect the dual trade-off banks now face. From a competitive logic perspective, proactively helping customers move idle funds into higher-yield products may not retain every low-cost deposit, but it helps maintain customers and preserve the overall financial services relationship.

China's financial agents accelerate rollout: banks pursue large-scale deployment

Compared with this contest surrounding third-party external agents in the U.S. market, domestic commercial banks are accelerating the rollout of agents, continuing to build out deployments in scenarios such as customer service, account management, and wealth management.

State-owned major banks are leading the way. ICBC has already built two major super-agent systems, one for customers and one for employees: "Gong Xiaocai," aimed at retail customers, served more than 22 million times cumulatively in the first half of the year; the credit review assistant "Gong Xiaoshen" assisted in approving more than 1 trillion yuan in loans. Agricultural Bank of China launched its self-developed agent "ABC-Claw" and the enterprise-level digital assistant "Yiming," supporting multi-agent collaborative invocation across business lines.

National joint-stock banks each have their own focus. China Merchants Bank has iterated its retail-side "Xiaozhao" intelligent service system, extending AI capabilities to high-end services such as private banking asset allocation; China CITIC Bank has built a company-level agent management platform, focusing on exploring AI applications in core areas such as investment research analysis, intelligent trading, and customer service; Ping An Bank launched 140 operations-review agents in the first half of the year, covering 55 business scenarios, used in middle- and back-office risk control processes such as anti-money laundering due diligence, account opening review, and cross-border remittance verification.

From these cases it is not hard to see that financial agents with cross-scenario execution capabilities have also emerged in China. One key difference is that domestic bank agents are embedded within their own closed business loops, whereas Muse mainly relies on browsers, external connectors, and the Plaid financial data channel to complete user instructions. But whichever path is taken, account data access, transaction authorization, and risk control are core issues that cannot be avoided.

Industry insiders believe that when the domestic financial sector previously discussed AI agents, the focus was mostly on reducing costs and improving efficiency, with implementation concentrated in processes such as customer service, marketing, trading, and risk control. The inspiration brought by Muse raises a deeper question: once users have a digital proxy that can autonomously check accounts, compare prices, and screen financial products, what impact will this have on the customer stickiness and low-cost deposit foundations that financial institutions have long relied on?

Looking at the domestic market, the migration of household deposits is currently driven mainly by interest rates and changes in asset prices. A recent CICC report shows that in the first half of 2026, household deposits increased by 7.6 trillion yuan, a year-on-year increase of 3.2 trillion yuan less, the smallest increase in nearly a decade; during the same period, funds invested by households in non-deposit financial products increased by a net 6.6 trillion yuan, a year-on-year increase of 3.3 trillion yuan more. The funds flowed mainly into insurance, private funds, and stocks, with net inflows of 2.2 trillion yuan, 1.3 trillion yuan, and 0.9 trillion yuan, respectively.

Against the backdrop of declining deposit rates already prompting residents to actively compare returns across asset classes, one variable worth continuing to watch is: if personal AI agents further reduce the friction costs of querying financial products, comparing prices, and switching between institutions, will this amplify the speed and scale of fund migration?

This also means that the potential challenges personal agents pose to banks will go far beyond deposit outflows. The deeper contest is over who will take on users' financial needs, screen and compare products on users' behalf, and ultimately hold the execution authority for fund operations. For the global banking industry, the question posed by AI may shift, from "how to make good use of agents to improve our own efficiency" to "how to serve customers who hold agents in their hands."

This article is from the WeChat public account "Caijing Mayflower" (ID: Caijing-MayFlower), authors: Cai Bingyan, Zhang Yingxin, published by 36Kr with authorization.