The next contest over artificial intelligence in finance may be decided, not by who builds the most capable agent but by who allows it to act. In October 2026, JPMorgan Chase, Citi and other US institutions drew a distinction between software that can inspect an account and software that can move its money. For a customer who requires an AI assistant to compare savings rates, pay invoices or rebalance cash automatically, the difference is fundamental. JPMorgan’s reported position is to block external agents that log in (using customers’ credentials) whilst continuing to support permissioned data sharing through established open-banking providers. Reporting on Citi and Bank of America indicates similar restrictions on direct third-party agent access and transactions. These are neither universal bans on AI nor proof that every bank permits the same read-only functionality; the policies concern particular external access methods, and their precise scope differs. Undoubtedly, the security argument deserves serious attention; a customer handing a password to an autonomous application may expose an entire account, not merely the information required for a particular task. An agent can misinterpret instructions, be manipulated by malicious content or initiate a transfer that the customer never intended. JPMorgan’s own guidance on deepfake payment fraud advocates layered verification and checks before funds leave an account: if money moves instantly, opportunities to stop or recall a fraudulent payment diminish.
Responsibility is less straightforward than a slogan about human approval suggests. For example, if an agent transfers money after a customer grants broad permission, was the instruction authorised? Who bears the loss if the software misunderstood a request, or a fraudster compromised the agent? Indubitably, banks are subject to conduct, anti-money-laundering, operational-resilience and customer-protection obligations, but a third-party model developer may not carry equivalent duties. Preventing unrestricted access can therefore be a defensible interim control, especially when consent and liability arrangements are unsettled. Yet the economic incentives are not entirely aligned with consumer protection: an agent able to monitor deposit rates continuously could shift balances from low-yield accounts to better alternatives, subject to fees, access conditions and deposit protections. What once required a motivated customer comparing several websites could become a routine software instruction. Moreover, for banks that rely on relatively inexpensive deposits to fund lending and other assets, automated switching may reduce the durability of customer balances and increase funding costs. That prospect does not mean an “agentic bank run” is inevitable; deposit rates, payment friction, trust, tax treatment and protection schemes influence behaviour, with many customers valuing convenience and security over marginal yield. Nor can every deposit leave the banking system simultaneously: transfers redistribute claims and liquidity among institutions, although particular banks can experience serious outflows. The important distinction is between faster competitive switching in normal conditions and correlated withdrawals during a period of uncertainty. The present debate also resembles an earlier dispute over access to banking data. In 2025, the Financial Times reported JPMorgan’s plans to charge fintech data aggregators, prompting concerns about competition and the cost of providing customer-authorised services. JPMorgan argued that maintaining secure application programming interfaces carries substantial expense and the eventual agreement between JPMorgan and Plaid preserved permissioned access under commercial terms. This history suggests the current disagreement may be less about whether agents ever gain access than about pricing, authentication, control and liability.
Security, competition and the right to act
The opportunity for fintechs is consequently real, but not automatic. A fintech cannot safely bypass a bank’s controls simply by describing its software as an agent. It must obtain lawful access, protect credentials, establish consent and comply with applicable payment rules. Companies able to provide verified agent identities, narrowly scoped mandates, transaction limits and auditable approval records may be better placed than those relying on screen scraping or shared passwords. The UK provides an instructive comparison: Open Banking Limited describes an ecosystem in which customers can authorise regulated third parties to access account information or initiate payments through secure interfaces. Furthermore, account-information permission is different from payment-initiation authority, and this distinction could support agentic finance. One permission allows an assistant to observe and recommend, and another, separately authenticated and constrained, permits a specific payment. Hence, it does not follow that every AI application automatically qualifies for those permissions: a practical framework might offer graduated capabilities, an unfamiliar agent begins with no access, and a verified assistant may receive read-only information limited to named accounts and a defined period. A stronger mandate could permit low-value payments to established beneficiaries whilst new recipients or larger transactions require independent human confirmation and permissions would be revocable, time-limited and recorded for audit. Essentially, such a model preserves a meaningful approval boundary without forcing a person to supervise every routine calculation.
Meanwhile, the technological challenge is not simply to identify the software, since an agent may genuinely represent a customer but still be acting outside its mandate. Banks and fintechs would need to distinguish identity, authority, transaction purpose and execution risk. Fraud screening would have to consider the recipient, device, unusual behaviour and the consequences of a mistaken transfer. Importantly, an AI system should not be allowed to approve its own expansion of privileges. Certainly, there are implications for programmable money and digital assets: a self-custodied wallet can permit software-driven transactions without asking a commercial bank for approval, which may encourage experimentation in automated treasury, settlement and investment. But the absence of a bank gatekeeper also removes some familiar opportunities for intervention: smart-contract permissions, lost keys and irreversible transfers can expose users to risks that a conventional bank might otherwise help contain. The competitive alternative must therefore provide safeguards, not merely freedom from restrictions.
For businesses, reliable agents could reduce treasury administration, reconcile payments and direct spare cash more efficiently. Fintechs might build neutral services that compare products across institutions instead of favouring one bank’s offerings. In the meantime, banks could develop their own agent interfaces, thus preserve customer relationships and generating new fee income. Whether this expands competition depends partly on interoperability: a bank-controlled assistant that works only with its own products may offer convenience without meaningful customer choice.
The wider economy could benefit if better information and lower switching costs improve competition for savings and business finance. Conversely, rapid movement of deposits could increase the liquidity-management costs of banks and create new concentration risks if millions of agents depend on a few technology providers. Notably, the Financial Stability Board’s work on AI-related vulnerabilities highlights third-party dependencies, cyber risk and potentially correlated behaviour. Efficiency gains therefore need to be considered alongside financial stability - the sustainable outcome is unlikely to be either unrestricted machine access or a permanent prohibition on action. It is more plausibly a negotiated architecture in which banks expose secure functions, fintechs compete to provide useful agents and customers grant specific, revocable authority. But such arrangements would need transparent liability rules, independent verification and effective recourse when things go wrong. As Prashant Kher, partner at EY, has reported: “The rail may become invisible to the client, but the experience will not. The winners will embed multi-rail payments into trusted workflows and build differentiated services above them.”
The question surely is not whether banks are frightened of AI - they have legitimate reasons to resist uncontrolled access, alongside commercial reasons to preserve deposit relationships. The test is whether restrictions evolve into safe, competitive interfaces or become enduring barriers that protect incumbents from customer choice. If banks do not establish credible ways for authorised agents to act, then fintechs and alternative settlement networks will have an incentive to build them elsewhere.
This article first appeared in Digital Bytes (17th of November, 2026), a weekly newsletter by Jonny Fry of Team Blockchain.
How can banks let AI agents manage money safely?