AI Agents in Banking 🚀: Security, Control & Regulation Now13 min read2,529 words

AI Agents in Banking 🚀: Security, Control & Regulation Now

Discover why AI agents in banking are being limited—security, control, and AI regulation. Learn how agentic finance reshapes automated payments safely.

AI agents in bankingagentic financebanking securityAI regulationautomated payments
AI Agents in Banking 🚀: Security, Control & Regulation Now

Why Banks Are Restricting AI Agents - and What Comes Next

AI agents could soon do the boring parts of personal finance - checking balances, comparing savings rates, paying bills, and moving money between accounts. The catch is that this only works if agents can access sensitive banking data and, eventually, be allowed to take action. Banks aren't comfortable with either step yet.

Major UK lenders have restricted or blocked some third-party agents from accessing customer accounts, especially when an agent could initiate transactions. This isn't really a "what if" fear. There's no sign that agents are already running customers' accounts at scale. The worry is what happens when access arrives before the basics - security, clear responsibility, and dispute processes - are in place.

There's also a competition issue. An agent that compares accounts and switches a customer to a better rate in seconds could chip away at the advantage banks get from customer inertia. And if lots of agents move deposits at the same time, the impact could spread beyond a single bank. This piece looks at the security and business pressures, the risk to financial stability, and the controls that could make a slower rollout possible.

Why are banks restricting AI agents?

A bank account isn't just a place to view information. It's where money can move, bills get paid, and financial obligations change. That means giving an external agent access creates a different kind of risk than letting an AI tool summarize a statement or draft a customer-service message.

Banks have practical reasons to be cautious. An agent might misunderstand an instruction, act on incorrect information, or be thrown off by misleading content it encounters while browsing. There's also the human angle: a customer could effectively grant an agent too much access by sharing login details, bypassing the protections the bank built for real account holders. And if an agent makes a payment the customer didn't intend, it may be unclear - customer, bank, agent provider, or someone else - who should take the hit.

This isn't an argument that agents are automatically unsafe. It's about drawing a clear line between using AI and authorizing AI to act. A model that helps a bank employee prepare a report doesn't touch customer funds directly. An agent that can initiate transfers does - and it can do so at scale, especially if it acts without a confirmation step.

Why account access is not a simple login problem

Traditional digital banking assumes a person is at the other end of the session. Authentication, confirmation screens, and fraud monitoring are built around that assumption. An agent may run a chain of tasks quickly and at scale, which doesn't fit those older guardrails.

One temptation is to hand the agent a customer's credentials. In practice, that can make it harder to pin down who initiated an action and what permissions were in effect. A safer setup is delegated access: the customer authorizes a specific agent, sets limits, and can revoke access without giving out a password.

Useful permissions could include:

  • Which accounts and data the agent can access.
  • Whether it can prepare actions, submit them, or both.
  • Transaction limits and approved recipients.
  • When the customer must confirm an action.
  • How the bank logs activity and handles revocation.

These controls won't prevent every mistake. What they do is make the authority clear, bounded, and trackable - so when something goes wrong, the bank can actually sort out what happened.

Why banks need common controls

If every department or product team builds its own agents, you end up with inconsistent rules for access, security monitoring, and oversight. Emirates NBD's Leap framework is one example of a bank-focused approach meant to govern authentication, sensitive data access, action permissions, audit trails, and human oversight across agent systems.

The principle matters more than any specific framework: an agent shouldn't get broad power just because it can technically use a banking interface. Access needs to be tied to the task and the likely consequences of failure. Banks are right to clamp down on poorly governed access - but a blanket ban probably won't hold if customers start expecting automated financial services as a baseline.

How could agents change banks' competitive position?

AI agents can change how customers choose and use financial products. A person might put up with a mediocre savings rate because switching takes time: comparing providers, opening an account, transferring money. An agent could monitor rates, check eligibility, and prepare the switch with far less effort.

That lowers the real cost of shopping around. Edgars Nemše, chief executive of GenLayer Foundation, has argued that agents could help people find and act on better offers quickly as they appear. For banks, it's not only that customers might leave. It's that the everyday relationship could shift - from the bank's app to an independent agent that represents the customer across multiple providers.

The customer interface is a strategic asset

Banks have poured resources into their mobile apps, loyalty features, and direct relationships. If a customer delegates routine decisions to an external agent, the bank risks becoming just the plumbing behind a service it doesn't control. The agent could decide where cash sits, which card to use, or when to refinance - while the bank provides the accounts and payment rails.

That shift isn't automatically bad. Better comparisons and fewer repetitive tasks can genuinely improve the customer experience. But it creates a new question of control: who sets the defaults, ranks offers, and influences the customer's choices?

Banks may respond by building their own agents or by offering secure interfaces for third-party agents. Either way has trade-offs. A bank-built agent can fit existing risk and compliance processes, but it may not compare offers in a fully neutral way. An independent agent may offer wider choice, but it needs reliable access, clear commercial incentives, and rules for handling conflicts of interest.

From assistance to delegated decisions

The most likely path looks staged, not a sudden handover of someone's finances. First, a customer might ask an agent to review spending or identify subscriptions. Next, it might prepare bill payments for approval. Only once the customer trusts it would they allow small, recurring payments without confirmation.

That matters because customers can judge an agent's performance before granting more authority. It also gives banks time to learn which tasks cause errors and how customers react when they do.

Corporate treasury is another early use case. Businesses already manage recurring payments and cash positions, but any agent touching those funds still needs company-specific approvals and an audit trail.

For banks, the competitive response isn't "add an AI chatbot." It's deciding what parts of the customer relationship they keep, what they can safely automate, and how they support agents without giving up control over sensitive actions. That's a product and operating-model decision, not a one-off feature release.

Could AI agents create systemic financial risk?

The concern doesn't stop at individual errors. If many agents chase the same objective at once, they could move money faster - and more uniformly - than humans usually do. Torsten Slok, chief economist at Apollo, has warned about an "agentic bank run," where automated systems shift household cash from low-yield accounts into higher-return options at scale.

The basic mechanism is familiar. Banks rely on deposits as funding for lending. If deposits leave quickly, affected institutions may have to replace that funding or adjust their balance sheets. The risk isn't that everyone withdraws cash immediately. It's that automated systems could make similar decisions in response to the same rates, news, or market signals, speeding up what might otherwise unfold more gradually.

Speed can turn individual choices into a collective event

An agent that moves one customer's money to a better account may be doing exactly what it's supposed to do. But thousands of agents making the same choice at the same time can produce a different result. This is a coordination problem: each system optimizes for its own customer, and no single agent accounts for the combined effect of all those decisions.

Automation can also create feedback loops. A market event might prompt agents to hunt for higher yields or move funds to institutions they judge as safer. Those moves can then affect what other agents see and recommend. That doesn't mean a bank run is inevitable. It does explain why speed, scale, and shared decision rules matter - not just whether one agent is reliable.

JPMorgan Chase chief executive Jamie Dimon has cautioned against unrestricted external access to bank accounts, warning about rogue agents. The practical lesson isn't "agents should never move money." It's that banks and regulators need to think about what permissions, transaction limits, and monitoring look like when activity is automated across many accounts.

Payment rails matter, but they don't solve the problem by themselves

Faster payment systems and blockchain-based settlement could make transfers more efficient. They could also reduce the time available to detect and stop an unwanted transaction. A stablecoin or other programmable payment instrument might change how settlement works, but it doesn't answer the bigger questions: who authorized the agent, who bears the loss, and how a customer disputes the action.

So the design question isn't just whether money moves over a traditional bank rail or a blockchain. It's whether the system has solid identity, authorization, limits, monitoring, and recourse. Faster settlement can improve the experience, but the controls around the instruction need to be just as dependable as the rails carrying the transfer.

What controls should banks put in place?

The right level of oversight depends on what an agent can actually do. The Bank of England and the Financial Conduct Authority reported in a survey that 80% of responding firms had deployed AI agents, while only 2% used agents with broad autonomy. Firms pointed to risks including data security, dependence on third parties, and hallucinations. That gap between deployment and broad autonomy is telling: companies are experimenting, but many are still keeping decision-making tightly constrained.

On November 4, the US Office of the Comptroller of the Currency issued a supervisory framework for banks' use of what it calls "super intelligence," including agentic systems. The framework is meant to help examiners evaluate how institutions use the technology; it doesn't create brand-new binding requirements. Still, the logic is straightforward: how much scrutiny you apply should reflect the consequences of an action, the agent's autonomy, and the level of human oversight.

Match permission to consequence

A practical control model starts by classifying actions, not agents. An agent that reads transaction history poses a different risk than one that can pay a bill, add a recipient, or move a large balance. Permissions should rise only when the use case justifies it - and when the control system can support it.

A staged policy could look like this:

  1. Read-only access: the agent can summarize accounts or identify patterns, but can't initiate transactions.
  2. Prepare for approval: it can draft a payment or transfer, while the customer reviews and confirms it.
  3. Limited autonomy: it can execute defined, low-value recurring actions within approved limits.
  4. Higher-impact decisions: larger or unusual transactions require stronger checks, confirmation, or human review.

Exact thresholds will vary by customer, product, and regulation. The main point is to avoid treating "AI access" like a switch that's either on or off.

Build monitoring and dispute handling before scaling

Agents can generate a lot of activity with little human involvement. If dispute processes were built for occasional, manual payments, they may struggle when automated actions become common. Banks should test not just whether an agent can complete a task, but whether the bank can reconstruct what happened when something gets disputed.

That means keeping records of the customer's authorization, the agent's permissions at the time, the instruction it received, the action it took, and the checks or approvals along the way. Banks also need a clear way to stop an agent, revoke access, and investigate a suspected mistake. A technically successful payment isn't proof that the customer intended it.

Third-party dependency is another operational issue. If an agent provider changes its model, has an outage, or suffers a security incident, the bank needs to know which services are affected and how to suspend access quickly. Controls have to cover the full chain, not just the bank's internal systems.

Conclusion: adoption depends on controlled access

Banks have real reasons to be cautious about agents that can act on customer accounts. Security failures, unclear liability, and high-volume disputes are all practical operational risks. But restricting access indefinitely won't preserve the traditional banking interface if customers start delegating routine work elsewhere.

The workable middle ground is controlled delegation: permissions tied to specific tasks, limits matched to financial consequences, records that support investigation, and clear ways to stop or dispute an action. Banks that build those controls can test useful automation without treating every agent as either harmless software or an unacceptable threat. The real question isn't whether AI belongs in banking. It's how much authority to grant, under what conditions, and who remains accountable when the system gets it wrong.

Frequently Asked Questions

Banks are concerned about account security, incorrect or unauthorized transactions, responsibility for losses and the limits of existing monitoring and dispute processes. Some have restricted third-party agents, particularly where they could act without customer involvement. There is no evidence that such agents have already been accessing UK accounts at scale.
An agentic bank run is a scenario in which many AI agents move customer deposits at once, perhaps from low-yield accounts to better-paying alternatives. Each agent may be acting for an individual customer, but coordinated movement could put pressure on bank funding and create wider financial-system effects.
The safer approach is delegated access controlled by the bank, rather than sharing login credentials. Customers should be able to limit which data an agent sees, what actions it can take, how much it can move and when approval is required. Access should also be revocable and recorded for later review.
Not necessarily. Agents may take over routine comparisons and transactions, but customers will still need ways to review permissions, resolve disputes and make consequential decisions. Banks may remain essential as account and payment providers even if an independent agent becomes the main interface for some customers.

Related articles