Artificial intelligence (AI) has moved beyond answering questions and generating content. Increasingly, AI systems are being designed to act. Agentic AI can pursue objectives, make decisions and take actions on a user’s behalf, potentially with limited human involvement in individual decisions.
For financial institutions, the debate has largely focused on their own use of AI. However, an increasingly important question is what happens when customers use AI to interact with them.
In July 2026, The Mills Review, commissioned by the UK Financial Conduct Authority (FCA), described retail financial services as moving from human-led services towards “AI-enabled, continuous and delegated services”. It anticipates consumers delegating more financial decision-making to AI systems which may “recommend actions, initiate transactions, and execute decisions within agreed parameters”. One of the review’s recommendations is to “enable the foundations for agentic finance”.1
This development could test assumptions underlying financial services regulation and firms’ controls. Existing frameworks generally contemplate an identifiable customer who gives instructions, receives information, makes decisions and can later explain or dispute them. Agentic AI introduces another actor into that relationship.
The FCA does not currently propose AI-specific regulation. It instead relies on existing frameworks, including the Consumer Duty, Senior Managers and Certification Regime, and firms’ governance and control obligations. Financial institutions therefore need to consider how those requirements apply as customers delegate increasingly significant decisions to software.
Who is giving the instruction?
Consider a customer who gives an AI agent access to a bank account with an instruction to manage the customer’s cash efficiently. The agent decides to transfer funds, convert currencies or purchase an investment.
At first sight, these may appear to be ordinary customer transactions. Agentic AI, however, exposes important distinctions between identity, authentication, authority and attribution.
Under regulation 67 of the Payment Services Regulations 2017, a payment transaction is regarded as authorised only where the payer has consented to the execution of the transaction or a series of transactions of which it forms part. Consent must be given in the form and in accordance with the procedure agreed between the payer and its payment service provider. Regulation 100 separately requires strong customer authentication in specified circumstances, including where a payment service user initiates an electronic payment transaction, subject to applicable exemptions.2,3
AI agents therefore raise a central question. Does establishing that an agent is legitimately connected to a customer also establish the customer’s authority for every action the agent takes?
The answer may depend on matters including the terms governing the customer relationship and the scope of the authority conferred on the agent. A customer might, for example, authorise an agent to optimise cash management while imposing limits on transaction values, counterparties or product types. The agent could be authenticated as legitimately acting through the customer’s account while nevertheless taking an action that the customer says fell outside its instructions.
Financial institutions will therefore need to consider not only how to authenticate AI agents but also how their authority is established, limited and revoked, and when further customer involvement should be required.
Know your customer, but what about their agent?
A related challenge arises for customer due diligence and fraud controls.
As agentic AI becomes more capable, firms may increasingly encounter interactions in which an AI agent collects information, compares products and initiates transactions for a customer. That creates a new layer between the institution and the person or entity with whom it has established a customer relationship.
This does not necessarily require a new regulatory concept of “know your agent”. Existing rules continue to focus on the customer and, where relevant, persons acting on its behalf. But firms may nevertheless need to understand aspects of the agent through which their customer is interacting.
Relevant questions include how the firm establishes the connection between the customer and the agent, what permissions the agent has, how those permissions can be changed or withdrawn, and whether the firm can distinguish legitimate autonomous conduct from fraud.
This distinction may become particularly important as agents move from retrieving information to exercising greater discretion over individual transactions.
Financial crime controls face a different type of actor
Agentic AI may also present new challenges for financial crime controls.
The Mills Review identifies the potential for AI to amplify fraud and cybersecurity risks. The implications may extend to how institutions assess customer activity for money laundering and sanctions risks.
Consider an agent instructed to make a payment to a supplier as quickly and cheaply as possible. Depending on its capabilities and permissions, the agent might select among payment methods, currencies, intermediaries or routes without the customer making each individual choice.
This could complicate transaction monitoring. An agent may take a series of individually explicable steps that, viewed collectively, produce an unusual pattern of activity. An agent optimising for cost or speed may also respond to a blocked transaction by seeking an alternative route.
The legal obligations themselves do not disappear because AI-system tools are involved. Section 330 of the Proceeds of Crime Act 2002, for example, creates a failure-to-disclose offence in the regulated sector where the statutory conditions are satisfied, including where a person knows or suspects, or has reasonable grounds for knowing or suspecting, that another person is engaged in money laundering.
Similarly, the UK Office of Financial Sanctions Implementation (OFSI) General Guidance outlines firms’ obligations under financial sanctions and OFSI’s approach to compliance.4 UK government guidance on Russian sanctions circumvention identifies red flags and sets out best practices on due diligence, safeguards and controls.5
Neither framework specifically addresses autonomous customer agents. The practical question is whether controls designed predominantly around human behaviour will remain effective when machines make transactional decisions.
In particular, firms need to consider when repeated or unusual agent behaviour should prompt human scrutiny and whether existing monitoring can distinguish legitimate autonomous customer activity from behaviour indicative of fraud, money laundering or sanctions evasion.
When something goes wrong, what is the evidence?
Agentic transactions also present a less obvious challenge involving the reconstruction of events. A customer may challenge a significant transaction by saying, “I never instructed the AI to do that.” Traditional transaction records may establish what instruction the institution received and what happened as a result. They may not, however, demonstrate why the customer’s agent took that action or whether the action fell within the customer’s instructions.
Relevant evidence may include the agent’s parameters and permissions, the instructions it received, subsequent interactions with the customer, communications between the agent and the financial institution, and records of the agent’s actions.
Responsibility for creating and retaining those records may not necessarily lie with the financial institution. However, institutions should consider what evidence they will require and retain when deciding whether to accept agent-initiated transactions.
Evidence could become important in internal investigations, regulatory enquiries, civil disputes and financial crime investigations, as well as in relation to customer complaints.
What should financial institutions be considering now?
Agentic finance remains at an early stage, and the regulatory framework will continue to develop. The FCA’s direction of travel suggests that financial institutions should not treat these issues as purely theoretical. The Mills Review found that one in five UK adults was already open to AI making decisions for them.
Firms should begin by considering the following questions:
- Can the institution identify when a customer is interacting through an AI agent?
- Can it establish and verify the scope of that agent’s authority?
- Do existing fraud, anti-money-laundering and sanctions controls adequately address autonomous agent behaviour?
- Will customer communications and protections remain effective where an AI agent receives and acts on information?
- Could the institution reconstruct an agent-initiated transaction if it were challenged months or years later?
These questions do not necessarily require a new regulatory regime, and the FCA isn’t currently proposing one. Financial institutions have already devoted considerable attention to their own use of AI and should now focus their efforts on considering how customers will use AI to interact with them. The existing concepts of consent, authority, customer understanding, due diligence and suspicion will need to operate in relationships increasingly mediated by autonomous systems.
