Retail financial services firms still manage much of the customer relationship through a succession of separate events: a balance check, an insurance renewal, a mortgage application, a savings transfer or a fraud alert.

By 2030, that episodic model could give way to continuous financial management, with AI agents monitoring customers’ needs and, within agreed limits, beginning to act on them. That is the prospect set out earlier this week in the Mills Review, commissioned by the UK’s Financial Conduct Authority (FCA) to examine how AI could reshape the sector. The report argues that this transition would move AI beyond a support tool and into the machinery of product distribution, servicing and financial decision-making.

If that model takes hold, AI will spread from internal workflows into customer interfaces and the infrastructure linking providers. The effects will reach product economics, bank funding, payment permissions, fraud controls, technology procurement and the evidence firms must provide of fair treatment.

Sheldon Mills, FCA executive director and chair of the review, said: “Artificial intelligence will transform financial services by 2030. It creates significant opportunities for consumers, firms and the wider economy. This report sets out a roadmap for how industry regulators and government can prepare for the next phase of AI-driven change in our world-leading financial services sector.”

From assistance to delegation

The review maps five levels of autonomy according to how far the human steps back from the decision. At one end, a person operates an AI tool. The relationship then moves through collaboration and consultation before the human becomes an approver of actions prepared by AI. At the far end, the system operates continuously within preset boundaries, and the person watches the results.

Few products or processes are likely to reach full autonomy. What changes is the nature of the risk. When AI drafts an answer or summarises a document, firms are chiefly testing its accuracy and whether anyone should rely on it. Once an agent can switch an account, reallocate savings or initiate a payment days after receiving its mandate, consent, accountability and redress become much harder to pin down. 

Research cited by the review shows AI moving beyond isolated pilots. The Cambridge Centre for Alternative Finance’s 2026 Global AI in Financial Services Report found that 81% of firms were adopting AI at some level, with 40% at the more advanced stages of scaling or transformation. More than half of industry respondents were also piloting or deploying agentic AI, although most uses remained limited in scope and subject to human oversight.

Within a retail financial institution, the possible applications span customer support, underwriting, claims, complaints, compliance, software development, and product design. AI could connect processes that previous automation left in separate silos, using information from servicing, risk and operations to identify patterns, prepare actions and escalate exceptions. Finance teams could use the same approach to test forecasts, challenge assumptions, flag emerging variances and prepare a response.

The control challenge begins when a model continues to work after deployment. A system that passed its initial tests may deteriorate as products, fraud patterns and customer behaviour evolve. Firms will need continuous monitoring for drift, unusual outputs and declining performance, with clear records showing what the model did, which data it used and when a person intervened. Simply placing somebody “in the loop” offers little protection unless that person has the information, authority and time to challenge the system.

The bill will not stop at software licences or lower headcount. Budgets must also cover data preparation, integration, testing, specialist skills, security, assurance and ongoing supervision. Any productivity dividend will reach customers only if controls work and competition forces firms to share the benefit. 

Demand emerges before trust

To test how far people are ready to delegate, the Mills Review commissioned Yonder Consulting to survey 5,026 UK retail financial services consumers. The research found that 20% would consider using AI capable of taking autonomous action within preset goals, equivalent to about 11 million people. That rose to 28% among adults who already use AI. Willingness was higher for tools that make recommendations or seek permission before each action, at 36% and 30%, respectively.

Actual use remains tentative. Although 67% of adults use AI, only 16% use it for financial services, mainly for information or suggestions. Among those users, 24% had uploaded personal financial data or documents. Across all adults, 13% would give an AI tool real-time access to financial information, rising to 36% among existing personal-finance users.

Trust falls away as the stakes rise. Some 68% were concerned about the misuse of personal and financial data, 67% about the lack of protection if something went wrong, and 65% about the concentration of too much power among a small number of large organisations. Only around two in five correctly understood the protection available when general-purpose AI tools are used for financial advice. 

That hesitation sits alongside some stubborn gaps in the market. Only 9% of consumers use traditional financial advice, around 900,000 people are unbanked, and the review estimates that £300bn remains in low-interest accounts. Just 30% hold life or income protection. An agent able to explain products, organise debt repayments, identify idle savings and prompt action could widen access to support that is currently too expensive, confusing or difficult to obtain. 

The commercial effect depends on who supplies the agent and whose interests it serves. A bank-controlled assistant may steer towards the bank’s own products. A general-purpose platform could rank providers based on technical integrations, commercial arrangements, or paid placement. An independent consumer agent might compare a wider range of the market and counteract behavioural biases, but it would still need reliable access to data and products.

Richer data makes personalisation more precise, but it can also make manipulation harder to see. Firms could tailor products and support more closely to individual circumstances while using the same information for opaque pricing, proxy discrimination and highly personalised digital steering. An apparently neutral interface may be optimised for conversion or revenue, changing the wording, defaults and friction presented to each customer in real time. 

Hyper-switching reaches the balance sheet

The gateway to the customer could become one of retail finance’s most valuable assets. Consumers may enter the market through an operating system assistant, a general-purpose AI application, a bank app, an aggregator, or an agent of their own choosing. Whoever interprets the customer’s objective may decide which products are visible, how trade-offs are presented and when a transaction takes place.

Providers will then compete for both machine and human attention. Product terms, regulated communications and servicing processes may need to be legible to agents. Distribution spending could move towards APIs, integrations and trusted-agent protocols, while brands that once benefited from inertia may find themselves screened out before a customer sees them.

The review’s most striking financial scenario is “hyper-switching”. Under this model, an agent could continuously monitor savings rates, insurance premiums, or investment fees and move a customer whenever a better option appears. That would strengthen price competition and reduce the value of passive customer balances. It could also compress margins, shorten product lives and make retention less predictable.

For deposit-taking institutions, the implications reach funding and liquidity. If agents can shift savings instantly between banks and building societies, retail deposits may behave less like stable relationship balances and more like rate-sensitive institutional money. Lower net interest margins could push up funding costs and, in turn, the price of lending to households and businesses. Simpler deposit-funded models, including some mutuals and credit unions, may face particular pressure.

Historical behaviour may no longer be a safe guide to deposit stickiness. Faster outflows could alter stress scenarios, contingency funding plans and the value assigned to different deposit segments. Firms may respond with stronger retention offers, product bundles and loyalty features, but agents may learn to look through those structures and compare the underlying value.

Insurers and investment platforms would face the same erosion of inertia. Automated comparison may place pressure on premiums and fees, while portfolio agents could rebalance holdings or sweep cash without the provider maintaining a direct relationship with the customer. Firms whose economics depend on complexity, poor switching or customer inattention are likely to feel the change first. 

A second concentration risk sits further up the technology stack. The report maps concentration across chips, data centres, cloud providers, frontier models and mobile operating systems. Pricing, data sovereignty and exit rights become central procurement issues. Larger incumbents may negotiate better terms, while smaller firms could gain from cheaper tools but struggle with assurance and vendor bargaining power. 

Fraud moves at machine speed

The same capabilities that make a financial agent persuasive and responsive can make a scam more convincing. The review expects AI mainly to intensify familiar forms of fraud by making attacks faster, cheaper and easier to personalise. Cloned voices, synthetic identities and automated social engineering allow criminals to target more people and refine their methods at speed.

The attack path may cut across a social platform, compromised personal data, a telecoms provider, an identity service, a payment rail and several financial institutions. No participant sees the whole pattern, while the criminal moves between channels faster than firms and authorities can exchange information.

Once agents begin transacting with other agents, fraud can disappear from human view altogether. Controls will need to identify abnormal behaviour and intervene at a comparable speed. Manual review cannot absorb unlimited alert volumes, so firms will need AI for detection and triage while retaining human judgement where an intervention could freeze money, deny service or harm a customer.

Delay has a balance-sheet cost. Fraud losses, remediation, cyber investment and operational resilience become part of the financial case for AI, alongside efficiency. Shared intelligence, interoperable data and clear escalation routes will matter as much as an individual firm’s model. A widely used AI or cloud provider could also become a common point of failure, exposing many institutions to the same outage, security flaw or behavioural change. 

Beyond the immediate fraud race sits the slower, more expensive problem of quantum security. The review urges firms to identify vulnerable cryptography and plan migration to post-quantum standards, creating a multi-year capital and supplier-management programme for organisations with legacy systems.

Seven recommendations for an agentic market

The review concludes that the UK’s principles-based framework remains a sound foundation. The Consumer Duty, Senior Managers Regime, and operational resilience rules can accommodate AI while keeping people close to decisions. The strain will show as human oversight shifts towards approval and observation, common providers create correlated risks, and influential AI services sit outside the regulatory perimeter.

Clarifying where that perimeter sits is the most immediate task. Its call to secure and adapt the regulatory perimeter would require the FCA to examine how consumers use general-purpose AI for savings, investments, pensions, mortgages and debt. The proposed review, beginning within three to six months, would consider when conversational guidance starts to resemble regulated advice, arranging or promotion and whether the FCA needs further powers or guidance.

Existing rules will also need to keep pace as humans step further back from individual decisions. The call to monitor the transition to autonomous models and adapt regulatory frameworks would track movement along the autonomy spectrum and clarify how the Consumer Duty, Senior Managers Regime and model-risk expectations apply at each stage. Firms would increasingly have to evidence controls and outcomes throughout a system’s life, rather than treating validation before deployment as the end of the process.

Firm-level accountability alone will not address the risks posed by common models, cloud providers, and infrastructure. That is why the review also wants to strengthen system-wide coordination and oversight through closer cooperation between UK and international authorities, monitoring of concentrated technology dependencies and a coordinated response to AI incidents that affect several firms or jurisdictions.

Supervision would have to become more continuous as well. Under the proposal to build and adopt an AI-enabled agentic supervisory model, the FCA could use AI across authorisation, supervision and enforcement while leaving important decisions with human supervisors. More structured, timely data could help it identify shared model exposures, complaint trends, and emerging harms earlier, moving oversight beyond periodic returns and ad hoc information requests.

Regulatory capability must develop alongside the market. The report also wants the FCA to scale up its AI Lab to support AI models and system innovation in financial services, giving the regulator greater independent capacity to assess emerging technologies, engage with firms and developers earlier, and publish practical lessons. The intention is to strengthen regulatory understanding without turning the FCA into a certification body for individual products.

The market infrastructure comes next. To enable the foundations for agentic finance, agents will need to be identifiable, their permissions verifiable, and their actions traceable. Clear rules will also be required for revoking mandates, resolving liability and establishing who is responsible when an autonomous transaction produces harm. The review favours developing trusted-agent standards through Open Finance, aligned with work on data, digital identity and payments.

The final proposal addresses access. A trusted public-interest AI-enabled financial capability service would bring together government, the Money and Pensions Service, consumer organisations, and industry to provide a free, reliable source of financial information and support. Without such an option, access to useful agents may depend on subscriptions, platform choice or commercial incentives, leaving consumers with the greatest need reliant on the weakest tools.

Taken together, the seven recommendations amount to a blueprint for an agentic market: clearer boundaries, common infrastructure, stronger supervision and a public-interest route for consumers. The aim is to keep innovation moving while preventing risk from disappearing into the gaps between firms, platforms and regulators. 

Preparing before autonomy arrives

Whatever model or platform eventually dominates, firms can prepare now by identifying where delegation changes the economics and risk of their business. The first job is to map each AI use case by autonomy, customer impact and reversibility. Systems that draft or summarise can tolerate different controls from those that approve credit, reject claims, move deposits or initiate payments. Permissions, escalation and human challenge should become stricter as actions become harder to reverse.

That map should extend into the technology supply chain. Firms should know which models, cloud services, and data providers underpin critical processes, what happens when prices or terms change, and how activity can be transferred if a supplier fails. Finance, procurement, risk, and technology teams will need a common view of the total cost and concentration associated with each deployment.

Customer-facing firms must also decide how they will meet agents at the front door. That includes machine-readable product information, secure data access, identity checks and rules governing what an external agent may view or execute. Otherwise, platforms with different commercial incentives may take control of the relationship.

Balance-sheet planning should begin before hyper-switching becomes widespread. Deposit stability, insurance renewal behaviour and investment-platform flows could all change as comparison and execution become continuous. Stress tests should ask how margins, liquidity, and funding behave when agents act faster than customers have historically.

The Mills Review ultimately describes a change in market architecture. AI will sit inside operating models, mediate customer access and connect decisions to execution. Firms that treat it solely as a productivity tool may capture early savings while missing the larger disruption to distribution, funding and control.

People may still own the most important financial decisions in 2030, but AI is likely to do far more of the work around them and execute more of what follows. That leaves firms with a dual task: make delegation useful, explainable and safe, and build enough resilience for the moment when millions of agents act at once.