AI in financial services: FCA good and poor practice and the road to agentic finance

On 19 October 2026, the UK Financial Conduct Authority (FCA) published examples of good practice and areas for improvement in firms’ use of AI, alongside a speech by its chief data, information and intelligence officer, Jessica Rusu. Together, the publications provide further detail on the FCA’s expectations for AI adoption, with a focus on consumer outcomes, governance and operational resilience.

These publications build on the final report of the Mills Review (see our coverage here) and HM Treasury’s Financial Services AI Adoption Plan (AI Adoption Plan) (see our coverage here). They also follow the Financial Stability Board’s (FSB) publication, on 15 October 2026, of its final report on sound practices for the responsible adoption of AI (see our coverage of the FSB’s related consultation report here).

The FCA continues to focus on applying existing regulatory requirements, rather than introducing a standalone AI-specific financial services regime. Drawing on engagement with firms and testing through its innovation initiatives, its latest publication sets out its view of good and poor practice under the existing framework. It creates no new requirements and prescribes no single governance model but provides a practical reference point for firms assessing their AI arrangements. The FCA intends to update these expectations as technology evolves.

In this briefing, we highlight the key takeaways from the FCA’s publications, including how it is taking forward the Mills Review’s recommendations, and approaching the opportunities and risks of increasingly autonomous systems.

Key points are as follows:

• Consumer Duty: firms remain responsible for the consumer outcomes delivered through AI-supported products, services, communications and support. The FCA highlights building tools “without establishing the consumer problem, target market or foreseeable harms” as a primary area for improvement.
• Governance: senior managers are expected to understand the AI risks within their areas of responsibility, rather than act as technical experts. However, the FCA warns that oversight remains an area where too many firms are still under-investing.
• Operational resilience: AI use and model dependencies should be fully integrated into operational resilience plans. Firms must stress-test AI systems, plan for failures and maintain robust fallback options, particularly when moving from low-risk to more material use cases.
• Agentic AI: existing incident reporting obligations extend to significant AI failures and near-misses. Looking ahead, the FCA will collaborate with industry, fellow regulators and frontier labs to develop a trusted framework for agentic finance.

1. The FCA takes forward most of the Mills Review’s recommendations

In her speech, Jessica Rusu reinforces the FCA’s focus on testing, assurance, monitoring and surveillance, working directly with industry to understand how AI systems are being deployed in practice. The emphasis is therefore on taking practical action and learning from real-world use, rather than relying solely on further consultation, study and reporting.

The speech confirms that the FCA is embedding six of the Mills Review’s seven recommendations into its live work. The exception is the recommendation to explore developing a public-interest AI-enabled financial capability service. Instead, the FCA intends to build on existing initiatives to help investors identify scams, verify information and challenge unreliable AI outputs.

One of the more significant recommendations in the Mills Review is for the FCA to secure and adapt the regulatory perimeter. The Mills Review recommends that, within three to six months of its publication, the FCA should consider conducting a review to assess the scale, nature and impact of general-purpose large language models (LLMs) outside the regulatory perimeter. This recommendation is also supported by HM Treasury’s AI Adoption Plan, which recommends that the FCA should undertake a comprehensive review of the consumer, competition and wider impacts of financial guidance and advice-like outputs generated by general-purpose LLMs. The speech does not specify how or when the FCA will undertake the particular review envisaged by the Mills Review.

Whilst the FCA recognises that general-purpose LLMs can significantly widen access to financial tools and information, it can blur the boundary between regulated and unregulated activity, to which the FCA’s answer is not simply to expand its remit, but to clarify the existing boundary. The FCA already provides tools such as InvestSmart and the FCA Firm Checker to help consumers understand the protections available to them.

However, the FCA recognises that these tools may be less effective when consumers encounter AI-generated financial information outside the traditional channels through which financial services are provided. Warnings may be more effective if presented at the point at which that information is delivered. The FCA is therefore calling on AI providers to help consumers understand the limits of financial information and advice generated by their models through clear, proportionate and consistent disclaimers.

The speech therefore identifies clearer communications by AI providers as an immediate priority. However, it does not include any other references to any next steps on perimeter review.

2. Good and poor practice: a practical guide to the FCA’s expectations

The FCA’s good and poor practice guidance is relevant for regulated firms that develop, deploy, procure or use AI in connection with regulated activities, consumer journeys, operations or governance. This includes AI tools developed in-house, supplied by a third party, embedded in another system or used by staff to support decision making.

Firms should read the examples as a helpful guide on the FCA’s expectations in particular scenarios, rather than a universal compliance checklist. The practical questions are whether the firm can explain why it is using AI, demonstrate the outcomes it delivers, identify who is accountable and show how it would respond if the system behaved unexpectedly or failed. Controls should reflect the nature of the use case and its potential impact, and evolve as the deployment changes. The examples cover three key areas: (i) Consumer Duty and consumers in vulnerable circumstances, (ii) governance and accountability, and (iii) operational resilience. They span a range of AI applications, from machine learning and generative AI to agentic AI, highlighting where firms are getting things right and where their approaches need to improve. Firms should use these practical examples to review their existing practices and identify potential gaps in their controls.

a. Consumer Duty and consumers in vulnerable circumstances

The FCA’s message is that technical performance alone is not enough. Firms remain responsible for the consumer outcomes delivered through AI-supported products, services, communications and support. A tool may be accurate or efficient yet still produce poor outcomes if consumers do not understand its responses, cannot obtain appropriate help or face barriers within an automated journey. Firms therefore need to assess how the service works for consumers in practice, rather than relying solely on model-performance measures.

Good practices include:

• defining the consumer need and target market before designing and deploying AI-supported products and services;
• testing outcomes across different consumer characteristics and scenarios;
• applying proportionate controls to manage AI bias and accuracy risks and ensure decisions can be explained;
• explaining to consumers how AI is being used and what its limitations are, and testing consumer understanding where appropriate; and
• monitoring live outcomes to identify and address issues promptly.

Poor practices include:

• excluding consumers in vulnerable circumstances by default, without providing appropriate alternative support;
• using guardrails that address one risk while creating another or steering consumers towards inappropriate outcomes;
• overstating an AI system’s capabilities; and
• creating inflexible automated customer journeys.

b. Governance and accountability

The FCA highlights five areas for firms to consider: (i) adopting proportionate governance models that cover the full AI lifecycle, (ii) ensuring ongoing monitoring, clear escalation routes and trained oversight functions, (iii) ensuring senior managers take reasonable steps to understand the AI risks within their areas of responsibility, seek appropriate information and expertise where needed, and ensure effective oversight, challenge and risk-appropriate governance and escalation arrangements, (iv) maintaining meaningful human oversight, rather than treating it as a box-ticking exercise, and (v) setting a clear risk appetite to guide decisions on where and how AI is deployed.

The FCA does not expect senior managers to be technical specialists, but expects them to take reasonable steps to understand the relevant risks and obtain appropriate information and expertise where necessary.

Good practices include:

• assigning clear SMF ownership for AI risks and giving senior managers targeted training so they can understand AI risks and discharge their responsibilities effectively;
• deciding which activities AI systems to support or automate, and what proportionate checks and human oversight are needed; and
• setting clear thresholds for when human intervention and escalation are needed, and defining how AI outputs can be challenged or overridden.

Poor practices include:

• using informal or insufficiently adapted governance arrangements that do not adequately test, challenge and approve AI use cases;
• senior managers delegating responsibility for AI development entirely to technical teams without sufficient understanding or effective challenge; and
• relying excessively on staff to oversee AI outputs, without addressing underlying model risk or governance responsibilities.

The FCA’s emphasis on records, explanations and escalation also has practical significance for complaints, investigations and disputes, where firms may need to reconstruct how a decision was reached and what action followed.

c. Operational resilience

The FCA’s key focus is on guiding firms to make sure their AI deployments are resilient and cyber-secure. The emphasis is on treating AI-related risks as part of firms’ wider operational resilience arrangements, rather than as a purely technical issue.

Good practices include:

• piloting lower-risk use cases before full deployment;
• mapping dependencies on AI models and providers into operational resilience planning, where applicable, including testing how important business services would be affected if an AI system failed or became unavailable; and
• proactively managing third-party relationships, including monitoring model updates, maintaining viable alternatives and planning for provider changes or exit.

Poor practices include:

• failing to account for AI dependencies in operational resilience assessments and contingency plans, where applicable;
• overreliance on certain AI providers and no exit options; and
• relying on vendors’ assurances without challenging or testing them.

3. AI agents: a new risk profile or just automation by another name?

Agentic commerce is becoming an increasingly prominent part of the AI adoption debate. Firms now need to consider not only how their customers may use AI agents to search for products, manage investments and initiate payments, but also how they may deploy their own agents to facilitate these activities.

Rusu’s speech emphasises the FCA’s view that “greater autonomy also changes the risk.” This acknowledges that agentic AI can introduce a different risk profile from existing forms of automation. The speech sets out that “[p]oorly governed AI can create consumer harm, cyber vulnerabilities and operational failures, potentially at much greater speed and scale.” The FCA therefore urges firms to invest in cyber security and operational resilience now. Rusu also highlights significant AI failures and near-misses as a reporting priority, against the backdrop of firms’ existing incident notification obligations.

The FCA also highlights the need to rethink assurance. An AI system that is safe at deployment may behave differently when operating at scale, under stress or as the underlying models change. Agentic systems therefore require continuous, conditional assurance, rather than a one-off assessment before deployment.

Importantly, Rusu’s speech confirms that the FCA intends to work with industry, other regulators and frontier AI labs to develop a “trusted framework” for agentic finance, addressing key issues such as agent identity, delegated authority, accountability and execution. This is a welcome development, consistent with the direction of travel set out in both the Mills Review and HM Treasury’s AI Adoption Plan. A trusted framework could help build confidence among financial services firms in participating in agentic commerce and agentic finance.

This work responds to a gap in the current UK framework: there is not yet a settled, generally adopted and interoperable framework for verifying AI agent identities, or linking them to the person responsible for their actions. Firms may therefore need to assess whether their existing credentials, contractual arrangements and verification processes adequately identify the relevant person and support reliance on agent-generated instructions.

The Mills Review identifies two approaches as the most viable for establishing this trusted framework: the building of a “trusted agent protocol” that firms would follow when designing and deploying AI agents, and the development of “trusted agent standards” into Open Finance. It recommends the latter, while recognising that the approaches could operate sequentially or in parallel, helping to establish the technical and governance foundations for wider adoption while ensuring that accountability remains clear and enforceable.

HM Treasury’s AI Adoption Plan further suggests that the trusted framework would need to address a range of foundational issues, including legal and liability frameworks, know your agent protocols, and authentication and governance. Legal and liability frameworks for AI agents are likely to be among the most challenging issues to resolve if financial services firms and consumers are to adopt increasingly autonomous AI use cases with confidence. The private law position is also developing. The UK Jurisdiction Taskforce’s legal statement (covered in our post) concludes that existing English law principles remain the starting point for liability for AI-related harm, while recognising that applying them to autonomous systems may raise difficult, fact-specific questions. As AI systems become more autonomous, questions of attribution, causation and the allocation of loss are likely to bring the application of those principles into sharper focus.

An important question is how the FCA’s proposed framework will interact with existing industry trust protocols and standards already being developed or used in agentic commerce. Whether the FCA builds on these initiatives or seeks to establish additional requirements will be an important area to watch as firms develop their strategies for agentic AI.

4. The FCA’s ambition to become an agentic regulator

Rusu’s speech highlights the FCA’s focus on using AI to transform its own operations, with a longer-term ambition to become an “agentic regulator”. The aim is to use the same technologies reshaping financial services to make supervision more targeted, timely and effective.

The speech identifies concrete projects, including a “market first responder” agent being developed to supervise around 9,000 small wholesale markets transaction-reporting firms, enabling supervisors to prioritise more complex activities.

The FCA is exploring how technology can improve the way it collects and uses regulatory data. Following the launch of its Handbook API in August, which allows firms to access machine-readable rules through their own AI tools, the FCA is now developing APIs for machine-readable data collection. It also sees an increasingly important role for its AI Lab and Sandbox in supporting system-wide simulations and developing new approaches to supervision.

More automated and data-driven supervision could enable the FCA to identify emerging risks and potential non-compliance earlier, and to target its supervisory attention more precisely. This places further practical emphasis on firms’ ability to identify, escalate and report issues promptly.

5. Looking ahead

Neither the speech nor the good and poor practice guidance changes the rules that apply to firms today. However, firms, AI developers and providers should keep a close eye on several developments in the coming months:

• The proposed perimeter review: the speech does not set out the timing or scope of the specific review envisaged by the Mills Review. Further detail on how the FCA takes this recommendation forward will be critical in understanding which activities need to be regulated and what rules they would be subject to.
• The trusted framework for agentic finance: further detail will help firms assess how the framework might apply to their own deployments of and interactions with agentic AI systems, especially in the context of agentic finance and agentic commerce.
• Expectations for AI developers and providers: the FCA has called on AI providers to help consumers understand the limits of financial information and advice generated by their models through clear, proportionate and consistent disclaimers. This signals the FCA’s interest in influencing practices beyond the regulatory perimeter, particularly where AI-generated outputs may influence and shape consumers’ financial decisions.