Agentic AI: Preparing for greater regulatory attention

The FCA’s Review into the long-term impact of AI on retail financial services (the Mills Review), published on 6 July 2026, explored how AI could reshape the sector and set out recommendations for the FCA to consider.

A central theme is the move towards greater autonomy. The Mills Review illustrates this through an AI autonomy spectrum, ranging from humans using AI as a tool, to systems capable of executing approved actions with limited human intervention.

As autonomy increases, AI systems may pursue objectives, interact with other systems and make decisions within defined parameters. This creates governance challenges that extend well beyond model performance alone.

As firms move further along the autonomy spectrum, existing governance and control frameworks are likely to come under increasing pressure, prompting greater supervisory focus on how firms manage the associated risks. Sarah Breeden, Bank of England Deputy Governor, Financial Stability noted in a speech on 30 June 2026:

'Our frameworks were not built to contemplate autonomous agents, and relying on a human in the loop for all agent actions is unlikely to be realistic. More sophisticated governance and accountability frameworks may be needed.'

While regulators consider whether the regulatory framework and supervisory approach should evolve, supervisors are likely to focus increasingly on how firms identify, manage and control risks that may emerge as AI systems become more autonomous.

In this article, we consider how greater AI autonomy may affect firms’ ability to manage AI-related risks and meet existing regulatory expectations. We focus on four key areas:

• Risk management 

• Governance and accountability  

• Security and operational resilience 

• Consumer outcomes

Risk management

Agentic AI changes the nature of risk management because firms are no longer assessing static technology tools but increasingly autonomous systems capable of making decisions, interacting with multiple systems and adapting their behaviour within defined parameters. Traditional model risk and technology risk frameworks remain essential, but they may no longer be sufficient on their own.

Rather than treating AI as a standalone technology risk, firms should consider how agentic AI affects existing risks across the enterprise. The objective should be to embed AI governance within existing enterprise risk management rather than creating parallel frameworks.

Supervisors are increasingly likely to ask:

• How has AI been incorporated into the firm's risk appetite? 

• Which AI use cases are considered material? 

• How are emerging AI risks identified after deployment?  

To strengthen governance, firms should:

• Update enterprise risk frameworks. Ensure AI-related risks are reflected within existing risk taxonomies, risk appetite statements and governance processes. 

• Maintain an inventory of material AI systems. Firms should understand where AI is deployed, the business activities it supports, its key dependencies and its accountable owners. 

• Develop meaningful risk indicators. Traditional technology metrics should be complemented by indicators covering issues such as model drift, unexpected autonomous behaviour, override rates, escalation frequency and customer outcomes. 

• Strengthen independent challenge. Risk and compliance functions should have sufficient capability to review AI deployments throughout their lifecycles, not simply at implementation. 

• Monitor continuously. AI governance should not end once a system is deployed. Firms should establish continuous monitoring capable of identifying changing behaviours, emerging risks and deteriorating controls over time. 

Ultimately, firms should move away from viewing AI governance as a one-off project and instead embed it within existing enterprise risk management processes.

Governance and accountability

As agentic AI systems operate with greater autonomy, existing governance and control frameworks may come under increasing pressure, making accountability more difficult to demonstrate. For regulated firms, agentic AI should therefore be fully integrated within existing governance and control arrangements, including the Senior Managers and Certification Regime (SM&CR).

Where AI materially influences regulated activities, important business services or customer outcomes, firms should ensure accountability remains aligned to the underlying business activity rather than ownership of the technology itself. 

PwC research found that accountability for AI-supported decisions sits with the technology function in 63% of organisations, despite the consequences often sitting with the business.

Supervisors are increasingly likely to ask: 

• Are governance and oversight arrangements proportionate to the autonomy of the system and the criticality of the activity it performs?

• Which Senior Manager Function (SMF) is accountable? 

• How have responsibilities been allocated, and what reasonable steps has the SMFs taken to oversee the use of agentic AI?

• When must AI systems refer decisions to a person?  

• How are changes approved following deployment? 

Firms should therefore:

• Allocate accountability according to the underlying activity. Where AI materially affects regulated activities, important business services or customer outcomes, responsibility should reflect the substance of the activity rather than ownership of the technology. 

• Calibrate governance and controls to autonomy and criticality. Governance and oversight should reflect both the degree of autonomy given to an AI system and the significance of the activities and decisions it performs, with clear business ownership, oversight and escalation arrangements. 

• Define delegated authority. Firms should establish what objectives systems may pursue, which actions they may take and when they must stop or refer matters to a person. SMFs should be able to demonstrate the reasonable steps they have taken to ensure these boundaries are appropriate, effectively controlled and subject to appropriate oversight.

• Maintain lifecycle oversight. Changes to models, prompts, data or connected tools may materially alter system behaviour. Firms should establish clear triggers for renewed testing, approval and governance throughout the system lifecycle. 

PwC has published further information on the central governance shifts needed in an agentic AI world of business here.

Security and operational resilience

The impact of AI on cyber resilience is already a significant area of regulatory focus. More broadly, however, increasingly autonomous AI systems may make operational disruption more difficult to identify, contain and recover from.

As firms automate more processes, fallback arrangements may prove weaker than expected. AI-driven transformation may also reduce operational capacity and specialist expertise available to perform manual workarounds when systems fail or require intervention.

The FCA and BoE explored cyber resilience challenges and controls for frontier AI including how it is accelerating the discovery of cyber vulnerabilities, in publications on 2 September 2026. Indeed, firms will need to ensure that validation, remediation and change-management capacity can keep pace with the volume and complexity of developments and findings.

Firms may also become increasingly dependent on a relatively small number of model and cloud providers, chip manufacturers and other critical infrastructure providers, creating concentration, sovereignty and supply chain risks. These dependencies as well as incidents will also need to be considered in the context of the FCA, PRA and BoE’s new operational incident and material third-party reporting regime, which comes into force in March 2027.

Supervisors are increasingly likely to ask:

• Which important business services depend on AI?

• Can services continue during AI-related disruption?

• Have resilience, fallback and recovery arrangements been tested?

• Do firms understand their critical third-party dependencies, and how prepared are key suppliers?

Firms should therefore:

• Ensure AI dependencies are captured in important business service mapping. Firms should review existing mapping to understand where AI supports service delivery and whether disruption could cause them to breach impact tolerances.

• Test degraded and anomalous behaviour. Scenario testing should assess whether unsafe behaviour can be detected, contained and investigated while continuing to deliver important services. 

• Validate fallback arrangements. Manual workarounds should be tested against the volume, speed and expertise required, particularly where AI adoption has reduced operational capacity. 

• Assess concentration risk. Firms should understand their dependence on a small number of providers and consider whether services could continue if critical suppliers became unavailable.

Delivering good consumer outcomes

As AI becomes increasingly embedded within customer communications, servicing and decision-making, firms may find it harder to demonstrate good outcomes under the Consumer Duty. PwC’s report on agentic AI in banking highlights the potential for agentic AI to deliver more personalised support and increasingly take actions on customers’ behalf. 

AI has the potential to affect all four Consumer Duty outcomes - products and services, price and value, consumer understanding and consumer support - but its most immediate impact is likely to be on consumer understanding and support. 

Generative AI may tailor explanations to individual customers, improving understanding while making assurance significantly more challenging. Firms cannot rely solely on pre-approved disclosures where AI systems can dynamically interpret, summarise or supplement information in individual customer interactions.

Similarly, agentic AI may influence customer outcomes through operational processes, including servicing requests and decision-making. Systems may technically operate as designed while nevertheless producing outcomes that are inappropriate in context or inconsistent across different customer groups.

Supervisors are increasingly likely to ask:

• How do firms evidence that AI-enabled journeys deliver good customer outcomes?

• Can firms explain and evidence AI-driven decisions and interactions?

• When is human support or intervention required?

• How are outcomes for vulnerable customers monitored and protected?

Firms should therefore:

• Test and monitor customer interactions. Assess whether communications remain clear, important qualifications are preserved and different customer groups continue to receive appropriate outcomes, using both pre-deployment testing and ongoing monitoring.

• Design for auditability. Firms should be able to reconstruct material AI interactions, understand the information used, demonstrate which controls applied and evidence compliance with approved parameters. 

• Set clear operating boundaries. Define when systems may act independently and when they must stop or refer matters to a person, including where uncertainty or customer vulnerability arises. 

• Design scalable human oversight. Human-in-the-loop controls may be appropriate for higher-risk decisions and defined escalation points, but are unlikely to be sustainable across every agent interaction as deployment scales. Firms should instead define where human judgement is required, embed appropriate monitoring and escalation triggers, and ensure people remain accountable for customer outcomes.

Ultimately, firms can only demonstrate good customer outcomes where they have clearly defined what AI systems may do, how outcomes are monitored and who is responsible for intervention.

Preparing for increased supervisory scrutiny

As firms move further along the AI autonomy spectrum, existing governance approaches may no longer provide sufficient assurance. Supervisors are increasingly likely to assess firms using existing regulatory expectations around governance, risk management, governance, accountability, operational resilience and customer outcomes.

The challenge for firms is therefore to ensure these frameworks remain effective as AI systems become more autonomous - with controls that are proportionate to the level of autonomy and capable of supporting safe deployment at scale.

Firms that build these capabilities now will be better positioned to unlock the value of agentic AI safely and at scale - giving the business confidence to extend autonomy where it delivers value, while demonstrating to supervisors that innovation is underpinned by effective governance and control.