UK financial services AI governance is still rulebook-light

At a glance

What this is: UK financial services AI governance is being built from existing FCA, PRA and ICO obligations rather than a standalone AI rulebook, with accountability, evidence and human control doing the heavy lifting.

Why it matters: IAM, governance and risk teams need to map each AI system to a named owner, control set and evidence pack because those controls now determine whether AI use is defensible under current regulation.

By the numbers:

• The Mills Review drew on 140 written submissions before concluding that lower-autonomy AI can still fit within the existing SM&CR model.
• HM Treasury designated the first four critical third parties on 10 July 2026.
• The Data (Use and Access) Act's automated decision-making rules have applied since 5 February 2026.

Context

UK financial services does not get a separate AI rulebook, so governance has to be assembled from existing obligations and applied to systems that now make or influence decisions at machine speed. That means the question is not whether AI is regulated, but which current regime owns each use case and what evidence proves control.

The identity and governance problem is practical rather than abstract: every material system needs a named accountable owner, a record of its autonomy level, and proof that human intervention is real rather than theoretical. The FCA, PRA and ICO each look at that evidence through a different lens, which is why firms need one operating model rather than three disconnected ones.

Key questions

Q: What should firms do first when AI governance is still covered by existing rules?

A: Start with a complete inventory of AI systems, assign a named owner to each one and record the business decision it affects. Once the inventory exists, classify systems by autonomy and customer impact so the control set matches the real risk. Without that baseline, firms cannot evidence accountability, oversight or safe escalation when a supervisor asks for proof.

Q: Why does higher autonomy make AI harder to govern under current financial services rules?

A: Higher autonomy weakens the assumption that a human can meaningfully review or stop the decision before it takes effect. That matters because current obligations rely on named accountability, reasonable steps and evidence of human control. When a system can approve or act on its own, firms must prove where oversight begins, not just that someone approved deployment.

Q: What are the signs that AI governance is failing in the enterprise?

A: Common warning signs include rapid growth in AI use without matching policy coverage, sensitive files being copied into personal accounts, and a large share of AI apps carrying high or critical risk. Another indicator is weak visibility into who is using which tools and what data they are sending. If teams cannot answer those questions, governance is not working as intended.

Q: Which is more important for regulated AI: model approval or decision accountability?

A: Decision accountability is more important because regulators care about outcomes, not just model sign-off. A model can be approved and still operate outside the intended control boundary if ownership, oversight and contestability are weak. Firms need both, but accountability for the customer-facing decision is the control that ties governance to regulation.

Technical breakdown

How AI autonomy levels change the governance burden

The Mills Review's autonomy scale is useful because it separates systems that assist staff from systems that act with only oversight. A drafting assistant may fit existing review processes, but an approver or observer creates a harder question: who actually controlled the outcome? In governance terms, the control focus moves from approval of the output to evidence of meaningful human involvement, decision traceability and override capability. That is why lower-autonomy systems can often sit inside existing SM&CR processes, while higher-autonomy systems force more explicit controls over delegation, monitoring and stop conditions. See the AI Governance Guide.

Practical implication: Map each AI system to an autonomy level and require different evidence for each tier.

Why model risk management now extends into agentic AI

Model risk management was built for systems that produce outputs, but agentic AI can also choose tools, sequence tasks and persist across interactions. That changes the failure mode from incorrect prediction to incorrect action, which is a wider operational risk. Existing model governance still matters, but it must be extended with tool access controls, human oversight thresholds and post-deployment monitoring that can detect unexpected behaviour. Without that, firms may have model documentation but no control over what the system is allowed to do with it. See the AI Risk Assessment Guide.

Practical implication: Treat agentic AI as model risk plus delegated execution risk, not as a standard analytics model.

How automated decision rules intersect with AI governance

The UK GDPR automated decision rules matter because they turn AI governance into a rights and remedies question, not just a technology question. If a decision is solely automated and has a legal or similarly significant effect, the firm must be able to explain the process, allow representations, provide human intervention and support contestation. That means AI governance records must connect the model, the decision path and the responsible owner. Where special category data is involved, the evidentiary bar rises further, so compliance teams need a clear trail from input data to decision outcome. See the AI and Privacy Guide.

Practical implication: Link each significant AI decision to the legal basis, human review path and contestation process.

NHI Mgmt Group analysis

AI governance in UK financial services is now an accountability architecture, not a policy memo. The article shows that regulators are not waiting for a bespoke AI regime before expecting control, ownership and evidence. That means firms need to treat each AI system as a governed business process with a named decision owner, documented autonomy level and reviewable records. The practical conclusion is that governance fails first when accountability cannot be pointed to a person and a system in the same sentence.

Named ownership is the control that separates acceptable automation from unmanaged delegation. The FCA's reliance on SM&CR and the Consumer Duty means the governance question is not whether AI was involved, but who owned the customer impact and whether they could demonstrate reasonable steps. Where autonomy increases, the burden shifts from approval of outputs to evidence of meaningful control over system behaviour. Practitioners should therefore make ownership records operational, not ceremonial, by tying them to approvals, monitoring and escalation routes.

The real governance gap is the evidence gap created by AI that acts before humans can intervene. The Mills Review's concern about higher-autonomy systems is not just conceptual. Once an AI system approves or observes decisions on its own, firms must prove where human oversight begins and ends, which is harder than writing a policy statement. In practice, that means control design should start with intervention points and stop conditions, because those are the records supervisors will ask for when the system's behaviour matters most.

Third-party AI turns governance into a supply-chain problem as well as a model problem. The article makes clear that many AI use cases are bought rather than built, and the critical third party regime means that provider oversight is now part of the firm's own accountability chain. That creates a new governance concept, delegated AI risk ownership, where the firm remains responsible even when the system arrives inside another vendor's product. Practitioners should therefore extend vendor oversight to AI services embedded in existing contracts, not just obvious standalone tools.

AI governance and privacy governance are converging around the same control question: what can the system do without a human in the loop? The ICO's automated decision safeguards, the FCA's customer outcome focus and the PRA's model risk expectations all converge on meaningful human control, but they express it through different obligations. That convergence means identity, access and decision evidence need to line up across governance, privacy and resilience teams. The practical conclusion is that a firm cannot claim control over AI if it cannot show who can review, override and contest the decision path.

From our research library:

• 63% of organisations surveyed lacked AI governance policies to manage AI or prevent shadow AI, according to IBM's 2025 Cost of a Data Breach Report.
• Read next: AI Governance Guide

What this signals

AI governance is moving from policy design to operating evidence. The firms that will stay ahead are the ones that can show inventory, ownership and intervention records for each material AI system, not just a written policy. That evidence model is now the practical boundary between controlled automation and unmanaged delegation.

Delegated AI risk ownership: as AI becomes embedded in third-party products, the governance problem shifts from buying a tool to governing a decision path. The task for practitioners is to keep vendor oversight, human review and escalation rights visible across the full lifecycle, including decommissioning and incident response.

Only 44% of organisations have implemented any policies to manage their AI agents, despite 92% agreeing that governing AI agents is critical to enterprise security. That gap means governance programmes are still catching up to the speed at which agents are already entering business workflows.

For practitioners

• Build a live inventory of AI systems and agents Record the owner, vendor, data sources, autonomy level and customer impact for every material AI use case, including embedded third-party features.
• Classify AI systems by decision autonomy Separate assistants, collaborators, consultants, approvers and observers so that governance evidence matches the real level of delegated decision-making.
• Document meaningful human intervention points Define where a person can review, override, pause or contest a system decision, then keep evidence that those interventions are available in practice.
• Extend third-party oversight to embedded AI services Add AI services to vendor registers, due diligence and exit planning, including cases where the capability arrives inside another product.
• Prepare a system-level evidence pack Keep approvals, risk assessments, testing records, monitoring results and incident playbooks together so the firm can show reasonable steps on demand. See the AI Risk Assessment Guide.

Key takeaways

• UK financial services AI governance is being forced to fit existing regulatory structures, so accountability and evidence matter more than a new rulebook.
• The most important control gap is not model accuracy but proving who owns the system, how far it can act and when a human can intervene.
• Firms that cannot demonstrate autonomy levels, oversight and third-party control will struggle to defend AI use under current FCA, PRA and ICO expectations.

Related NHI Mgmt Group resources: AI Governance Guide · AI Risk Assessment Guide · AI Governance Platform Buyer's Guide · Agentic AI Compliance Guide

Key terms

• AI Governance: AI governance is the set of controls used to discover, classify, approve, restrict, monitor, and revoke AI-enabled access. It connects identity, data, and policy so organisations can manage what AI can reach, what it can share, and when it should be stopped.
• Meaningful Human Control: Meaningful Human Control is human oversight that is real, informed, and able to affect an AI system’s operation or output. It is not a token review step. Practitioners should define when humans must approve, override, or halt AI actions, and ensure they have the context and authority to do so.
• Autonomy Level: A measure of how independently an AI agent can decide, select tools, and execute actions. In practice, autonomy level determines how much approval, monitoring, and rollback capability the organisation needs before the agent is allowed to touch business systems.
• Critical Third Party: An external provider whose services support an important or mission critical function for a regulated organisation. If that provider fails, the customer can face operational disruption, compliance exposure, or service unavailability, so the relationship usually needs tighter oversight and stronger contractual controls.

Deepen your knowledge

The NHI Foundation Level course, the industry's only accredited NHI security programme, covers NHI governance, agentic AI identity and workload identity in a way that supports regulated operating models. It gives security, IAM and governance teams a shared vocabulary for accountability, evidence and control design.