Why agentic AI widens the accountability gap in UK finance

Agentic AI does not simply generate an answer and leave a person to act on it. It can plan a sequence of steps and execute actions inside a workflow — approving a payment, declining a claim, closing a compliance alert — though the degree of autonomy varies: some firms require approval before each action, others let an agent act and log the result for later review, and a smaller number remove meaningful human involvement almost entirely. UK financial firms are adopting this kind of system, and reported evidence points to a gap between that adoption and the controls in place to govern it.

EY surveyed 50 UK financial-services C-suite executives between March and June 2025 and found 48% said their firms were already using agentic AI, with a further 16% planning to do so within six months. In the same survey, 26% reported no or limited controls to ensure their AI systems complied with laws and regulations, and 24% reported insufficient controls against unauthorised access or corruption. The sample is small — 50 executives, not a census of the sector — so these figures are reported evidence of a problem, not a sector-wide prevalence estimate. But they point to something concrete: firms are letting systems act inside regulated processes before they can fully evidence that those actions are controlled, traceable and attributable to an accountable person.

This matters because accountability in UK financial services does not move with the technology. The obligations that applied when a person made a decision still apply when an agent does; what becomes harder is proving it.

From answers to actions

Most AI discussion in financial services has concerned systems that produce an output — a credit score, a fraud flag, a settlement figure — for a person to review. Agentic AI pursues a goal across multiple steps, with limited supervision at each one.

Three workflows show the range. A fraud agent might freeze an account rather than surface a risk score to an analyst. A claims agent might assess a submission, request evidence, and approve or reduce a settlement. A compliance agent might triage and close suspicious-activity alerts.

The accountability gap widens as a firm moves from an agent that recommends, to one that acts with after-the-fact review, to one with no meaningful human step in between — each step removes a chance to catch an error before it reaches a customer. What follows treats firm-operated agents separately from consumer-operated agents, addressed later.

What the regulator evidence shows

The most detailed UK regulatory picture comes from the Bank of England and Financial Conduct Authority's (FCA) survey of 118 regulated firms — banks, insurers, non-bank lenders, investment firms, payments firms and financial market infrastructures — published in November 2024. It found 75% of respondents were using AI, with a further 10% planning to within three years, and the median number of reported use cases was expected to rise from 9 to 21 within three years.

Within those use cases, 55% involved some automated decision-making, but fully autonomous decision-making accounted for only 2%, with a further 2% combining automation with dynamic models. Third-party implementations made up a third of all use cases, up from 17% in 2022, and 64% of risk and compliance use cases specifically. Partial understanding of the AI in use was reported by 46% of firms against 34% reporting complete understanding, with weaker understanding linked mainly to third-party models. 84% of respondents had an accountable person for their AI framework.

These data sets are not a like-for-like contradiction: they measure different things, on different samples and dates, using different definitions of autonomy. A firm can have an accountable executive on paper while still lacking controls that make that accountability meaningful.

Why existing accountability still attaches

The FCA has said it does not currently plan AI-specific regulation. Its approach, confirmed when it launched the Mills Review and restated in that review's final publication on 6 July 2026, is to rely on its existing principles-based, outcomes-focused framework while adapting supervision as AI becomes more embedded. The final review named Consumer Duty, the Senior Managers and Certification Regime (SM&CR), operational resilience and the Critical Third Parties regime as the foundation for an AI-enabled market.

That means an agent does not create a gap in the rulebook; it creates a gap in evidence. Under SM&CR, firms must maintain current statements of responsibilities and clearly identify shared or divided responsibilities among senior managers. A responsible senior manager can be held accountable for a breach if they failed to take reasonable steps to prevent or address it — harder to satisfy for an action an agent took unsupervised, unless the firm can show what it was instructed to do, what it did, and what oversight existed.

Consumer Duty, which took effect on 31 July 2023 for open products and 31 July 2024 for closed ones, applies to retail outcomes a firm determines or materially influences. Foreseeable harm can arise through an act or omission, and can involve more than one firm in a distribution chain. If an agent materially influences a retail outcome, the firm remains responsible for meeting the Duty.

Human oversight that can actually intervene

The Information Commissioner's Office (ICO), in guidance covering agentic AI specifically, says a human step satisfies accountability only if the person can genuinely challenge and change the outcome, not rubber-stamp a decision already effectively made. Older ICO guidance described this as part of a general restriction on solely automated significant decisions under data-protection law. That position changed on 5 February 2026, when the main automated-decision-making reforms in the Data (Use and Access) Act 2025 took effect, broadening when such decisions may be made using personal data. Safeguards remain: the affected person must be told about the decision, make representations, and obtain meaningful human intervention. A firm deploying such an agent needs to check the decision falls within the Act's permitted circumstances and that these safeguards are actually available, not merely documented.

The audit trail problem

The FCA has not set a universal logging specification for agentic AI. Practical evidence that can help a firm demonstrate compliance with the obligations applicable to its workflow, drawn from how the FCA expects firms to evidence mapping, testing and remediation, and from ICO guidance on explaining AI-assisted decisions, is a record detailed enough to reconstruct what happened: the instruction given, the data accessed, the tools called, any approval or override, the action taken and the outcome for the customer. This is governance practice, not a statutory checklist — the fields needed depend on architecture, rulebook and workflow.

Operational resilience and third parties

FCA operational-resilience rules required in-scope firms to identify important business services, set impact tolerances, map the people, processes, technology, information and third-party relationships supporting those services, and test severe but plausible disruption. From 31 March 2025, firms were expected to remain within those tolerances. An agent supporting an important business service belongs in that mapping.

Third-party implementation reached 64% of reported risk and compliance use cases in the 2024 survey. A third party's failure does not transfer the regulated firm's responsibility for staying within its impact tolerance. From 13 July 2026, four major cloud and technology providers were designated as critical third parties, supplementing rather than replacing each firm's own outsourcing responsibilities.

For a narrower set of firms — UK-incorporated banks, building societies and PRA-designated investment firms with approved internal models — PRA Supervisory Statement SS1/23, current from 23 April 2026, brings vendor and in-house models inside a model-risk framework with responsibility allocated to a Senior Management Function holder. This is not a universal rule for every firm or system; whether an agent counts as a "model" depends on its function.

Workflow stress tests

A fraud agent that freezes an account on a false-positive signal locks a customer out of their own money without a human check first; Consumer Duty support obligations still apply during that freeze.

The FCA's letter on Consumer Duty in general insurance identified unreasonable claims delays and harmful settlement practices as foreseeable harm. An autonomous claims agent that delays, undervalues or creates undue barriers to a claim risks exactly that. In July 2025 the FCA reported £200m in compensation to over 270,000 motorists after insurers improved motor-claims processes.

An agent that wrongly closes a suspicious-activity alert creates a risk of missed escalation or regulatory non-compliance; whether that becomes an actual breach depends on the applicable rules and facts. Reconstructable records of what the agent reviewed let a firm assess that risk afterwards.

Limits and open questions

EY's 48%-and-26% figures come from 50 executives and do not disclose which agent capabilities counted as "agentic," how many firms each executive represented, or whether use was production, pilot or experimental. The 2024 Bank of England/FCA survey predates much of the 2025–26 agentic-AI adoption wave and may understate the current position. PRA SS1/23 applies to a restricted set of firms, not every regulated business using AI.

A further, less settled question concerns consumer-operated agents. FCA research published in July 2026, surveying more than 5,000 UK retail-finance consumers, found 20% would be likely to use AI capable of acting autonomously within pre-set goals — equated by the FCA to around 11 million UK adults. When a customer's own agent instructs a regulated firm, questions of identity, authority and consent become harder to resolve, and this remains unsettled.

What to watch next

The FCA has indicated further good-and-poor-practice material on AI was expected following the final Mills Review published on 6 July 2026, and HM Treasury published its own Financial Services AI Adoption Plan on 14 July 2026. Whether that FCA material had appeared by 4 October 2026 was not established by the research behind this article; readers should check the FCA's AI hub for the current position. For primary detail on the frameworks discussed here — Consumer Duty, SM&CR, operational resilience, PRA SS1/23 or the ICO's guidance on automated decisions — go to the FCA, PRA and ICO's own published material rather than a commercial intermediary.