Respondents to the Bank of England and FCA's 2026 AI Survey Most Often Relied on Vendor Assurance and Contracts to Understand High-Materiality Third-Party AI

On October 26, 2026 the Bank of England published the report of a survey of artificial intelligence and machine learning in UK financial services that it ran with the Financial Conduct Authority. It is their fourth, after surveys in 2019, 2022 and 2024. The report calls itself "a descriptive summary of the aggregate survey results," and says its findings "should be read as evidence from participating firms, rather than as a complete assessment of AI adoption across the whole UK financial sector."

The survey received 196 responses. 105 respondents completed the full survey, and unless the report says otherwise, its figures come from those full responses. Parts of the report follow, not the whole.

Third-Party AI Systems

"Third-party AI systems have become both more prevalent and more material," the report says. Third-party implementations accounted for 39% of the AI systems respondents reported in 2026, against 33% in 2024. Among the systems that firms considered to be of high materiality, the share was 53%, against 35% in 2024. The report sets its own limit on such comparisons: they "should therefore be interpreted with care where question wording, definitions, scope or respondent populations differ."

Most respondents, the report says, "applied additional evaluation and integration processes to third-party AI systems, rather than relying on the same procurement processes as for other third-party services." AI-specific risk identification and management, and additional security and data-protection controls, were the measures reported most widely. Enhanced operational-resilience expectations, "such as joint incident testing or exit planning," were "considerably less common."

Understanding and Updating Them

The survey asked firms how they reach "a sufficient level of understanding" of their high-materiality AI systems. For systems developed internally, respondents most commonly used deep knowledge of the system, assurance of its behavior and outcomes, and constrained autonomy or bounded action spaces. For third-party systems, the report says, respondents "most frequently relied on assurance by the vendor and contractual arrangements (81%), although assurance of behaviour and outcomes and constrained autonomy were also widely reported." Deep knowledge was reported by 24% of respondents for third-party systems, against 77% for systems developed internally.

The survey also asked how firms handle updates to vendor-provided models. 48% of respondents reported always undertaking regression testing for high-materiality updates to vendor-provided models or their guardrails, "with a further 25% stating that they usually do so." 73% reported always or usually having defined rollback plans for high-materiality problematic updates. And 25% reported always or usually allowing high-materiality vendor updates "to automatically propagate into production." The report adds that "given the comparatively small number of firms that responded to this question (48 firms), these results should be treated with caution."

An update that reaches production on its own can change what a system does between one review and the next. Where that is allowed, the question is what record shows that anyone looked at the change.

Agents, Governance and Accountability

The survey asked about AI agents too. 80% of respondents reported agents deployed for decision support, and 2% reported agents deployed with broad autonomy. "AI agent deployment is widespread, but autonomy is limited," the report says.

Accountability for AI systems and outputs, it says, "most commonly sits with business area owners/users and executive leadership." Firm-wide AI policies, principles or standards were reported by 94% of respondents, while "fewer than half reported incident-management and response playbooks." Among the controls firms used to manage data risk in AI systems, "Pre-deployment review, human approval and vendor due diligence were each reported by more than 90% of respondents." Ongoing monitoring, leakage testing and periodic review or recertification "were each reported by around half of respondents."

"The most common governance challenges concerned oversight of third-party AI systems," the report says. Applying model risk management principles to third-party or vendor-supplied models was the challenge selected most often, by 54% of respondents, followed by overseeing systems provided or materially influenced by third parties, at 49%.

When a vendor's update reaches a high-materiality system, what does the firm keep to show what changed?