Of every AI system that UK financial firms told their regulators was high materiality this year, 53% was built by somebody else. Of those third-party systems, 24% of firms claimed deep knowledge of how they work. The two numbers are in the same document, four sections apart, and the Bank of England and FCA report both without connecting them.

The fourth joint AI survey of UK financial services, published 26 October 2026, drew 196 responses: “Of the 196 respondents, there were 101 larger firms and 95 smaller firms. In total, 105 respondents completed the full survey.” All responding UK banks and international banks reported using AI. That makes it the closest thing to a census of how regulated British finance actually runs this technology, and it is a survey of firms rather than an inspection, which the authors say outright: the findings “should be read as evidence from participating firms, rather than as a complete assessment of AI adoption across the whole UK financial sector.”

What moved, counted in both directions

I went looking for a clean story about dependency rising. It is not clean, and the report contains the counter-examples.

Measures that rose against the 2024 edition: adoption, 75% to “86% of respondents reported using artificial intelligence (AI), up from 75% in 2024”; third-party share of all systems, “39% of all reported AI systems, up from 33% in 2024”; third-party share of the high-materiality subset, “53% of high-materiality AI systems, up from 35% in 2024”; cloud concentration, where the top three providers go “from 72% in 2024 to 79% in 2026”; and model-provider concentration, “from 41% in 2024 to 55% in 2026”.

Measures that fell: data-provider concentration, from 30% to 28%. High materiality as a share of all systems dropped from 16% to 7%, though the report immediately notes “the absolute number of AI systems reported as high materiality in 2026 was greater than in 2024”, so that is a denominator effect rather than a retreat. And two business areas moved against the trend hard enough that the report flags them: third-party share is up everywhere except “risk and compliance, where it falls from 64% to 25%, and operations and IT, where it falls from 56% to 46%”.

So five up, three down, one that is arithmetic rather than behaviour. A sweeping claim that outsourcing rose on every axis would be false on the report’s own charts. The honest version is narrower: outsourcing rose where it carries the most weight, and fell in the two areas where firms had already outsourced most.

The comparison also comes pre-hedged by its authors, and the hedge is load-bearing: “Comparisons with 2024 should therefore be interpreted with care where question wording, definitions, scope or respondent populations differ.” The 2024 survey had 118 responses against this year’s 196, and 2024 counted “use cases” where 2026 counts “AI systems”. Every year-on-year figure above inherits that caveat.

The gap that needs no comparison

Set 2024 aside and one finding stands on a single year’s data. Firms were asked how they reach sufficient understanding of their high-materiality systems, split by who built them. For internal systems, “Deep knowledge of the system is selected by 77% of firms”. For third-party systems the same answer is 24%, and the substitute is contractual: “Vendor assurance and contractual comfort shows the opposite pattern, at 6% and 81%.”

That is the whole mechanism. Firms are not claiming to understand bought systems and failing; they have switched to a different instrument, from inspection to paperwork, for the half of their material estate that came from a vendor. The report calls it out as the top governance problem its respondents name themselves: “Applying model risk management principles to third-party or vendor-supplied models was the most frequently selected challenge (54%)”, with overseeing vendor-influenced systems second at 49%.

What the contracts are holding up

If paperwork is the control, the question is whether the paperwork has been tested. Three figures say not yet, consistently.

Exit plans. “Of the 169 reported distinct cloud and compute relationships, 37% were deemed both high criticality and hard to substitute. Only 35% of that group had tested exit plans.” Read the chart’s own breakdown and it is worse than a shortfall: “exit planning is absent for 47% of cloud/compute relationships, in progress for 18% and present for 36%.” Nearly half of the relationships firms describe as critical and hard to replace have no exit plan at all.

Resilience terms. Most firms do apply extra process to AI vendors, led by “Identification and management of AI-specific risks is most common (77%)”. But the operational half lags badly: “Enhanced operational-resilience expectations, such as joint incident testing or exit planning, were considerably less common, reported by 27% of firms.”

Incident playbooks. Of the governance components surveyed, “Incident management and response playbooks are the least common named option (49%)”, against 94% for firm-wide AI policies. Policy is near universal; the document you reach for when a vendor model misbehaves at 3am exists in half of firms.

The same shape appears inside firms’ data controls, which is the part I found most convincing because it is not about vendors at all. “Human review or approval for higher-risk use cases and predeployment review of privacy and data security are the most common, each reported by 95% of firms”, while “Periodic review or recertification is the least common control shown, at 51%”. The authors name the pattern: “Pre-deployment data risk controls were more widely reported than post-deployment monitoring and review controls.”

Where the report is harder on its subject than I would be

One number undercuts any reading that firms are sanguine about vendors, and it is the vendor-update figure: “25% of respondents reported always or usually allowing high-materiality vendor updates to automatically propagate into production.” A quarter is not nothing, but three quarters do not, 73% keep rollback plans, and 48% always regression-test such updates. That is a real control, functioning. The report then shrinks its own finding: “given the comparatively small number of firms that responded to this question (48 firms), these results should be treated with caution.” A vendor would not have added that sentence.

The rating that went the other way

Against all of the above, firms rated the danger lower. “On average, respondents rated risks as less severe in 2026 than in 2024, both currently and looking one to three years ahead.” Third-party and supply-chain risk is still among their three highest-rated current risks, and the systemic picture is sharp, with cyber amplification rated high or severe by 57% today and 80% in one to three years, critical third-party dependencies 40% rising to 63%. So this is not complacency about the category. It is a lower severity score on a larger, more outsourced, more concentrated estate, reported by the same firms that name vendor oversight as their hardest governance problem. Those can all be true together. The report does not reconcile them, and neither will I.

For context on scale outside the UK, the ECB’s written overview to the Eurogroup of 9 November 2026 states in a footnote that “More than 90% of banks under ECB supervision use AI tools and 85% use generative AI”, and that “all banks under the ECB’s supervision were asked to submit action plans that address risks related to frontier AI models by the end of October 2026.” Those plans are not public, so nothing here speaks to what they contain.

What I could not check

The providers are unnamed throughout. The report gives shares, not identities, so the claim that cloud concentration sits at 79% across three firms cannot be attached to specific companies from this document, and I have not done so. All materiality ratings are firm self-assessments, with no definition of high materiality applied by the regulators for comparability. The autonomy figures are self-reported too: 80% of respondents run agents for decision support and “only 2% reported having agents deployed with broad autonomy”, which is reassuring if the categories mean the same thing at every firm, and the report does not test that. Chart values here were read from the charts’ own alt text, because the figures are images and a plain text extract of the page returns the prose with the chart numbers missing. I did not obtain the underlying response data, which is not published.