Sebastien Rousseau

THIRD-PARTY AI RISK

Third-Party AI Is Now the Governance Problem for UK Financial Firms

AI use among surveyed UK financial firms has reached 86%, but the numbers that matter most sit beneath the headline: third parties now supply more than half of the high-materiality systems, and overseeing them is the governance challenge firms report most often.

16 min read
Banner for: Third-Party AI Is Now the Governance Problem for UK Financial Firms

AI use among surveyed UK financial firms has reached 86%, but the numbers that matter most sit beneath the headline: third parties now supply more than half of the high-materiality systems, and overseeing them is the governance challenge firms report most often.

Executive Summary

  • 86% of respondents reported using AI, up from 75% in 2024. Generative AI made up 56% of reported systems.
  • Around 7% of reported systems were assessed as high materiality, down from 16% in 2024, yet the absolute number of high-materiality systems was greater in 2026.
  • Third-party implementations accounted for 53% of high-materiality systems, up from 35% in 2024. The top three cloud/compute providers account for 79% of reported cloud/compute relationships.
  • 80% of respondents have AI agents deployed, but only 2% reported agents with broad autonomy.
  • Respondents expect both benefits and risks to rise across all categories, with cyber risk amplification and critical third-party dependencies rated the top systemic risks.
  • Oversight of third-party AI systems was the most commonly reported governance challenge.

What the survey is, and what it isn't #

The publication, the fourth Bank of England and FCA survey of AI and machine learning in UK financial services, is a descriptive summary of aggregate survey results, and it's worth holding that framing in mind before drawing any conclusion from it. The findings are evidence from participating firms, rather than a complete assessment of AI adoption across the whole UK financial sector. Nothing below should be read as a census.

The survey received 196 responses: 101 larger firms and 95 smaller firms. For the first time, firms with fewer than 100 employees could choose a shorter questionnaire, and 105 respondents completed the full survey. Unless otherwise stated, the figures in the report are based on responses to the full survey, which aids comparability with previous surveys. Percentages are calculated using valid question-level responses only, so firms that didn't answer a particular question are excluded from that question's denominator. Responses from international banks and broker/dealers are given with reference to their UK operations.

The 2026 survey retained core questions from previous surveys for continuity, while adding or amending questions to reflect developments in AI adoption and capability since 2024. That's why the report can set 2026 figures against 2024 ones for adoption, materiality and third-party share, and also why the report asks readers to interpret comparisons with 2024 with care where question wording, definitions, scope or respondent populations differ. Some topics, such as the short annex on frontier AI and cyber risk, carry no 2024 comparison at all.

These caveats matter because several of the most striking numbers below are shares of shares. A reader who ignores the denominators will mistake a change in composition for a change in volume. The survey itself offers a clean example of that trap, which we'll reach shortly.

Adoption is close to universal, and generative AI dominates #

Start with the headline. 86% of respondents reported using AI, up from 75% in 2024. That figure covers both the full and short-form surveys, and the report itself warns that the 2026 and 2024 percentages do not relate to the same population of firms and should be compared with caution. Firms are increasingly reporting use of generative AI, AI-enabled agents and third-party implementations of AI systems, and those three threads run through everything else in the survey.

Generative AI accounted for 56% of all reported AI systems. The comparison the report draws is with foundation models, which accounted for 17% of reported AI use cases in 2024. The report says these measures are not directly comparable, so it would be a mistake to call it a 39-point surge in one technology; taken together, in the report's words, they suggest a substantial increase in firms' reported use of general-purpose capabilities since 2024.

Where does the AI sit within firms? Operations and IT accounted for approximately 31% of all reported AI systems, up from around 22% in 2024, and no other individual area accounted for more than 11%. That's a striking concentration. Customer-facing and front-office use cases get most of the public attention, yet the largest single block of reported systems is in the plumbing of the business. A related figure from the small-firm section points the same way: among short-form survey respondents, operations was the most commonly reported domain for current or planned AI use, cited by 64%.

The perceived benefits line up with that picture. Operational efficiency, automation and productivity were the highest-rated areas of perceived benefit from AI today. Firms are, on the evidence of what they report, using AI mainly to run themselves more efficiently.

Adoption isn't evenly distributed either. Almost half of firms reported ten or fewer AI systems, while the ten firms reporting the most systems accounted for 43% of all reported AI systems. So the 86% figure describes breadth of use, not depth. A small number of firms carry a large share of the total, and a large number of firms are running only a handful of systems. Small firms are in the mix too: 73% of short-form respondents reported using AI, and of those, 58% reported predominantly using general-purpose large language models.

The materiality puzzle #

Here's the trap mentioned earlier. Around 7% of reported AI systems were assessed by respondents as high materiality, compared to 16% in 2024. Read quickly, that looks like AI becoming less important to firms, or less risky.

It doesn't mean that. The absolute number of high-materiality AI systems reported in 2026 was greater than in 2024. The share fell because the total grew faster than the high-materiality subset. As firms deploy more systems, most of the additions are assessed as lower materiality, which dilutes the percentage while the count of high-materiality systems still rises. Those systems are also concentrated: ten firms accounted for 62% of all reported high-materiality AI systems.

That's a useful corrective for anyone building a governance case on percentages alone. A falling share of high-materiality systems says nothing reassuring about the workload of the people responsible for them. There are more of them to oversee than there were in 2024.

Where the important systems come from #

The most consequential result in the survey is who builds and runs the systems that matter most. Third-party implementations accounted for 39% of all reported AI systems, up from 33% in 2024 and 17% in 2022. That's a steady move. The shift among high-materiality systems is much sharper: third-party implementations accounted for 53% of them, up from 35% in 2024.

So, across the reported systems as a whole, the ones firms consider most material are now more likely than not to come from outside. An 18-point rise in two years, in the category where failure would hurt most, is the clearest signal in the data. Combined with the materiality finding above, the picture is of a growing number of important systems, a majority of which firms don't build themselves.

The concentration among providers sharpens it. The three largest cloud and compute providers accounted for 79% of all reported cloud/compute relationships, and 98% of respondents reported using one of those three. Model API and hosting provision was also concentrated, though less so: the three largest providers accounted for 54% of reported relationships, and the top three cloud/compute providers were all also in the top five model API/hosting providers. Firms may have many AI systems, but they're drawing on a small set of suppliers for the infrastructure and models underneath.

Substitution is the harder question. Of the 169 reported distinct cloud and compute relationships, 37% were deemed both high criticality and hard to substitute, and only 35% of that group had tested exit plans. For the 226 model API/hosting relationships, the high-criticality, hard-to-substitute share was 14%. Enhanced operational-resilience expectations for third-party AI, such as joint incident testing or exit planning, were reported by only 27% of firms.

Respondents raised the systemic version of this concern themselves: cyber risk amplification and critical third-party dependencies were the two highest-rated systemic risks, both today and over the next three years, and respondents expected them to intensify as AI use becomes more widespread. The share rating critical third-party dependencies as a high or severe concern rose from 40% today to 63% in the next one to three years. That expectation is a forecast by respondents, not an observed outcome, but the supplier concentration figures give it an obvious basis.

Firms aren't unaware of the constraint. Concerns about information security and operational resilience were identified as a major constraint on third-party AI adoption by 40% of respondents, followed by limited transparency into third-party AI systems (training data, model design, performance limitations, failure modes) at 34%. Counting major and moderate constraints together, difficulty negotiating contractual protections was the second most cited challenge. Taken together, these describe firms leaning heavily on outside providers while saying they can't always see inside them.

Agents: widely deployed, mostly in a supporting role #

The agent numbers need careful reading, because the headline and the substance point in different directions. 80% of respondents reported having AI agents deployed, while a further 14% reported plans to deploy them over the next 12 months. On those figures, only a small remainder has neither deployed nor planned to.

But look at what the agents do. 80% of respondents reported agents in production that provided decision support, defined as AI agents recommending actions while humans decide and execute. 46% reported agents with constrained execution, acting within predefined limits and rules, and 22% reported agents with conditional autonomy. Only 2% reported deploying agents with broad autonomy. The dominant pattern is agents that advise or act within tight limits.

Customer communications is where the line is starting to move. 16% reported current use of agents for automated customer communications, rising to 34% when planned deployment was included. That more than doubles the share if firms follow through on their plans, though plans are plans, and the survey can't tell us how many will be delivered. By contrast, even including planned deployment, no more than 9% of respondents expected agents to approve or decline customer outcomes, execute trades or place orders, initiate, execute or block payments, or change limits, thresholds and controls.

My reading is that the 2% figure is the more telling number, and it deserves more attention than the 80%. Deploying an agent that suggests and waits for a person is a modest step. Deploying one with broad autonomy changes who's accountable for an outcome. Firms have so far been willing to take the first step widely and the second almost not at all. Whether the customer-communications plans signal the start of a change in that posture is one of the more interesting open questions in the data.

Benefits and risks both expected to rise #

Respondents don't see AI as a one-sided bet. They expected benefits to increase across all areas over the next three years, with the largest anticipated increases in cybersecurity and operational resilience, and in revenue generation and competitive positioning. These are expectations, and the survey presents them as such. None of the listed benefits was widely considered to be more than moderate in scale today.

The risk side is just as consistent, in the sense of direction. The highest-rated areas of perceived risk to firms from AI today were data protection and security, third-party and supply chain risk, and GenAI-specific risks. Over the next one to three years, respondents expected risk levels to increase across all categories, although on average they rated risks as less severe in 2026 than in 2024.

The source frames the overall picture in balanced terms. These developments may bring benefits including greater efficiency, improved decision support and stronger risk management. They may also create or amplify risks related to governance, operational resilience, data protection, consumer outcomes, third-party dependency and financial stability. The hedging is the source's, and it's appropriate; the survey measures perceptions and reported practice, not realised gains or losses.

One cluster of results concerns frontier AI and cyber capability. 88% of respondents assessed evolving frontier AI cyber capabilities over the next two years as having a moderate or high impact on their firm's risk profile. That's a forward-looking assessment rather than a measured effect, though 19% already named increased effectiveness of supply-chain or third-party attacks as the channel through which frontier AI currently poses the greatest risk. What firms have already done in response is easier to check: 64% selected vulnerability management and patching cadence as the area of cyber resilience that had received the greatest increase in management attention as a result of awareness of frontier AI.

Constraints: more practical than regulatory #

If you expected regulation to top the list of obstacles, the survey is more nuanced. The most cited non-regulatory constraints were difficulties integrating AI into existing business processes and decisions, and legacy systems or infrastructure, identified as large constraints by 23% and 22% of respondents respectively. Those are modest shares, which suggests that no single non-regulatory obstacle dominates, but they're ordinary engineering and organisational problems rather than exotic ones.

On the regulatory side, the most cited perceived constraints were UK data protection and privacy, and international regulation, which 42% and 29% of respondents respectively perceived as posing a high compliance burden. 28% indicated no regulatory constraints.

Small firms show a similar pattern: 30% of short-form respondents identified data protection, confidentiality and IT considerations as constraints. Data protection appears in the risk ratings, the regulatory burden figures and the small-firm constraints. If a single theme links the survey's different sections beyond third parties, it's that one.

Governance: broad on paper, thinner after deployment #

The formal structures are close to universal. 94% of respondents reported firmwide AI policies, principles or standards, and 84% reported centralised AI/model-risk governance fora. Among data risk controls, human approval, pre-deployment review and vendor due diligence were each reported by more than 90% of respondents.

Now compare the controls that apply after launch. Ongoing monitoring, leakage testing and periodic review or recertification were each reported by around half of respondents. The contrast is plain. The data risk controls that happen before or at the point of deployment, where a person signs something off, are close to universal. The ones that keep working after deployment are reported by roughly half.

I'd argue that's the central governance finding, and it connects directly to the third-party story. The survey shows how differently firms understand what they buy. For high-materiality systems, 77% of respondents reported relying on deep knowledge of internally developed systems, compared with 24% for third-party systems; for third-party systems, 81% instead relied on vendor assurance and contractual arrangements. A system that's bought rather than built is one where pre-deployment review and vendor due diligence are the obvious levers. Monitoring and recertification depend on visibility into a system, and limited transparency into third-party AI systems was already flagged by 34% of respondents as a constraint. The survey doesn't state this causal link, and the data can't prove it, but the pieces fit together uncomfortably well.

The firms say as much themselves. Oversight of third-party AI systems was the most commonly reported area of governance uncertainty or implementation challenge. Applying model risk management principles to third-party or vendor-supplied models was selected by 54% of respondents, and overseeing systems provided or materially influenced by third parties by 49%. Roughly half of firms, then, report a challenge in applying model risk management principles to vendor-supplied models, and roughly half in overseeing systems that third parties provide or materially influence.

Put the three threads side by side. Third parties supply a majority of high-materiality systems. A handful of providers dominate the underlying relationships. And the firms' own governance, strong at the gate, is thinner on continuing oversight, with the third-party case flagged as the hardest. That isn't a criticism of any firm; it's a description of where the reported effort sits against where the reported exposure is.

What to take from it #

For technical leaders, the practical reading is that buying has overtaken building for the systems that matter, at least in this sample: 53% of high-materiality systems are third-party implementations. The engineering effort therefore shifts to what can be done around a system you don't control: contractual transparency, testing at the boundary, and monitoring that doesn't depend on access to internals. The survey doesn't prescribe any of these, and none should be mistaken for a finding. They're the questions its numbers raise.

For founders selling AI into financial services, the data suggests where friction will be. Respondents report information-security and operational-resilience concerns (40%) and limited transparency (34%) as major constraints on third-party AI adoption. A vendor that can answer those two concerns directly is addressing what a substantial share of respondents say holds them back. And because more than 90% of respondents report vendor due diligence as a data risk control, a supplier should expect to be examined.

For anyone tempted to read the 86% adoption figure as a sign that the hard part is over, the survey offers a more sober view. Adoption is broad, but a few firms carry a large share of the systems, agents are mostly in supporting roles, and respondents expect risk to rise in every category over the next one to three years. Respondents also expect benefits to rise in every area, so neither optimism nor caution has the evidence to itself.

The caveats from the start still apply. This is a descriptive summary of 196 responding firms, only 105 of which completed the full survey, and it isn't a complete assessment of the UK financial sector. Within those limits, the pattern is consistent enough to be useful: the systems firms consider most important increasingly come from outside, and the oversight they report is weakest where that dependence is greatest.

The question the data raises but can't answer is what happens at the autonomy boundary. With 34% of respondents either using or planning agents for automated customer communications, against only 2% deploying agents with broad autonomy today, the next survey will show whether agents stay in supporting roles when the systems behind them are, more often than not, supplied by someone else.

Frequently Asked Questions #

What is the Bank of England and FCA survey of AI in UK financial services? It is the fourth joint survey of artificial intelligence and machine learning in UK financial services, following surveys in 2019, 2022 and 2024. The 2026 edition received 196 responses, 105 of them to the full questionnaire. The authors present it as a descriptive summary of evidence from participating firms, not a complete assessment of the UK financial sector.

How much of firms' AI comes from third parties? Third-party implementations accounted for 39% of all reported AI systems in 2026, up from 33% in 2024 and 17% in 2022. Among high-materiality systems the share was 53%, up from 35% in 2024.

How concentrated are AI providers? The three largest cloud and compute providers accounted for 79% of reported cloud/compute relationships, and 98% of respondents used one of them. The three largest model API/hosting providers accounted for 54% of reported relationships in that category. Of the 169 reported cloud and compute relationships, 37% were deemed both high criticality and hard to substitute, and only 35% of that group had tested exit plans.

What governance challenges do firms report with third-party AI? Oversight of third-party AI systems was the most commonly reported area of governance uncertainty or implementation challenge. 54% of respondents selected applying model risk management principles to third-party or vendor-supplied models, and 49% selected overseeing systems provided or materially influenced by third parties. For third-party high-materiality systems, 81% relied on vendor assurance and contractual arrangements, while only 24% reported deep knowledge of the system.

Are AI agents acting autonomously in UK financial firms? Rarely. 80% of respondents reported having AI agents deployed, and 80% reported agents in production for decision support, where agents recommend actions and humans decide and execute. Only 2% reported agents with broad autonomy, and even including plans, no more than 9% expected agents to approve or decline customer outcomes, execute trades or place orders, initiate, execute or block payments, or change limits.

References #

Last reviewed .

Syndicate this article

Format for Medium

# Third-Party AI Is Now the Governance Problem for UK Financial Firms

> Originally published at [https://sebastienrousseau.com/2026-11-21-third-party-ai-governance-uk-firms/](https://sebastienrousseau.com/2026-11-21-third-party-ai-governance-uk-firms/)

Third parties supply 53% of high-materiality AI systems at surveyed UK financial firms, and overseeing them is the governance challenge firms cite most often.

Read the full article on sebastienrousseau.com: https://sebastienrousseau.com/2026-11-21-third-party-ai-governance-uk-firms/

Format for Mastodon

Third-Party AI Is Now the Governance Problem for UK Financial Firms

Third parties supply 53% of high-materiality AI systems at surveyed UK financial firms, and overseeing them is the governance challenge firms cite most often.

https://sebastienrousseau.com/2026-11-21-third-party-ai-governance-uk-firms/

Copy formatted for LinkedIn

Third-Party AI Is Now the Governance Problem for UK Financial Firms

Third parties supply 53% of high-materiality AI systems at surveyed UK financial firms, and overseeing them is the governance challenge firms cite most often.

Here are the key strategic takeaways:

- Adoption is near-universal. 86% of respondents reported using AI, against 75% in 2024, though the two populations differ.
- Dependence is shifting outward. Third parties supplied 39% of all reported systems and 53% of high-materiality ones.
- Supply is concentrated. The three largest cloud and compute providers accounted for 79% of reported cloud/compute relationships.
- Governance has a gap. Over 90% report human approval, pre-deployment review and vendor due diligence as data risk controls, but only around half report ongoing monitoring.
- Read with care. The findings come from participating firms, not the whole UK financial sector.

What is your organisation's approach to the challenges outlined in this piece?

→ https://sebastienrousseau.com/2026-11-21-third-party-ai-governance-uk-firms/

#ThirdPartyAiRisk #AiGovernance #BankOfEngland #Fca #UkFinancialServices

Sebastien Rousseau | CC-BY-4.0
Cite this article

Third-Party AI Is Now the Governance Problem for UK Financial Firms

Third parties supply 53% of high-materiality AI systems at surveyed UK financial firms, and overseeing them is the governance challenge firms cite most often.

BibTeX

@online{rousseau2026third,
  author  = {Rousseau, Sebastien},
  title   = {{Third-Party AI Is Now the Governance Problem for UK Financial Firms}},
  year    = {2026},
  url     = {https://sebastienrousseau.com/2026-11-21-third-party-ai-governance-uk-firms/},
  urldate = {2026}
}

RIS

TY  - GEN
AU  - Rousseau, Sebastien
TI  - Third-Party AI Is Now the Governance Problem for UK Financial Firms
PY  - 2026
UR  - https://sebastienrousseau.com/2026-11-21-third-party-ai-governance-uk-firms/
ER  -

Vancouver

Rousseau S. Third-Party AI Is Now the Governance Problem for UK Financial Firms. sebastienrousseau.com. 2026 Nov 21. Available from: https://sebastienrousseau.com/2026-11-21-third-party-ai-governance-uk-firms/

Chicago

Rousseau, Sebastien. "Third-Party AI Is Now the Governance Problem for UK Financial Firms." sebastienrousseau.com. November 21, 2026. https://sebastienrousseau.com/2026-11-21-third-party-ai-governance-uk-firms/.

APA

Rousseau, S. (2026, November 21). Third-Party AI Is Now the Governance Problem for UK Financial Firms. sebastienrousseau.com. https://sebastienrousseau.com/2026-11-21-third-party-ai-governance-uk-firms/

Republish this article

Third-Party AI Is Now the Governance Problem for UK Financial Firms

Third parties supply 53% of high-materiality AI systems at surveyed UK financial firms, and overseeing them is the governance challenge firms cite most often.

This article is licensed under Creative Commons Attribution 4.0 International. Republication requires attribution to the canonical URL.

Third-Party AI Is Now the Governance Problem for UK Financial Firms

Third parties supply 53% of high-materiality AI systems at surveyed UK financial firms, and overseeing them is the governance challenge firms cite most often.

Originally published at https://sebastienrousseau.com/2026-11-21-third-party-ai-governance-uk-firms/ by Sebastien Rousseau.
Licensed under CC-BY-4.0.