Eighty-six per cent of surveyed firms now use AI, but the more telling numbers sit underneath that headline: most deployed agents advise rather than act, outside suppliers provide a growing share of the systems that matter most, and oversight of those suppliers is the governance challenge firms report most often.
Executive Summary
- The survey received 196 responses, and 86% of respondents reported using AI, up from 75% in 2024.
- Generative AI accounted for 56% of reported AI systems; operations and IT accounted for approximately 31%.
- 80% reported decision-support agents in production; only 2% reported agents with broad autonomy.
- Third-party implementations made up 53% of high-materiality systems, up from 35% in 2024.
- 94% reported firmwide AI policies, principles or standards, but ongoing monitoring is reported by only around half of respondents.
- Respondents expect both benefits and risks to grow, with cyber risk amplification and critical third-party dependencies the two highest-rated systemic risks.
What the survey is, and what it isn't #
The publication is a descriptive summary of aggregate results from the latest AI survey run as part of the Bank and FCA's wider programme of work on AI. That programme includes previous AI/ML surveys in 2019, 2022 and 2024. The 2026 survey kept core questions from earlier rounds for continuity and added or amended others to reflect developments in AI adoption and capability since 2024.
A total of 196 firms responded, split into 101 larger firms and 95 smaller firms. Of those, 105 completed the full survey. Unless otherwise stated, the figures are based on full-survey responses, to aid comparability with previous surveys, and percentages are calculated from valid question-level responses only, which means firms that skipped a question are excluded from that question's denominator. Responses from international banks and broker/dealers refer to their UK operations.
Two caveats shape how far any of these numbers can be pushed. 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. And comparisons with 2024 should be interpreted with care where question wording, definitions, scope or respondent populations differ. I'll flag the year-on-year movements below because they're the most striking figures in the report, but they carry that qualification every time.
For the purposes of the survey, an AI system is a machine-based system that, for explicit or implicit objectives, infers from the input it receives how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. That's a broad definition, and it's worth keeping in mind when a figure like 86% appears. It describes use of a wide category of systems, not necessarily use of any single technology.
Adoption has become the default #
The headline is simple. 86% of respondents reported using AI, up from 75% in 2024. The report's own summary is that firms are increasingly reporting use of generative AI, AI-enabled agents and third-party implementations of AI systems. Those three themes organise most of what follows.
Among the full-survey respondents, deployment is close to total: 94% reported at least one AI system in production, rising to 99% once pilots and deployments planned over the next 12 months were included. The most widely reported current use was enterprise productivity, such as coding assistants or summarisation tools, selected by 86% of respondents.
Smaller firms are not far behind. Among respondents to the new short-form survey for firms with fewer than 100 employees, 73% reported using AI, and of those, 58% said they predominantly use general-purpose large language models. For a smaller firm, that is a notable pattern: the technology most associated with the current wave of adoption is also the one a majority of small adopters lean on most. Operations was the most commonly reported domain for their current or planned AI use, at 64%.
Where the systems sit tells a similar story. In terms of reported systems, Operations and IT accounted for approximately 31%, up from around 22% in 2024. No other individual business area accounted for more than 11%. That's a wide gap. Whatever the public conversation about AI in customer-facing or trading roles, the survey's weight sits in the back office.
The perceived benefits line up with that placement. Operational efficiency, automation and productivity were the highest-rated areas of perceived benefit from AI today. It would be odd if firms had concentrated their systems in operations and IT and then reported their biggest gains elsewhere. They didn't.
Generative AI is now the majority of the estate #
The shift in technology mix is large. Generative AI accounted for 56% of all reported AI systems, compared with 17% for foundation models in 2024. I'd be careful about reading that as a clean like-for-like jump: the 2024 figure counted use cases rather than systems, and the report says the two measures are not directly comparable, only that together they suggest a substantial increase in firms' use of general-purpose capabilities. But even with that caution, a category that makes up more than half of all reported systems is no longer a niche experiment.
It also isn't evenly spread across respondents. The top ten respondents accounted for 43% of all reported AI systems. A small group of firms is running a very large share of the systems counted, so the aggregate picture is partly a picture of those ten. A reader looking for what a typical firm does should keep that concentration in mind.
Another change moves in the opposite direction from what you might expect. Around 7% of reported AI systems were assessed by respondents as high materiality, compared to 16% in 2024. That is a falling share, not a falling count: the report states that the absolute number of high-materiality systems reported in 2026 was greater than in 2024. The survey doesn't say why the share fell. A growing estate dominated by operational and IT uses could plausibly be diluting the share that matters most, but that's my speculation rather than something the data establishes. The high-materiality systems are concentrated too: ten firms accounted for 62% of them.
Agents: plenty of advice, very little autonomy #
The agent figures are the clearest example of firms adopting something new while keeping a hand on it. 80% reported having AI agents in production that provided decision support, meaning agents recommend actions and humans decide and execute. Between those poles, 46% reported agents with constrained execution (acting within predefined limits and rules) and 22% reported agents with conditional autonomy (acting independently but needing human approval for certain actions). Only 2% reported deploying agents with broad autonomy.
I think the 80% and the 2% are the most important pair of numbers in the report, because they show where the line currently sits. Agents are common; autonomous agents are rare. The dominant practice among respondents is to put agents into workflows as advisers, with execution confined to predefined limits where it is allowed at all.
Customer communications offer a more specific view of where that line may move. 16% reported current use of agents for automated customer communications, rising to 34% when planned deployment was included. Planned deployment more than doubles the current figure. Plans aren't outcomes, so the 34% is a statement of intent, not adoption. But it points to customer communication as one area where firms expect to give agents more to do. The harder decisions stay off the plan: even including planned deployment over the next 12 months, no more than 9% of respondents expected agents to approve or decline customer outcomes, execute trades, initiate or block payments, or change limits and controls.
This matters for the control discussion further down. The decision-support model depends on a human deciding, and automated customer communications is exactly the kind of use where a human may be less directly in the loop. Whether the controls firms report today are built for that shift is a question the survey's numbers raise but can't answer.
The third-party question #
If one trend deserves a closer look, it's this one. Third-party implementations accounted for 39% of all reported AI systems, up from 33% in 2024. That's a modest rise. The figure for high-materiality systems is not modest: third-party implementations accounted for 53% of high-materiality AI systems, up from 35% in 2024.
Put plainly, outside suppliers provide fewer than four in ten systems overall but more than half of the ones firms consider most material. The dependency is concentrated precisely where the stakes are highest, and it grew by eighteen percentage points in two years. Given the caveat about comparing rounds, I wouldn't lean on the exact size of that jump, but the direction and the 53% level are reported plainly.
Respondents see the constraint themselves. Information-security and operational-resilience concerns were identified as a major constraint on third-party AI adoption by 40% of respondents, followed by limited transparency into third-party AI systems at 34%. So the sector leaning harder on suppliers for its most material systems also names supplier security and opacity as its main brakes on third-party adoption.
The risk ratings point the same way. 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. Looking ahead, respondents expected cyber risk amplification and critical third-party dependencies to intensify as AI use becomes more widespread, and these were the two highest-rated systemic risks both today and over the next three years. For cyber risk amplification, 57% rated their concern high or severe today, rising to 80% in the one to three years ahead; for critical third-party dependencies, the share rose from 40% to 63%. That is an expectation, and the survey frames it as one, but it's consistent in kind with the dependency figures above.
For anyone selling AI into financial services, the implication I'd draw is commercial rather than regulatory: buyers say transparency into third-party systems limits them, and a supplier that can answer that concern directly addresses something 34% of respondents named. That's my reading, not a survey finding.
Governance: strong on paper, thinner in practice #
On structure, respondents look well organised. 94% reported firmwide AI policies, principles or standards, and 84% reported centralised AI/model-risk governance fora. Among the controls used to manage data risk in AI systems, human approval, pre-deployment review and vendor due diligence were each reported by more than 90% of respondents.
The picture changes once you move from controls applied before a system goes live to controls applied afterwards. Ongoing monitoring, leakage testing and periodic review or recertification were each reported by around half of respondents. A gap of roughly forty percentage points separates the front-loaded safeguards from the continuing ones.
I'd argue this is the weakest point in the reported control environment. Checks before deployment are easy to describe and easy to audit. Monitoring a system in production is a continuing cost, and the figures suggest around half of firms are reporting it. Where third-party systems dominate the high-materiality tier, the gap matters more, because firms can't always inspect what they didn't build. The survey shows this directly: for high-materiality systems, 77% of respondents reported relying on deep knowledge of internally developed systems, against 24% for third-party systems, where 81% relied instead on vendor assurance and contractual arrangements.
Firms recognise the difficulty. Oversight of third-party AI 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 as a challenge by 54% of respondents, and overseeing systems provided or materially influenced by third parties by 49%. Due diligence at the point of purchase is close to universal; the follow-through is where firms say they are least sure of themselves.
Constraints that aren't about regulation #
It would be easy to assume regulation is the main brake on adoption. The survey offers a mixed picture. 42% of respondents perceived UK data protection and privacy as posing a high compliance burden, and 29% said the same of international regulation. Those are real shares, but they're minorities, and 28% of respondents indicated no regulatory constraints at all.
Practical constraints show up at lower rates. The most cited non-regulatory constraints, difficulties integrating AI into existing business processes and decisions, and constraints imposed by legacy systems or infrastructure, were identified as large constraints by 23% and 22% of respondents respectively. Among small firms, 30% identified data protection, confidentiality and IT considerations as constraints.
Read together, the numbers suggest that the most commonly cited obstacles are data protection and supplier security, with integration and legacy infrastructure further behind. I wouldn't over-interpret the ordering, since the questions weren't posed identically (some ask about compliance burden, others about large constraints), but the absence of any single overwhelming barrier fits with 86% adoption.
Expectations: more benefit, more risk #
Respondents 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 treats them as such; they suggest where firms hope AI will go, not where it has gone.
Risk expectations move in step. Over the next one to three years, respondents expected risk levels to increase across all categories. The source states that AI developments may bring benefits including greater efficiency, improved decision support and stronger risk management, and may also create or amplify risks related to governance, operational resilience, data protection, consumer outcomes, third-party dependency and financial stability. Both halves of that sentence are hedged, and rightly so.
There's something worth noticing in the pairing. Cybersecurity and operational resilience appear among the largest anticipated benefit increases, and cyber risk amplification appears as the highest-rated systemic risk. Firms expect AI to help with cyber resilience and to make cyber risk worse at the same time. The survey doesn't reconcile those expectations, and I wouldn't pretend to. They sit side by side.
Frontier AI and cyber #
The survey asked specifically about frontier AI cyber capabilities. 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, not a measured outcome, and should be read as respondents' expectation.
It has already changed some behaviour, though. 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. So the concern has produced a concrete shift in management attention towards patching, according to nearly two-thirds of those asked.
A smaller, separate finding sits alongside: 57% of respondents reported considering the sustainability impacts of AI use. It's a majority, but a bare one, and few say so publicly: 82% do not report the sustainability impacts of AI externally, though 36% of those plan to within two years.
What to watch #
The numbers describe an industry that has adopted AI broadly, kept humans in charge of decisions, and handed a growing share of its most material systems to outside suppliers. The tension is in that last point. The controls that firms report most often, such as policies, approvals and due diligence, happen early. The ones that would watch a supplier's system over time are reported by around half.
Two figures will be worth checking in the next round, assuming comparable wording. The first is the customer-communications number: 16% today, 34% with planned deployment included. If the plans turn into production use, the 2% share of broad autonomy and the 80% share of decision support will tell us whether the human-decides model survived contact with volume. The second is the 53%. If third-party implementations continue to take a larger share of high-materiality systems while ongoing monitoring stays near half, the survey's own respondents, who already rate critical third-party dependencies as a leading systemic risk, will have described the problem before it arrived. The open question is whether monitoring practice closes the gap before the dependency widens further.
Frequently Asked Questions #
How many firms use AI according to the Bank of England and FCA 2026 survey? 86% of respondents reported currently using AI, up from 75% in 2024 and 58% in 2022. The survey received 196 responses, 101 from larger firms and 95 from smaller firms. The report warns that differences in respondent composition and question scope mean the 2026 and 2024 percentages do not relate to the same population and should be compared with caution.
How much autonomy do UK financial firms give AI agents? Very little. 80% of full-survey respondents reported agents deployed for decision support, where agents recommend actions and humans decide and execute. 46% reported agents with constrained execution, 22% with conditional autonomy and only 2% with broad autonomy.
How much of their AI do firms get from third parties? Third-party implementations accounted for 39% of all reported AI systems, up from 33% in 2024 and 17% in 2022. They also accounted for 53% of high-materiality AI systems, up from 35% in 2024.
What do firms see as the biggest AI risks? At firm level, the highest-rated current risks were data protection and security, third-party and supply chain risk, and GenAI-specific risks such as hallucinations and prompt injection. At system level, cyber risk amplification and critical third-party dependencies were the two highest-rated risks, both today and over the next three years.
Where are AI governance controls weakest? Human approval, pre-deployment review and vendor due diligence were each reported by more than 90% of respondents, while ongoing monitoring, leakage testing and periodic review or recertification were each reported by around half. Oversight of third-party AI was the most commonly reported governance challenge.
References #
- Bank of England and Financial Conduct Authority, (2026). Artificial intelligence in UK financial services - 2026 ⧉. Bank of England.
- Bank of England and Financial Conduct Authority, (2024). Artificial intelligence in UK financial services - 2024 ⧉. Bank of England.
- Bank of England and Financial Conduct Authority, (2022). Machine learning in UK financial services ⧉. Bank of England.
- Bank of England and Financial Conduct Authority, (2019). Machine learning in UK financial services ⧉. Bank of England.
- Bank of England, Prudential Regulation Authority and Financial Conduct Authority, (2022). DP5/22: Artificial Intelligence and Machine Learning ⧉. Bank of England.
- Bank of England, Prudential Regulation Authority and Financial Conduct Authority, (2023). FS2/23: Artificial Intelligence and Machine Learning ⧉. Bank of England.
Last reviewed .
