Financial institutions are putting AI into production faster than many are maturing ongoing oversight. A UK survey found that 86% of respondents use AI, while 80% reported having AI agents deployed and another 14% planned deployment within 12 months. Only 2% reported agents with broad autonomy. Third-party systems represented 53% of reported high-materiality AI use.
What Happened
AI use has continued to increase across UK financial-services firms responding to a 2026 survey from the Bank of England and Financial Conduct Authority. Across the full and short-form surveys, 86% of respondents said they currently use AI, up from 75% in 2024. The report cautions that differences in respondent composition and question scope mean the percentages are not directly comparable. The survey received 196 responses, including 101 larger firms and 95 smaller firms. A total of 105 respondents completed the full survey used for many of the detailed findings.
Generative AI accounted for 56% of all reported AI systems. Operations and IT represented approximately 31% of reported systems, more than any other individual business area. AI agents were already widespread: 80% of respondents reported having agents deployed, while another 14% planned to deploy them within the next 12 months. Most deployed agents provided decision support, with people deciding and executing the action. Only 2% of respondents reported agents with broad autonomy.
Third-party AI systems became both more common and more material. They represented 39% of all reported AI systems and 53% of high-materiality systems. Provider concentration was substantial. The top three cloud and compute providers accounted for 79% of reported relationships, while the top three model and API providers accounted for 54%.
Formal governance was common. Ninety-four percent of respondents reported firmwide AI policies, principles, or standards, and 84% reported centralized AI or model-risk governance forums. However, the control picture was less mature after deployment. Human approval, pre-deployment review, 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.
Why It Matters
The findings show a shift from experimental AI to operational dependency. AI is increasingly embedded in enterprise productivity, operations, IT, risk, and compliance. Even when an agent has limited autonomy, it can influence decisions, process company information, or use tools connected to business systems. The risk is therefore shaped by data access and integration, not only by whether the agent can act without approval.
Third-party concentration raises a separate resilience concern. Firms may rely on a small number of cloud, compute, model, and API providers while having less direct knowledge of those systems than they have of internally developed technology. Seventy-seven percent of respondents reported relying on deep knowledge of internally developed high-materiality systems, compared with 24% for third-party systems. For third-party systems, 81% instead relied on vendor assurance and contractual arrangements.
That imbalance matters because third-party AI represented more than half of reported high-materiality systems. A strong contract or assurance package can support governance, but it does not provide the same operational visibility as direct knowledge. Firms need to know how they will detect degradation, data leakage, manipulation, or service disruption after deployment.
What to Watch
The most important development is whether post-deployment controls catch up with adoption. Formal policies and approval processes are already common, but ongoing monitoring, leakage testing, and recertification are less consistently reported. Firms may need to move from one-time approval toward lifecycle oversight as their AI estates grow.
AI-agent autonomy also deserves attention. Broad autonomy remains rare at 2%, but 80% of respondents already have agents providing decision support, and another 14% plan agent deployment. If organizations expand what those agents can access or execute, existing approval and monitoring models may need to change.
Frontier AI cyber capabilities are another pressure point. Eighty-eight percent of respondents assessed their likely impact over the next two years as moderate or high. Sixty-four percent said awareness of frontier AI had increased management attention most strongly around vulnerability management and patching cadence. Watch whether that attention produces measurable improvements in exposure reduction and recovery readiness.
What Defenders Should Do
Inventory first: maintain a current record of AI systems, agents, models, business owners, data sources, tools, external providers, and production status. Separate experiments from systems that influence operations, customer communications, compliance work, or other material decisions.
Map access: document what each AI system can read, generate, recommend, and execute. Include credentials, plugins, APIs, files, and connected business applications. Limited autonomy does not mean limited exposure if an agent can reach sensitive information or shape a consequential decision.
Strengthen the operating cycle: pair pre-deployment review with ongoing monitoring, leakage testing, periodic review, and recertification. Define conditions that trigger suspension, rollback, human escalation, or renewed approval. Assign ownership for monitoring results rather than assuming the vendor will identify every problem.
Assess concentration: identify business processes that depend on the same cloud, compute, model, or API provider. Review what happens if that provider becomes unavailable, changes behavior, or cannot supply enough evidence for an investigation. Vendor assurance and contracts should be supported by technical monitoring and tested contingency plans.
Connect AI and cyber governance: include AI systems in vulnerability management, incident response, access review, data-protection controls, and operational-resilience exercises. AI should not remain a separate governance topic once it is part of production operations.
The Larger Lesson
AI governance is becoming an operational discipline. Policies, committees, and pre-launch reviews establish accountability, but they cannot see what changes after a system is deployed. Continuous monitoring and tested response processes are what turn governance into resilience.
The report also shows that limited autonomy is not the same as limited risk. Decision-support agents, third-party models, and concentrated infrastructure can materially affect an organization even when a person still approves the final action. Effective oversight must follow data, access, dependency, and business consequence.