Bank of England and FCA survey published October 26, 2026; news report October 27, 2026
UK Finance Runs Agents Everywhere and Trusts Them Almost Nowhere: 80 Percent Deployed, 2 Percent With Broad Autonomy
The Bank of England and the Financial Conduct Authority published their fourth survey of AI in UK financial services on October 26, drawing 196 responses, with most detailed figures from the 105 firms that completed the full questionnaire. AI use reached 86 percent of respondents, up from 75 percent in 2024, although the regulators caution that the respondent mix differs between editions and that the results describe participating firms rather than the whole sector. Generative AI now accounts for 56 percent of reported AI systems, and operations and IT host about 31 percent of them. AI agents are widespread but tightly held: 80 percent of respondents have agents in production for decision support, a further 14 percent plan to deploy within 12 months, and only 2 percent report agents running with broad autonomy. Even counting planned deployments, no more than 9 percent expect agents to approve or decline customer outcomes, place trades, move payments or change limits and controls. The third-party shift is the other headline: vendor-supplied systems are now 39 percent of all reported AI systems, up from 33 percent, and 53 percent of the systems firms rate as high materiality, up from 35 percent. The three largest cloud and compute providers hold 79 percent of reported relationships. Firms named information security and operational resilience as the main brake on third-party AI at 40 percent, limited transparency into vendor systems next at 34 percent, and oversight of vendor AI was the governance area firms most often found difficult.
Executives
Two numbers belong together in the board pack: 80 percent of firms run agents, and more than half of their most material AI systems come from outside vendors. Autonomy is being held back deliberately, while dependency is not. The governance question has moved from which model you run to how you oversee someone else's model, its updates and its concentration in a small set of cloud providers. That is vendor governance work, and the survey suggests supervisors will look there first.
CIO / CISO
Benchmark your own AI register against the survey shape: agents in decision support with human action, consequential actions fenced off, and a named control for vendor model changes. Where a high-materiality system is vendor-supplied, record what transparency you actually hold, evaluation evidence, update notice and exit options, because limited transparency and model risk management applied to supplied models were the difficulties firms reported most. Track cloud and model provider concentration as a single third-party risk view, in line with the Govern and Map functions of the NIST AI RMF.
Engineers
Build for the oversight firms say is hard: log vendor model version changes against your workflow outputs, keep a regression set you can run when a supplier updates a model, and instrument agent actions so a human approver sees the data scope and destination before anything consequential runs. The survey pattern, broad deployment with narrow autonomy, only stays safe if the technical fences, tool permissions and approval gates, match the policy claim.
Everyone (non-IT)
If your bank or insurer uses an AI assistant, this survey says it is most likely helping a staff member decide, not deciding on its own: almost no firm lets agents approve outcomes, move payments or place trades by themselves. That restraint is intentional, and it is reasonable to ask any financial firm you use what its agents are and are not allowed to do without a person.