Agents everywhere, autonomy rare: what UK finance reveals about AI in practice
The Bank of England and FCA's 2026 survey finds 86 percent of UK financial firms use AI and 80 percent have deployed agents, yet only 2 percent allow broad autonomy and oversight of third-party models remains thin.
TL;DR
- The Bank of England and the FCA published the fourth edition of their AI survey on 26 October: 86 percent of respondents use AI (75 percent in 2024), and 80 percent have already deployed AI agents, with a further 14 percent planning to within 12 months.
- Autonomy is the exception. Only 2 percent report agents with broad autonomy, and most agents support human decisions or act within predefined limits.
- Third-party systems now account for 53 percent of the AI that firms rate as highly material (35 percent in 2024), while ongoing monitoring and incident playbooks lag behind pre-deployment checks.
The survey drew wide coverage in financial and trade press this week. Finance is a useful bellwether: it is heavily regulated, data-rich, and already deep into agentic AI. What its firms do about control is a preview for other sectors.
Deployed, but on a short leash
| Survey finding (2026) | Figure |
|---|---|
| Firms using AI | 86% (75% in 2024) |
| Firms with AI agents deployed | 80% (plus 14% planning within 12 months) |
| Agents with broad autonomy | 2% |
| Agents in constrained execution | 46% |
| Share of highly material AI that is third-party | 53% (35% in 2024) |
| Firms overseeing third parties mainly via vendor assurance and contracts | 81% |
Figures come from the regulators' report as summarised in secondary coverage. The sample is small and self-reported: 196 responses, of which about 105 completed the detailed questionnaire, so some percentages apply to that smaller base.
The control gap is after launch
Over 90 percent of firms use pre-deployment review and human approval. But ongoing monitoring sits at roughly half, and fewer than half have incident playbooks. Controls are strongest at the gate and weakest once an agent is running, which is exactly where behaviour drifts. Add growing dependence on outside models and platforms, overseen mostly through contracts, and the risk concentrates in places firms cannot directly inspect.
The FCA has also said existing rules already apply and that it is not writing new ones. For firms, that means accountability sits with them, not with the vendor.
What this means for leaders
- Define an autonomy ladder for agents (suggest, act within limits, act freely) and record where each one sits.
- Fund post-deployment monitoring and incident response with the same rigour as pre-launch approval.
- Map which material processes rely on third-party models, and test exit and fallback options, not just contract terms.
- Name an accountable owner for each agent, as regulators will expect one.
- Treat "agent count" as a vanity metric; track outcomes and incidents instead.
Related reading
- The BoE-FCA AI survey: agents widespread, autonomy limited (Resultsense)
- The ownership gap: who actually runs AI decisions
Source: Artificial intelligence in UK financial services - 2026, Bank of England and Financial Conduct Authority.