Bank of England and FCA survey: 80% of respondents run AI agents, but few expect them to execute trades

Most UK financial firms now use AI agents, but very few let them act on their own, and fewer still expect them to trade, according to the fourth joint survey of AI and machine learning by the Bank of England and the Financial Conduct Authority, published on Monday 26 October.

The survey had 196 respondents, 101 larger firms and 95 smaller ones. For the first time, firms with fewer than 100 employees could complete a shorter questionnaire. The 105 firms that completed the full survey are the base for the figures on agents, risks, providers and governance below. The authors say 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”.

Adoption is high

• Across full and short-form responses, 86% of respondents said they use AI, up from 75% in 2024. The report warns that the two figures cover different groups of firms and should be compared with caution.
• Generative AI accounted for 56% of all reported AI systems.
• Operations and IT accounted for about 31% of reported AI systems. No other business area accounted for more than 11%.
• Current and planned use stayed comparatively limited in algorithmic execution, market making and hedging automation, and in liquidity and treasury decisions.

Agents: widespread but on a short lead

80% of respondents have AI agents deployed, and a further 14% plan to deploy them within 12 months. Most of those agents support decisions rather than take them:

• 80% have agents deployed for decision support, where agents recommend and humans decide and execute.
• 46% have agents with constrained execution, acting within predefined limits.
• 22% have agents with conditional autonomy.
• 2% have agents with broad autonomy.

In trade and execution, 7% of respondents reported decision-support agents, 5% constrained execution, 1% conditional autonomy and none broad autonomy. Even including planned deployment over the next 12 months, no more than 9% expected agents to execute trades or place orders, approve or decline customer outcomes, make or block payments, or change limits, thresholds and controls.

Where the risks sit

Cyber risk amplification was the top-rated systemic risk. 57% of respondents rated their concern as high or severe today, rising to 80% for the next one to three years. Critical third-party dependencies came next, rising from 40% to 63%.

Every listed systemic risk was expected to rise. The largest expected increase was in interconnectedness, such as automated trading and common failure modes, followed by AI-enabled market manipulation or surveillance evasion.

Concentration in providers

• Third-party implementations made up 39% of reported AI systems and 53% of high-materiality systems, up from 33% and 35% in 2024.
• The top three cloud and compute providers accounted for 79% of reported cloud and compute relationships. 98% of respondents used one of the three.
• Of 169 cloud and compute relationships, 37% were rated both highly critical and hard to substitute. Only 35% of that group had tested exit plans.

Governance gaps

94% of respondents reported firmwide AI policies. Pre-deployment controls such as human approval were widespread, but ongoing monitoring, leakage testing and periodic review were each reported by only around half. For high-materiality third-party systems, 81% relied on vendor assurance and contracts, and 24% reported deep knowledge of those systems, compared with 77% for systems built in-house.

The survey follows FCA chief executive Nikhil Rathi’s Mansion House speech on 22 October. The published text of that speech, which the FCA notes may differ from the version delivered, warned: “Unprecedented volatility, concentration risks, new forms of market abuse and rampant cross-border financial crime: fast-emerging technologies putting these on steroids.”

What we are watching

• Whether the Bank or the FCA follows the survey with guidance on AI agents in trading, or on third-party AI oversight.
• Whether supervisors ask trading firms to test exit plans for critical cloud and model providers.