Eighty percent of UK financial firms now run AI agents in production, according to a Bank of England and Financial Conduct Authority survey published on 26 October, yet almost all deployments keep humans in charge of payments, trades and customer outcomes.

The fourth joint AI survey shows adoption continuing to climb: 86 percent of respondents use AI in some form, up from 75 percent in 2024. Generative AI accounts for 56 percent of reported systems, a sharp rise from the 17 percent share attributed to foundation models two years earlier.

Where agents actually work

Agents in this regulatory taxonomy are software components that can plan multi-step tasks. Most banks deploy them in operations and IT, which together represent about 31 percent of reported systems. Customer-facing automation is growing—16 percent of firms use agents for automated communications today, rising to 34 percent when planned deployments over the next year are included—but autonomous payment approval remains rare.

Only 2 percent of respondents said agents operate with broad autonomy. The dominant model is decision support: machines recommend, humans execute. That aligns with FCA expectations that accountability stays with authorised persons even when models draft emails or prepayment files.

Materiality and concentration

Regulators care about materiality because high-impact failures can move markets or harm consumers. Seven percent of systems were rated high materiality in 2026, down as a share but up in absolute count as firms deploy more models overall. Cloud concentration remains pronounced: the top three hosting providers account for 79 percent of relationships, while the top three model API vendors account for 54 percent.

Information-security worries top the list of constraints on third-party AI, cited as a major barrier by 40 percent of firms. Transparency into vendor training data and evaluation logs trails close behind, echoing speeches from BoE officials about systemic risk in shared model stacks.

UK policy backdrop

The survey lands days before Chancellor John Healey’s budget, which is expected to include further AI security spending through the AI Security Institute and a National Cyber Shield programme. City ministers want the UK to host model evaluation while keeping banks from experimenting recklessly on retail customers.

Unlike the EU AI Act’s product rules, UK supervisors still rely on principles-based regulation supplemented by thematic reviews. The BoE-FCA data gives supervisors a benchmark to challenge boards who claim they are “just experimenting” while agents touch live ledgers.

Implications for customers

Retail users may notice faster chat responses or slicker fraud alerts long before agents approve mortgages alone. Banks market AI assistants cautiously after past mis-selling scandals. The survey suggests even aggressive adopters keep payment initiation behind dual controls.

SME borrowers could see quicker credit decisions where agents gather documents, but underwriters still sign off limits. Insurers experiment with claims triage, yet final payouts usually require human signatory thresholds.

What supervisors want next

Both agencies plan deeper dives into agent logging, kill switches and third-party contract clauses. Firms should expect questions about whether vendor models can be disconnected within minutes if benchmarks go rogue, a lesson from cyber exercises rather than hypothetical ethics debates.

For investors, widespread agent adoption without autonomy may disappoint futurists but reassures risk officers. The UK’s largest lenders can show regulators they are modernising back offices while keeping the blunt tools of banking—moving money, changing limits—under human thumbs.