The UK government wants banks, insurers, and financial-market companies to move artificial intelligence from isolated pilots into core operations, although dependence on overseas models and cloud infrastructure places a limit on the control those organisations can retain.

HM Treasury has published a Financial Services AI Adoption Plan covering regulation, operational resilience, skills, agentic payments, and sovereign capability. It proposes an industry support hub and further coordination among government, regulators, and companies.

The plan arrives as British financial institutions confront a practical example of model-access risk. UK banks were reported to have had limited access to Anthropic’s advanced Mythos model, prompting government AI adviser Harriet Rees to describe the episode as a warning over sovereignty and dependency.

The adoption plan says financial companies already use AI for fraud detection, operations, risk management, and customer service, but implementation remains uneven. Its sector adoption figure draws on earlier Bank of England and Financial Conduct Authority work that is now being updated.

Production systems cannot simply be paused

Moving AI into production changes the risk calculation because a pilot can be suspended when a model becomes unavailable or performs poorly, whereas a system embedded in fraud detection, credit, compliance, payments, or customer service requires continuity plans, monitoring, human escalation, and an alternative when the underlying service changes.

Model access is therefore more consequential than ordinary software availability. Frontier services may be restricted through capacity decisions, safety policies, commercial terms, export controls, national-security concerns, or government pressure. A regulated company can satisfy domestic rules and still lose access to an important capability because a decision was made elsewhere.

Sovereignty does not require Britain to build a domestic equivalent of every advanced model. It can also mean retaining data, preserving workload portability, understanding system behaviour, and maintaining critical functions when one supplier becomes unavailable. Architecture, procurement terms, and operational testing carry as much weight as the nationality of the vendor.

Financial institutions should also decide where broad general-purpose models offer a genuine advantage over smaller or more specialised systems. A general model may improve document analysis and internal knowledge work, while a defined risk process may need a narrower system that can be tested and reproduced more consistently.

Regulatory clarity cannot repair weak implementation

The plan responds to industry requests for greater regulatory certainty, but uncertainty is only one barrier. Many institutions still operate fragmented data estates, ageing core systems, complex outsourcing arrangements, and approval processes built for slower software releases. AI often exposes those weaknesses rather than bypassing them.

Agentic payments create an especially demanding test because software able to initiate or manage a transaction raises questions about identity, authority, consent, fraud, liability, revocation, and dispute resolution. A technically capable agent is not automatically an authorised payer, and every automated action requires an accountable chain of approval.

Skills policy must also extend beyond recruiting machine-learning specialists. Risk staff, compliance teams, product managers, procurement professionals, auditors, and operations workers need enough understanding to challenge the technology and recognise when an efficient process is producing an unacceptable outcome.

UK resilience supervision is moving in the same direction. Regulators have begun directly overseeing several major technology providers designated as critical third parties to financial services, acknowledging that infrastructure concentration can create sector-wide exposure. AI suppliers may acquire comparable importance even where they do not fall neatly within the same regime.

Support hubs and sandboxes can help companies test systems, but they cannot compensate for unclear ownership, poor data, weak fallback arrangements, or dependency that has never been exercised under failure conditions. The plan provides an implementation framework rather than evidence that scaled deployment has already been achieved.

British finance has the resources and regulatory experience to become an advanced AI user, although credibility will depend on whether institutions can deploy powerful systems without allowing a vendor’s access decision, model update, or policy change to interrupt regulated services.