Summary
- Reuters reports increased supervisory questioning around bank AI deployments.
- European focus includes recovery capability and dependence on external technology.
- Details attributed to unnamed sources should not be presented as formal new rules.
Banking supervisors in Britain and the eurozone are paying closer attention to how financial institutions use advanced artificial intelligence and whether they can recover from cyber incidents affecting automated systems, according to a Reuters investigation published on 13 October. The report describes efforts by the Bank of England and the European Central Bank to test resilience planning as the use of AI grows across financial operations.
Reuters cites people familiar with the supervisory discussions, rather than announcing a new regulation or a formal deadline applying across every bank. The distinction matters: supervisory questions and institution-specific expectations may be consequential even before a public rule changes, but the underlying account should remain attributed to those sources where officials have not independently confirmed the details.
The concern involves the increasingly complex dependencies created when a financial service relies on AI models, external providers, automated decisions and conventional software controls in the same operational process. An AI system used for fraud assessment, customer operations or cyber defence may need access to sensitive data and the ability to initiate actions. If its behaviour becomes unreliable, the bank must understand which systems it has touched and how to restore a trustworthy operating state.
Reuters reports that European authorities are pressing institutions to examine recovery capability, including backups and alternative ways to bring services back online. Those arrangements can be expensive and difficult to maintain across a large estate. Having a second copy of data is different from retaining the application dependencies, credentials, operational knowledge and tested processes needed to restart a payment or customer service safely.
The supervisory questions sit within an existing European regulatory framework. The Digital Operational Resilience Act places obligations on relevant financial entities for ICT risk management, incident management, testing and third-party risk. In the UK, operational resilience requirements likewise focus on the ability to deliver important business services through severe but plausible disruption. Neither regime assumes that preventing every cyber incident is possible.
AI introduces questions about the boundary of an incident. A maliciously manipulated model might generate an incorrect recommendation, disclose information through a connected tool or trigger transactions outside an intended workflow. Even without a malicious actor, a badly configured autonomous agent could cause consequences requiring investigation and restoration. Recovery then depends on having usable evidence of the chain of actions as well as backups of underlying systems.
The technical difficulties are likely to vary with the role AI plays in a banking service. A model used only to summarise internal documents creates different exposure from an agent authorised to change account settings or trigger another software system. In the second case, maintaining records of model instructions, tool calls and final transactions can be necessary to establish what happened after a failure. That record may be distributed across cloud systems, vendor logs and the bank’s own applications.
Recovery plans designed to bring back a server from backup do not automatically resolve questions about whether transactions initiated during the incident were valid. The supervisor’s view of materiality may also depend on whether a tool is used by a small experimental team or is incorporated into a service whose disruption could affect customers. Under DORA and UK resilience frameworks, firms already identify important services and test severe but plausible scenarios. The Reuters account describes a further area of scrutiny within that operational context, rather than an entirely new legal concept. It also leaves unresolved how many institutions have been formally required to change a recovery plan, which specific providers prompted concern and whether supervisors will publish more detailed expectations.
Concentration among external technology providers can complicate the recovery arrangements under examination. Several banks may depend on the same cloud infrastructure or model supplier, creating the possibility that one technical failure affects multiple institutions. The Reuters report does not establish that such a systemic event has occurred. It indicates that central banks and regulators are assessing those dependencies against recovery capabilities that were often developed before agentic AI became part of ordinary operations.
