The Bank of England's four-part warning for agentic finance

The Bank of England's July 2026 Financial Stability Report frames AI risk through four connected channels: use in core financial decisions, use in markets, reliance on AI service providers, and a changing cyber threat environment. The important shift is that AI is no longer treated solely as a model-risk question. It is also an operational-resilience and concentration-risk question.

As agents gain the ability to use tools and complete longer chains of work, a control failure can travel further before a person intervenes. That makes the permissions surrounding an agent as important as the quality of its output.

A research assistant that drafts a memo has a limited blast radius. An agent that can access payment systems, modify forecasts, or trigger downstream workflows has a much larger one. Treating both as the same class of AI use leaves a governance gap.

The report also highlights shared infrastructure. If many institutions depend on the same model or service provider, an outage, vulnerability, or unexpected behaviour can affect multiple firms at once rather than remaining an isolated incident.

Create an agent register that records each system's data access, available tools, permitted actions, human approval points, and shutdown mechanism. Prioritise review based on what the agent can change—not how impressive its benchmark scores appear.

For critical workflows, test provider failure and degraded-output scenarios just as you would test a core technology dependency. The practical question is not simply whether the agent works, but whether the finance process remains safe when it does not.