Editor's note — September 12, 2026: Removed the published TOPIC PITCH drafting label, added the named speech and distinguished the author's inference from the source.
The Governance Gap
Issue 19: Agentic AI Is Breaking the Old Governance Model
Sarah Breeden's June 30, 2026 Bank of England speech discusses how AI is changing finance. My inference is that enterprises should reassess delegated authority when AI systems can use tools and act across workflows.
The board-level issue is no longer just “What did the AI say?” It is “What did the AI do, who authorized it, and can we prove it?”
THE TAKE
Agentic AI is not just another AI feature.
It is a change in operating model.
A chatbot gives an answer.
An AI agent can do work.
That distinction matters. Once AI can access systems, use tools, make recommendations, execute tasks, or interact with customers, it becomes more than software. It becomes a digital actor inside the business.
Existing governance programs should be tested against that authority.
Agents may use familiar application and service-account infrastructure while introducing more dynamic behavior. The controls need to address both.
That creates a control gap.
Who owns the agent?
What data can it access?
What decisions can it influence?
What actions can it take without approval?
What happens when it is wrong?
Where are the logs?
Who can shut it down?
Those questions need answers before agentic AI scales across the enterprise.
THREE THINGS BOARDS SHOULD KNOW
Agentic AI creates a new accountability problem — If an AI system takes action, the organization needs clear ownership, authority limits, monitoring, and evidence.
Human oversight cannot be vague — “Human in the loop” only works when the company defines when review is required, who performs it, and what proof is retained.
AI agents need enterprise-grade controls — Identity, access management, logging, exception handling, kill switches, and periodic review should apply to agents just like other critical systems.
THE QUESTION I'M HEARING
“How do we let the business use AI agents without slowing everything down?”
You do it by defining bounded autonomy.
Not every AI action needs the same level of approval.
A low-risk agent summarizing internal knowledge does not need the same oversight as an agent touching customer records, financial workflows, legal documents, security systems, or regulated decisions.
The key is to classify use cases by risk.
Boards should expect management to define:
- Approved agent use cases
- Business owners
- Data access boundaries
- Systems the agent can touch
- Actions that require human approval
- Actions the agent is never allowed to take
- Logging and monitoring requirements
- Escalation and shutdown procedures
- Review frequency
This defines who can act and when intervention is required.
The degree of autonomy should increase only when evidence supports it.
The danger is letting AI agents spread like shadow IT with better branding. Once that happens, the organization loses visibility into where agents are operating, what authority they have, and what risk they are creating.
THE BOTTOM LINE
Agentic AI is exposing a major governance gap.
If an AI system can act on behalf of the business, it must be governed like part of the business.
Boards should ask for the AI agent inventory before the incident forces the conversation.
Source and further reading: Primary reference. Framework guidance is not a statement that a particular law applies to every organization.