Executive briefing: AI risk
Agentic AI Risks and AI Agent Governance: The Controls Boards Must Require
Agentic AI is AI that acts, not only advises: it sends email, changes records, places orders and moves money without a person approving each step. The main agentic AI risks are autonomy drift, excess access, chained errors, data leakage, an accountability gap and agents switched on inside vendor software. AI agent governance means ten controls in place before any agent acts, from an agent inventory and a named owner to spend limits, an audit trail and a kill switch, and an autonomy ceiling the board sets for each agent.
Read the full briefingIn brief
- Generative AI drafts and a person decides. Agentic AI plans, decides and acts, often with no person between the decision and the action.
- Agents bring six risks: autonomy drift, excess access, chained errors, data leakage, an accountability gap and third-party agents.
- Ten controls should be in place before any agent acts. The first two, an inventory and a named owner, come before everything else.
- Autonomy is not all or nothing. The board sets a ceiling for each agent, from level 1 (assist) to level 5 (act autonomously), by impact.
- Most organisations meet agents first through their suppliers. New vendor agent features should stay off until an owner approves them.
Why the board should care
The organisation answers for its AI
In Moffatt v Air Canada (2024) a tribunal held the airline liable for its chatbot's wrong advice. "The AI did it" is not a defence. An agent that could have paid the refund would have turned one error into a payout.
Agents act at machine speed
An agent with broad access and no ceiling can repeat the same mistake for every customer, order or record it touches before anyone notices.
Agents arrive without a project
Vendors add agent features to tools you already use, sometimes switched on by default, with access to everything the tool can reach.
Full briefing
Read the full briefing
Free for executives. Confirm your work email once and every AIPEX executive briefing opens for 12 months.
- The video briefing
- The full written briefing
Next step
Set the limits with your leadership team
The Executive AI Strategy Day (£1,499 + VAT) builds the agent inventory, autonomy ceilings and control set with your leadership team, in one day. Prefer to talk first? A free 30-minute scoping call with a senior partner.
FAQ
Frequently asked questions
What is agentic AI?
Agentic AI is AI that can plan a task, decide the steps and carry them out across other systems, such as sending email, changing records, placing orders or moving money, without a person approving each step.
What are the main agentic AI risks?
Six risks: autonomy drift, excess access, chained errors, data leakage (including manipulation through prompt injection), an accountability gap, and agent features switched on inside vendor software.
What is AI agent governance?
AI agent governance is the set of controls, owners and limits that decide what an AI agent may do, who answers for it, and how the organisation sees, stops and recovers from its actions.
What controls should be in place before an AI agent goes live?
Ten: an agent inventory, a named owner, least-privilege access, action and spend limits, human approval points, testing before release, a full audit trail, a kill switch, live monitoring and incident response.
How much autonomy should an AI agent have?
Set a ceiling for each agent by impact, from level 1 (assist) to level 5 (act autonomously). Anything touching customers, money or regulated decisions should normally sit at level 3 (act with approval) or below until the controls are proven.
Who is accountable when an AI agent makes a mistake?
The organisation. In Moffatt v Air Canada (2024) a tribunal held the airline liable for its chatbot's wrong advice. That is why every agent needs one named, accountable executive.
Does the EU AI Act cover AI agents?
The Act regulates AI systems by risk and use, not by whether they are agents. An agent used for a high-risk purpose carries the high-risk duties, including human oversight and logging, which apply from 2 December 2027. The ten controls help build that evidence.
This briefing is general information on governing AI agents. It is governance guidance, not legal advice. AIPEX is not a certification body. For decisions about specific AI systems, take advice from qualified advisers.