An AI agent should have only as much autonomy as the organisation can monitor, reverse and answer for. Boards should set an autonomy ceiling for each agent, tied to the value and reversibility of what it can do, and raise that ceiling only on evidence. Capability alone is never a reason to grant more freedom.
In a recent interview, Mustafa Suleyman discussed safety and the economics of AI agents. This article sets out the AIPEX view on the practical question it raises for any board: how far should we let AI agents act on our behalf?
Why is autonomy a board decision, not a technical setting?
Autonomy decides who acts in the organisation’s name. When an agent sends a customer email, changes a price, approves a refund or places an order, the organisation has acted, and the organisation answers for it. Courts and regulators do not accept “the system did it” as a defence. That makes the level of autonomy a question of delegated authority, which has always been a board matter.
It is also a question the technology will keep pushing. Each new generation of models can do more, and vendors increasingly switch agent features on inside software you already use. If the board does not set limits, the limits are set by default, by whoever configures the tool.
The commercial record supports caution. Gartner has predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls (Gartner, 25 June 2025). Weak controls are one of the named causes, and they are the one the board can fix.
What does an autonomy ceiling look like?
An autonomy ceiling is the highest level at which a given agent may operate. We use a five-level scale, set out in full in our briefing on agentic AI controls. The board, or a committee it delegates to, approves the ceiling for each agent, and management may not exceed it without coming back.
| Level | What the agent does | Typical fit |
|---|---|---|
| 1. Assist | Suggests. A person does everything. | Drafting, research, summaries |
| 2. Recommend | Proposes an action. A person decides. | Credit, pricing or hiring shortlists |
| 3. Act with approval | Prepares the action. A person approves each one. | Customer replies, supplier orders |
| 4. Act within limits | Acts alone inside hard limits, and reports. | Low-value refunds, routine scheduling |
| 5. Act autonomously | Acts alone, reviewed after the event. | Rarely justified today |
Most organisations should expect the majority of their agents to sit at levels 1 to 3 for some time. Level 4 is where the real productivity gains begin, and it is also where the controls must be strongest.
How should the board decide the right level for each agent?
Two tests do most of the work. The first is value at risk: what is the largest amount of money, data or reputation a single wrong action could put in play? The second is reversibility: if the agent gets it wrong, can the action be undone quickly and completely? A misfiled internal note is reversible. A payment, a published statement or a deleted record may not be.
Low value and reversible actions can go higher on the scale. High value or irreversible actions stay low, whatever the vendor says the model can do. Add a third test for anything customer-facing or regulated: would we be comfortable explaining this decision to the person affected, and to a regulator, without a human having looked at it?
Recognised frameworks point the same way. The NIST AI Risk Management Framework (AI RMF 1.0, January 2023) places accountability and human oversight in its Govern function, and ISO/IEC 42001 expects top management to assign roles and authorities for AI. Neither tells you what level to choose. Both expect you to have chosen deliberately, and to be able to show it.
What must be in place before an agent moves up a level?
Evidence, not enthusiasm. Before an agent moves from level 3 to level 4, management should be able to show the board:
- A named executive who owns the agent and answers for its actions.
- A track record at the current level: volume handled, error rate and how errors were caught.
- Hard limits on value, volume and reach, enforced by the system rather than by policy.
- Access restricted to the systems and data the task needs, and nothing more.
- A complete audit trail that attributes every action to the agent.
- A kill switch that a named person can use at once, and a tested plan for reversing its actions.
If any of these is missing, the answer is “not yet”. Ownership in particular is where agent programmes fail first; our article on who owns the AI agent when it breaks sets out how to write the mandate.
What should the board ask management now?
- How many agents act on our behalf today, including agent features inside vendor software?
- What autonomy level is each one operating at, and who approved it?
- Which agents can move money, change customer records or publish anything without a person approving it?
- What evidence would we need to see before raising any agent’s ceiling?
- When did we last test a kill switch, and how long did it take?
These questions are not a brake on ambition. Organisations that can show control over their agents are the ones that can safely give them more to do. Those that cannot will find their agent programmes stalled by the first serious error, or cancelled before they reach it.
Where to start
Start with an inventory and a ceiling for every agent you already have, before approving new ones. Our briefing on agentic AI risks and the controls boards must require sets out the ten controls and the five autonomy levels in full, with a board checklist. For most organisations the next move is Step 3 on the AIPEX roadmap, Plan: the Executive AI Strategy Day (£1,499 + VAT) produces an agent inventory, autonomy ceilings, the control set and a 90-day plan in one working day.
Frequently asked questions
How much autonomy should an AI agent be given?
Only as much as the organisation can evidence control over. Start agents at low autonomy, such as drafting or recommending with human approval, and raise the level only when monitoring, audit trails and a tested way to stop the agent show it is performing within agreed limits.
Who should own an AI agent's decisions?
A named executive, in the same way a manager owns decisions delegated to their team. The agent acts under delegated authority, so its limits, approvals and escalation routes should be written down and reviewed. For more, see Who Owns an AI Agent's Decisions? →
What controls should be in place before an agent acts on its own?
Clear limits on what it can do and spend, logging of every action, monitoring that flags unusual behaviour, human approval for high-impact steps, and a tested kill switch. Boards should also expect regular review of whether the autonomy level is still justified.
Correct as of 28 September 2026. This content reflects the law, official guidance and deadlines as published on that date. The EU AI Act and its guidance, standards and deadlines are still being amended and clarified. If you are reading this after 28 September 2026, check for later amendments, European Commission guidance and harmonised standards before relying on it. It is general information, not legal advice, and AIPEX accepts no liability for decisions taken on it without specific professional advice.