Entrusting an AI agent as you would a senior postThe delegation charter in seven questions

Consider a multi-site business: some thirty branches, a few hundred people, a head office. The executive team does not see the order placed this morning in Lyon, the exception granted to a customer in Nantes or the recruitment begun in Lille. It remains accountable for them. That is why it delegates, and it delegates in two ways. To its directors, a mandate (decisions, financial limits, escalation, reporting, accountability). To its employees, tasks and procedures. The mandate belongs to the director, the procedure to the employee. The whole of governance rests on that distinction.

When an AI agent arrives in the business, this discipline is often lost. It is given an objective, a few instructions and access rights, and people watch what it produces. If it succeeds, its remit grows. A team may even connect it to its own data without anyone at head office knowing: a delegation that no one granted. The risk is not that the agent lacks intelligence. It is that the business has entrusted it with, or allowed it to take, a power it has never governed.

After twenty years in general management, and since I began designing and running AI agents in production myself, my view is simple: an AI agent should never be given more latitude than a human executive in a comparable role. Here are the seven questions I ask, and which you can put to your next executive committee.

The seven questions

1. The mandate. Could someone outside the project read the mandate and explain what the agent is permitted to do and what it is not? ‘Optimise our purchasing’ is not a mandate: analysing, proposing, amending and ordering are different powers.

2. Autonomy, action by action. Does each sensitive action have an explicit level (autonomous; autonomous with a report; subject to approval; prohibited)? The level depends on the consequence, not on the technical difficulty. Anything that commits the business, affects a third party or is difficult to reverse requires human approval.

3. Evidence. Can the reported result be confirmed by data independent of the text the agent has produced? An agent can produce a faultless report that is wrong. I have had an agent report six files indexed when all six had failed. Nothing that has not been checked is presented as complete.

4. The log. In the event of an incident, can we reconstruct the facts without relying on the memory of those present? Each agent has its own identity, each alert an acknowledgement, each silence a follow-up.

5. Shutdown. Who can stop the agent, in what circumstances, and how does the business carry on afterwards? A stop button that has never been used is a hypothesis.

6. Memory. Do we know what information currently influences the agent’s behaviour, and who approved it? Ungoverned memory turns a coherent system into a series of decisions that no one can any longer explain.

7. The person accountable. Is one named individual answerable for both the risk and the value produced?

Where people go wrong in good faith

A director who discovers what an agent can do will, quite reasonably, want to make the most of it. He or she will hand its supervision to a member of staff who monitors the tool, reads its reports, adjusts its settings and flags anything amiss. That member of staff is carrying out a monitoring procedure. They are not accountable for what the agent decides or sets in motion. That accountability remains with the director who authorised the agent.

An obstetrician entrusts the monitoring of a pregnancy to a midwife: a useful delegation that frees up medical time. If something goes wrong at the birth, no one considers that the obstetrician has shed responsibility for the birth by delegating the monitoring. A task was delegated; accountability for the decision was kept. The agent is the pregnancy being monitored, the employee overseeing it is the midwife, and the director remains the doctor. A committee that cannot say, for each agent, who the doctor is has not delegated. It has installed a tool and hoped for the best.

What this means for the head of an owner-managed or mid-sized business

You do not need to be a technical specialist to govern your agents. You need to do what you already do with your directors: write the mandate, set the thresholds, require evidence and name the person accountable. ISO/IEC 42001, the first management system standard for AI, says nothing different. It makes this a certifiable framework, which is useful once your larger customers begin to ask for it. The EU AI Act has deferred its most demanding obligations to late 2027 and 2028. That delay is no reason to wait: making an inventory of your agents and writing their mandates creates value now.

If any of the seven answers remains unclear, the agent is not ready for more autonomy. That is not a failure. It is the information that was missing in order to decide.