AI agents: five questions every executive committee must settle

An AI assistant answers a question. An AI agent is given an objective, works across your applications, carries out a sequence of steps and reports back on what it has done. The difference looks technical. It is not. The moment software acts on behalf of the organisation, the question becomes one for leadership.

These are the five questions we put on the table at the start of every engagement.

1. Which processes do we hand over – and which do we keep?

The best candidates are frequent, well-documented processes where errors are visible and can be corrected: account reconciliations, contract review against standard clauses, handling of routine requests. The worst candidates are rare, poorly described, or commit the organisation irreversibly.

A good use case fits in three lines: a trigger, what the agent does, what the human decides. If it takes a page to explain, it is not ready.

2. Who owns an agent?

An agent without an owner is a risk without an owner. Every agent needs a named business owner who answers for its results just as they answer for those of their team. IT keeps it running; it does not own the decision.

3. Where does autonomy stop?

An agent’s autonomy is calibrated, much like a delegation of signing authority. Up to what amount can it propose a credit note? Can it write to a customer, or only draft the message? These thresholds are set down in writing before go-live – and revisited in light of experience.

A well-governed agent follows the same rules as an employee: a role, an owner, clear limits and accountability.

4. How will we know it works?

A convincing demonstration proves nothing. Before go-live, the agent is tested on real cases – difficult ones included – against criteria set in advance: accuracy, turnaround time, cost and the rate at which a human has to step in. After go-live, the same measures continue. What is not measured cannot be managed.

5. What data does it use, and where does that data go?

An agent reads your data and sometimes passes it to a model hosted elsewhere. You need to know which data, where it goes and under what contract. The EU AI Act also classifies uses by level of risk: some, such as recruitment or creditworthiness assessment, are deemed “high-risk” and carry specific obligations. Better to know before you build than after.

The bottom line

None of these five questions requires technical expertise. Each of them requires a decision. That is why agentic AI starts in the boardroom, not in the lab.