How many people do you need to supervise AI agents?

There is no fixed ratio of people to agents. The number comes from the work the agents send back: held actions, drafts to send, escalations and the samples you check. Multiply each by the time a proper review takes, add the roles every agent needs, and plan for your peak with the agents one depth lower than normal.

Why is agent count the wrong starting point?

Two agents can need very different supervision. One that reads documents and files summaries at depth 1 may need an hour of sampling a week. One that drafts every customer reply at depth 2 may need several full-time reviewers. What drives supervision is the volume and type of work that comes back to people, and how carefully each item must be checked.

Depth changes the answer more than anything else. Moving a capability from depth 2 to depth 3 can cut review work sharply, because a person no longer sends every item. Moving it back, after a failed certification run, restores all of it overnight.

What roles does every agent need, whatever the volume?

• An owner who answers for the agent and can narrow or stop it.
• Named reviewers for held items, with deputies.
• Someone on call during the hours the agent works.
• A platform owner for access, versions and connections.

One person can hold several of these roles for a small, low-risk agent. They cannot all be the same person for an agent that moves money, because the owner should not be the only reviewer of their own agent's work.

How do you work out the reviewing hours?

Use hold and escalation rates from shadow runs, not from a vendor's estimate. Then add a margin for peaks and for the day the agent drops a depth.

What does the calculation look like?

Say a bank's business support team runs one agent at depth 2 across 1,500 customer messages a week. It drafts a reply to 1,200, escalates 300, and the team samples 50 of the drafts in depth each week.

Reviewers spend about two minutes per draft, 40 hours. Escalations take ten minutes each, 50 hours. The sample takes another five hours. With the owner's weekly review, the agent needs nearly 100 hours of supervision a week, about three people. When half the messages later move to depth 3 after certification, draft reviews fall to 20 hours and total supervision to about 80. The team keeps the third person rostered for six more weeks, until the new depth has held through a month-end.

Supervising well is a skill that has to be built. The EU AI Act expects deployers to ensure staff have enough AI literacy for their role, covered in Article 4 for teams that supervise agents, and our view is simply that approvers need training too.

Which controls stop supervision from thinning out?

• A load limit per reviewer, so volume above it calls in a deputy rather than speeding people up.
• Sampling of approved items, to show whether reviews are real.
• No reduction in supervisors until a new depth has held for a full cycle.
• A written decision on which work stays with people, as set out in deciding what stays with people when an agent joins a team.

As agents multiply, supervision becomes a team design question; see managing a team of AI agents and people and scaling from one AI agent to many. The wider sequence is in our guide to deploying an AI agent.