How to assign an owner to every AI agent

Every agent needs one named person who answers for it: the owner of the business process it serves. Not the builder, not IT, not the vendor and not a team mailbox. The owner accepts the role in writing, names a deputy, and is replaced the day they move on. An agent without an owner is reduced to read-only.

What does an agent owner do, and who appoints them?

The owner is accountable for the agent's outcomes. They approve its capabilities and depth, name the people who approve its held actions, review its refusals and incidents, sign its certification and attest to its inventory entry each quarter. When something goes wrong, they are the person a regulator, auditor or customer complaint reaches.

The executive accountable for the business area appoints owners, usually on the agent operations lead's recommendation. The natural choice is the process owner: the person who already owns the work the agent does. That keeps accountability where the knowledge is, and it is why the process owner should sign off certification, not the vendor.

What are the steps?

1. Agent, depth 1: lists agents without an owner, or whose recorded owner has left, changed role or is a group rather than a person.
2. Agent, depth 1: identifies the process each agent serves, from its brief and the systems it acts in.
3. Person, agent operations lead: proposes an owner for each: the person who owns that process.
4. Person, accountable executive: appoints the owner.
5. Person, owner: accepts in writing, confirming they understand the duties and have read the agent's brief, capabilities and depth.
6. Person, owner: names a deputy who acts in absence.
7. Agent, depth 3: records owner and deputy on the inventory and as sponsor on the agent's identity.
8. Agent, depth 1: watches for changes: the owner leaving, moving role or going on long leave triggers a handover task.
9. Person, agent operations lead: applies the orphan rule. If an agent has no accepted owner after a short grace period, it is limited to read-only until one is appointed.

Which controls make ownership real?

One person, by name. A team cannot be accountable. Evidence: a named individual on every inventory entry.

Written acceptance. Ownership that was assigned but never accepted tends to be ignored. Evidence: the signed acceptance.

Duties tied to actions. Certification, approver rosters and quarterly attestation all require the owner's sign-off, so the owner cannot be passive. Evidence: the owner's signatures on each.

Leaver and mover triggers. HR events update the inventory. Evidence: handover records within days of each change.

The orphan rule. Evidence: depth reductions applied to agents without owners.

An owner is not the same as a person approving each action. Whether every agent needs someone in the loop for each action is a separate question, answered in should every AI agent have a human in the loop? The owner is accountable whether or not individual actions are approved.

Which signs show ownership is going wrong?

• The owner is "IT" or "the vendor". The technology team runs the platform; it does not own the outcomes of a collections process or a claims decision.
• The owner cannot describe the agent's limits. Ask an owner which of their agent's capabilities need depth 3. If they do not know, they are not owning it.
• Shared agents have several owners who disagree. An agent serving three teams with three owners and no tie-breaker drifts. Name one owner and give the others a say in their capabilities.
• Owners rubber-stamp. Quarterly attestations come back in minutes with no changes, year after year.
• Agents outlive their owners. The owner left months ago and nobody noticed.

What should you measure?

Count agents with a named, accepted owner (the target is all of them), days from an owner leaving to a new acceptance, agents under the orphan rule, and attestations completed on time. Also track how many owners hold more agents than they can follow; a single person owning dozens of agents is a warning sign.

Where does it usually go wrong?

Ownership is assigned at build, then forgotten. The person who requested the agent becomes owner by default and later moves on.

Owners are too junior. The owner must be able to stop the process if needed. If they cannot suspend the agent or change the process, they are not the right person.

Ownership is split by phase. One person owns it in pilot, another in production, and the handover is lost. Name the production owner before the pilot starts.

Depth is treated as the owner's preference. The owner sets caps per capability, but evidence sets what the agent has earned; see how to assign depth to each of an AI agent's capabilities.

Where Pryme fits

In Pryme, certification and depth are per agent, and the owner signs off the evidence from Attest rather than a vendor's claim. Certificates lapse after 90 days, which brings the owner back to the agent on a fixed rhythm. The deployment path is described in deploying an AI agent: train, shadow, certify, widen.