An operations manager joining a workspace where agents already run must own four things by the end of week two: a map of every agent and its depth, a capacity plan that includes review of held work, the weekly correction rhythm, and a plan for the day an agent drops to depth 1. Week one builds the picture.
What does an operations manager oversee when agents run?
Agents in an operations team do routine work at depth 3 within the business's limits, draft work for people at depth 2, and hold or refuse what is outside their bounds. People review held work, decide exceptions and correct the agents. The operations manager keeps the whole system running: enough reviewers, trained to the right level, a cadence that turns corrections into improvements, and a fallback when an agent's depth changes.
That is a different job from managing a team that does every task by hand. The role is described in what an AI operations manager does, and the term itself in what is an AI operations manager. Onboarding is owned by the head of operations; the outgoing manager or the agent owner guides.
What happens in the first two weeks?
- Day 1. Person (head of operations): hands over the agent register: agents, owners, certified depth, certificate expiry, customer limits.
- Day 1 to 2. Agent, depth 1: summarises the last month per agent from the record: allowed, held, refused, withheld verdicts; corrections by reason; any drops to depth 1.
- Day 2 to 3. Person (new manager): sits with reviewers on the busiest queue, timing how long held items take.
- Day 4 to 5. Person (new manager): rebuilds the capacity plan, adding review time per queue to the existing workload model.
- Week 2. Person (new manager): chairs the weekly correction review with team leads and agent owners, deciding which patterns to bank, watch or close.
- Week 2. Person (new manager): writes the depth-drop plan per agent: if it fell to depth 1 tomorrow, where would the work go and who would do it?
- Week 2. Person (new manager): maps reviewer training against the agents each person reviews, and books renewals.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Routine work | Executes within limits | 3 | Exceptions |
| Drafts | Prepares for review | 2 | Approval |
| Monthly summary | Reads the record | 1 | Interpretation |
| Capacity plan | Nothing | none | Manager |
| Correction review | Supplies grouped reasons | 1 | Bank, watch or close |
| Depth-drop plan | Nothing | none | Manager |
Which numbers should the new manager learn first?
Held volume per queue. This sets review workload. It changes when bounds change or new work types arrive.
Time to decide a held item. With held volume, this gives reviewer hours.
Correction rate by reason. This shows where agents need improvement, and whether banked changes worked.
Certificate dates. Each agent's certificate lapses after 90 days. A re-certification that clears no depth drops it to depth 1 the same day and returns all its routine work to people. Knowing the dates lets the manager schedule re-certification away from peaks.
Reviewer certificates. People's certificates last 12 or 24 months depending on level, and a reviewer whose certificate has lapsed may no longer hold approval rights if the business ties the two.
What decisions belong to the operations manager?
Staffing and priorities, the cadence of correction reviews, and the timing of changes. Not the wording of an agent's brief, which belongs to the agent owner, and not raising an agent's depth, which is earned through Attest evidence within the role's ceiling. A new manager under queue pressure will be asked to "let the agent do more"; the answer is to gather the evidence, not to change a setting.
The manager also decides how the team learns from incidents. When a held action is approved wrongly, or an agent's verdict reaches a customer in a way nobody intended, the record shows what happened step by step. The new manager should run the first such review with the agent owner present, focus on which bound or person decided, and end with one change: to the agent, to a reviewer's training, or to the process. Reviews that look for causes rather than culprits keep reviewers recording what really happened.
Which controls and evidence should onboarding leave?
- A reviewed agent register. Evidence: the manager's annotated copy.
- A capacity plan including review. Evidence: the plan, signed off.
- Correction review minutes. Evidence: patterns and decisions from week two.
- Depth-drop plans. Evidence: one per agent.
- Training map. Evidence: reviewers matched to certificates and expiry.
What should the manager measure from then on?
Held items waiting beyond target. Repeat corrections after a banked change. Unplanned drops to depth 1. Reviewers approving without a current certificate. Time from a correction pattern appearing to a live change.
Where does it usually go wrong?
Treating agents as headcount. A manager who counts agents as reduced staff forgets the review work they create and the fallback they need.
Cancelling the correction review. It is short and never urgent, so it slips, and agents stop improving.
No plan for depth drops. The first failed re-certification lands in a peak week and the backlog takes days to clear.
Managers in nearby roles face the same questions; see the plans for a new risk manager and a new property manager. Where an organisation trains its people in order, operations is usually early, as who in a wealth manager needs AI agent training first shows for one sector. The wider approach is in training teams for AI agents.
Questions people ask
Should an operations manager be able to review held items?
Yes. They should be trained to the same level as their reviewers, so they understand the work and can step in.
How often should the capacity plan change?
Whenever an agent's bounds or depth change, and at least monthly while agents are new.
What if an agent owner disagrees with the manager's priorities?
Settle it in the correction review with evidence from the record: volume, harm and cost of each pattern.