Running the rhythm is the operations manager's part. Before agents join the floor, they should learn to read the workspace, plan reviewer capacity for held work, track certificates before they lapse, interpret verdicts and the bound that decided each one, and manage usage and cost. Reviewers judge single cases; the operations manager keeps the whole system flowing and safe.
How does an operations manager's job change?
Operations managers already balance demand, staffing and service levels. Agents add a new kind of worker with unusual properties. It works at any hour, it never tires, and its authority can change overnight. When a certificate lapses or a run fails, an agent that was executing routine work at depth 3 drops to depth 1, and everything it was doing lands in a human queue.
The manager who understands this plans for it. The manager who does not finds out at nine on a Monday morning, with a backlog nobody forecast.
What should operations managers understand?
- Held work is a capacity problem. Every held action waits for a named person. If agents draft 2,000 items a day and 300 are held, someone must review 300 items well. That is a staffing number, not a technical one.
- Depth and certificates set throughput. Depth is earned per agent and certificates last 90 days. A calendar of certificate expiry dates is as important as a holiday rota. Re-certification needs real cases, so it must be planned before expiry, not after.
- Verdicts are management information. Allowed, held, refused and withheld, each with the bound that decided, tell the manager where work is getting stuck. A rise in "held: own limits" may mean a limit is set too tight; a rise in "refused: sources" may mean a data feed has broken.
- The minimum chain. Role scope, certificate, sources, the organisation's own limits and the platform guardrails apply to every action, and the lowest wins. Managers should know which bounds they own (usually the limits) and which they do not.
- Usage and billing. Agents consume capacity that costs money. Managers should read usage by agent and by task, and know which work is worth automating and which is cheaper to leave with people.
What does the agent do, and what does the manager own?
| The agent | The operations manager |
|---|---|
| Executes routine work within bounds | Decides which work goes to agents at all |
| Holds anything outside its bounds | Staffs the held queue and sets review targets |
| Reports its own verdicts in the workspace | Reads the patterns and acts on them |
| Drops to depth 1 when a certificate lapses | Plans re-certification before expiry |
Risk managers look at the same workspace from another angle; compare what risk managers should learn first. Property managers face a version of this with tenants and contractors, set out in what property managers should learn first.
What does a well-run week look like?
Take an insurer's operations team running a claims intake agent and a correspondence agent. On Monday the manager checks the certificate calendar and sees the intake agent's certificate expires in twelve days; re-certification is booked for Thursday on last month's cases. The held queue ran at 180 items a day last week, so three reviewers are rostered with protected time.
On Wednesday refused verdicts naming sources jump on the correspondence agent. The manager traces it to a policy library update that has not been approved for agent use, and asks the knowledge owner to publish it. By Friday the refusals clear. Nothing dramatic happened, which is the point.
E-commerce operations in India and insurance operations in Germany run this rhythm under different pressures; see operations manager training in e-commerce in India and in insurance in Germany.
Which controls should operations managers insist on?
A named owner for every agent. A held queue with a service level and enough reviewers to meet it without rushing. A certificate calendar visible to the whole team. A rule that nobody raises a limit to clear a backlog without the risk owner's sign-off. And an incident route that everyone knows. The argument for managers as the deciding check is in why operations managers are the real control, and the role map in AI agent skills by role.
Where the Academy fits
The Operator track at Pryme Intelligence Academy was built around this role. Its eight modules include reading the workspace, keeping knowledge current, when something goes wrong, usage and billing, and running the rhythm. A manager preparing a whole team can start with the free readiness assessment for up to 25 people, which reports scores by concept, the gaps and the first agent to start with.
Questions people ask
How many reviewers does a held queue need?
Work it out from held volume and the time a careful review takes for that kind of item. Then add slack for lapses, when held volume jumps.
Should operations managers be able to change agent limits?
They should be able to propose changes. Approval should sit with the risk owner, so pressure on throughput never quietly loosens a control.
What is the first thing to check each morning?
The held queue and any certificate lapses overnight. Those two tell you whether the day will run as planned.