Agents in production generate operational work every day: held items waiting for people, refusals to route, certificates to renew, evidence to weigh before widening, incidents to contain. If that work belongs to everyone, it belongs to no one. Organisations running more than a couple of agents need a named agent operations lead whose job it is.
Why does this work fall through the gaps?
Because it sits between teams that each own part of it. Technology built the agents and keeps them running. Process owners answer for outcomes. Risk sets the framework. Frontline teams approve held items. Each does its part, and nobody looks at the whole.
The symptoms are familiar. Held queues grow because nobody is watching staffing. A certificate lapses on a busy Monday because nobody tracked the date. An agent stays at draft for months after the evidence would support widening, because nobody assembled it. A pattern of refusals goes unnoticed because nobody reads refusals. An incident takes hours to contain because nobody is sure who may lower an agent's depth.
None of these is a technology failure. They are operational failures, and they need an operational owner.
What does an agent operations lead actually do?
The role is closer to an operations manager than an engineer. It runs the agents as a working part of the business.
| Area | Daily and weekly work |
|---|---|
| Held queues | Watch volumes and waiting times; match reviewer capacity to demand; escalate spikes |
| Refusals and verdict patterns | Read refusals grouped by deciding bound; route each cluster to the owner who can fix it |
| Certification | Track every agent's certificate date; schedule re-tests ahead of lapse; act on failures the same day |
| Widening and narrowing | Assemble evidence for depth increases; recommend to agent owners; lower depth quickly when needed |
| Sources | Work with source owners to keep policy documents, terms and procedures current |
| Incidents | Lead the first response for wrong-action incidents; run the query for affected cases |
| Reporting | Produce the agent pack for risk committees and the board |
The lead does not own every decision. Agent owners still approve their agents' depth, and named approvers still decide held items. The lead makes sure those decisions happen, on time and on evidence.
Why is widening the part that most needs an owner?
Because widening is where value comes from and where risk enters. An agent kept at draft forever delivers a fraction of its potential. An agent widened on enthusiasm rather than evidence is an incident waiting to happen. Somebody has to sit between those two failure modes.
The lead's job is to make the evidence case. That means gathering scored results from real cases, reviewer edit rates, refusal patterns and incident history for the agent, and presenting them to the agent owner with a recommendation. When the owner agrees, the lead makes sure the change is recorded and the new depth monitored closely for the first weeks.
For agents being prepared for work where real cases are scarce, the lead also has to judge what evidence is good enough. Our piece on certifying an AI agent without real data looks at the options and their limits.
What skills does the role need?
Operational judgement first. The lead needs to understand how work flows through the business, where queues form and what customer harm looks like. People from operations management, process improvement or operational risk often fit well.
Second, enough fluency with the agent governance model to read the verdict record and reason about bounds: what depth means, how the five bounds interact, what a held item versus a refusal tells you. This does not require building agents or understanding how models work in detail.
Third, the authority and confidence to act. The lead must be able to lower an agent's depth during an incident without waiting for a committee, and to push back when a sponsor wants to widen without evidence.
Where should the role sit?
In operations, reporting to the executive who owns the processes the agents run. Not in technology, which would pull the role towards uptime and away from outcomes. Not in risk, which would compromise risk's independence as a second line.
In a smaller organisation, one person may cover all agents. In a larger one, there may be a lead per business area with a central function setting common practice. What matters is that every agent in production has a named lead, and that lead has the time to do the work rather than fitting it around another job.
How does the role change across markets?
The core work is the same everywhere. What changes is the regulatory reporting the lead supports and the operational context. In payments, for example, the lead's day is shaped by settlement windows, scheme rules and fraud patterns that differ by market. Our briefs on agents for payments operations in Australia and in India show how the same role meets different local demands.
What is the trade-off?
A dedicated role is a cost, and it can feel like a contradiction: agents were meant to reduce headcount, and now there is a new job. Some organisations will try to spread the duties across existing roles instead.
That usually fails for the reasons above. The duties are continuous, they cut across teams and they need someone watching the whole. The cost of the role is small next to the cost of lapsed certificates, unmanaged queues, missed widening and slow incident response. If you run agents in production today, write down who does each item in the table above. If the answers are "everyone", "the vendor" or blank, you have found the job description. The operating model it supports is set out in our guide to deploying an AI agent: train, shadow, certify, widen.
Questions people ask
How many agents can one operations lead manage?
It depends on the depth and volume of the agents more than their number. A handful of agents executing high-volume work at depth 3 or 4 needs more attention than many agents that only read or draft. Watch queue health and certification timeliness as signs the lead is stretched.
Should the agent operations lead approve held items?
Generally not as a routine approver, because it blurs their role as the person watching the queue. They may step in to clear spikes or handle escalations, but held items should go to named approvers in the business function the action belongs to.
Is this the same as an AI product manager?
No. A product manager decides what agents should be built and how they should work. The operations lead runs them once they are live. The two should work closely, especially when operational evidence suggests a change to the agent's design.