Most plans for AI agents in operations ask one question: how many fewer people will we need? The better question is what the remaining people will do. Agents take over routine processing, which shifts human work towards reviewing drafts, handling exceptions, owning sources and deciding held items. Roles, skills and career paths have to be redesigned around that shift.
What work moves to agents, and what stays?
Agents take on the high-volume, well-documented middle of a process: reading inbound requests, gathering the facts, applying the rules, drafting the outcome, and for proven capabilities, executing routine actions within limits. That is the work most operations staff spend most of their day on.
What stays with people falls into a few kinds. Decisions that need judgement or authority: approving held items above a limit, handling vulnerable customers, making exceptions. Cases the agent cannot process: unusual documents, missing data, conflicting instructions. Ownership of what the agent relies on: keeping procedures, terms and policies current. And supervision of the agents themselves: watching queues, reading refusals, assembling evidence.
The result is that the volume of human work falls, but its difficulty rises. The easy cases go. The hard ones remain, and arrive more concentrated.
How does the shape of the team change?
A traditional operations team is wide at the bottom: many processors handling routine work, fewer team leaders, a small number of specialists and managers. Agents remove much of that base.
| Role group | Before agents | After agents |
|---|---|---|
| Processors handling routine cases | Largest group | Much smaller; many move to review or exceptions |
| Reviewers and named approvers | Part of team leaders' work | A distinct, skilled role with protected time |
| Exception handlers | Informal, whoever is free | A defined team for cases agents cannot process |
| Source and procedure owners | Occasional, often neglected | Ongoing work, because procedures are now agent instructions |
| Agent operations lead | Did not exist | Owns queues, certification, widening and incidents |
| Managers | Supervise people doing tasks | Supervise people supervising agents |
The team ends up narrower at the base and thicker in the middle. More of its people need judgement, product knowledge and the confidence to refuse.
Why is this different from earlier automation?
Earlier automation, such as rule-based robotic processes, typically replaced narrow, fixed steps: copy this field, move that file. People still did the reasoning in between. Agents reason across the whole case, read unstructured documents and draft whole outcomes. They take on more of what was considered skilled work, which is why the team's shape changes and not only its headcount. Our explainer on whether an AI agent is just a chatbot with tools sets out what distinguishes agents from those earlier tools.
What happens to how people learn the job?
This is the problem most plans miss. In a traditional team, people learn by processing routine cases for months before they handle hard ones. The routine work is the training ground. If agents take the routine work, new staff go straight to exceptions and review without the grounding that made experienced reviewers good at it.
There are workable answers. New starters can shadow the agent's work, reviewing its drafts with a mentor before they approve anything themselves. A share of routine cases can be kept with people deliberately, as training rather than production. And reviewers can be rotated through exception handling so their judgement keeps developing. What does not work is assuming the next generation of reviewers will appear on its own.
How should managers plan the transition?
As a change programme, not a technology rollout. People need to know what their job will become, what they will need to learn and how they will be measured. Our explainer on what change management is covers the general discipline; for agents, three things matter most.
First, redesign roles before the agent widens, not after. If an agent is about to move from drafting to executing a capability, the people who were reviewing every draft need a new role ready.
Second, protect time for review. A reviewer expected to approve held items between other tasks will rubber-stamp. Review needs to be the job, or a defined part of it.
Third, change measures. A processor measured on cases closed becomes a reviewer measured on decision quality, or an exception handler measured on resolution. Keeping old measures pushes people back into old behaviour.
Does this look different by market or sector?
The direction is the same, but the pace and detail vary with regulation, labour markets and the type of work. In payments operations, for example, the exception work that remains with people is shaped by scheme rules, fraud patterns and local supervisory expectations. Our brief on agents for payments operations leaders in Singapore shows how the same shift plays out under one market's rules.
What is the trade-off?
Reshaping a team costs more than shrinking it. It means retraining people, creating roles that did not exist, keeping some routine work with people for learning, and accepting that the remaining team is more skilled and may be more expensive per head.
Shrinking without reshaping costs more in the end. The team that remains is not equipped for the work that remains, review becomes a formality, exceptions pile up and the agent's bounds stop meaning anything because nobody can exercise them. Before your next agent widens, draw the team as it will be: who reviews, who handles exceptions, who owns sources, who runs the agents and how new people will learn. Then plan the move from today's shape to that one. The deployment approach this supports is in our guide to deploying an AI agent: train, shadow, certify, widen.
Questions people ask
Do agents always reduce operations headcount?
Not always, and not immediately. Some organisations use the capacity to handle more volume or better service with the same team. Where headcount does fall, it usually falls among routine processors while review and exception roles grow.
Which existing staff make the best reviewers?
People with deep process knowledge and the confidence to disagree. Experienced processors and team leaders often fit well, provided they are given the time, information and authority that real review requires.
Should operations staff learn to build agents?
Some will want to, and it helps to have people who understand both. But the more important skill for most is supervising agents: reading their drafts critically, understanding why an action was held, and knowing when to escalate.