AI agents, software that carries out tasks by itself, are starting to do daily work across company departments. Once that happens, letting teams use them informally becomes a management problem. Someone has to own the agents’ behaviour, quality and risk, including how their mistakes get caught. Harvard Business Review argues that companies need that person to thrive in the AI era, and calls the role the agent manager. The job is to manage how agents learn, collaborate, perform and work safely alongside people.
HBR’s example comes from Salesforce, where the job exists in practice. Its support agent manager oversees generative AI support agents across support, sales and marketing. His work includes dashboards, scorecards and tracking what agents do and where they fail. That gives the business a way to see how each agent performs before small errors become larger problems.
The economic case is direct, and it cuts both ways. A well-managed agent can reduce repeated work, speed up handoffs and help teams respond faster without adding the same level of headcount. An unmanaged agent can send wrong answers, duplicate tasks and create extra review work. The value comes from the agent’s performance inside the team’s everyday workflow.
That puts operational discipline at the centre of the role. A customer support agent needs clear rules, and so does a finance, legal or marketing agent. The agent manager sets those rules before teams rely on agents for important work. Without them, every department builds its own habits, measures quality differently and handles errors in its own way.
This job is management work. A manager of people looks at output, quality, training and risk, and an agent manager applies the same discipline to AI systems. Writing the instructions agents follow is part of it, but the deeper task is coordination. That takes someone who understands the work, the customer impact and the risks across departments.
The role does not need a deeply technical person at the start. Technical knowledge helps, but the core skill is judgement: knowing when an agent can act alone, when it needs approval and when a human should stay in control.
