Your company is going to need someone to manage the AI agents.
Not build them. Not write the prompts for them.
Watch them.
Harvard Business Review published a piece in February naming this as the highest-priority new role in companies moving AI from pilot to production. They call it the agent manager. When you read what the job actually requires, you're looking at a program manager with a different team.
The difference: the direct reports are AI agents, not people.
The Role: What Harvard Found
The HBR article tracks Zach Stauber at Salesforce. His full-time job is managing a fleet of generative AI agents across support, sales, and marketing on a platform the company calls Agentforce.
His daily routine: "Data, Data, Data. I start and end my day in dashboards, scorecards, and agent observability monitoring."
That is a program manager's day. The only thing that changed is what he's monitoring.
An agent manager does five things:
- Set goals for each agent. What is this agent supposed to accomplish? What does "working well" look like?
- Define the numbers that matter. Resolution rate. Escalation rate. Output review rate. The metrics that tell you the agent is on track.
- Catch drift early. Agents don't break cleanly. They degrade. The output slowly gets worse until something fails. The agent manager sees the numbers moving before the failure happens.
- Retrain or adjust. When an agent drifts, the agent manager investigates and makes the fix: prompt adjustment, new examples, parameter change.
- Handle cases the agent can't. Every agent has an edge. The agent manager is the escalation path.
Salesforce doesn't have one person doing this. They have a team.
Why This Role Is Not Technical
The HBR finding that got attention: domain expertise matters more than technical skill for this role.
Companies don't want engineers. They want people who deeply understand the work being automated.
That makes sense. An agent running customer support gets evaluated on response time, resolution rate, escalation frequency, and satisfaction scores. None of those require knowing how a language model works. They require knowing what good customer support looks like.
If you already do that job well and you use AI regularly, you're qualified for the agent manager role.
The technical layer, model selection, prompt structure, API setup, is a setup task. It happens once. The ongoing job is judgment, observation, and correction. That's management. You already know how to do it.
The Mistake Most Businesses Make
They try to hire a technical person first.
They post for an "AI engineer" or "LLM developer" and wait for someone who can build agents from scratch. Meanwhile, the agents they already have are drifting, producing inconsistent results, and nobody is watching the numbers.
The person you need to watch the agents is not a developer. It's your best operations person. The one who already knows how everything is supposed to work. The one who notices when a process goes sideways before anyone else does.
Give them a clear charter: these three agents are yours. You own the performance. You flag the problems. You propose the fixes.
That person exists on your team right now.
What "Good" Looks Like: 2 Numbers to Start
If you have agents running and nobody is monitoring them, start with two metrics.
Escalation rate. What percentage of tasks does the agent escalate to a human? If it was 12% last month and it's 18% this month, something changed. Find out what.
Output review rate. If a human reviews the agent's output before it ships (a reply goes out, a document gets sent), what percentage of those reviews result in an edit? Rising edits mean the agent is drifting.
These are not sophisticated metrics. They're what a new manager tracks in their first month with a new team. Start here. Add more once you have a baseline.
Honest Limits
This article covers the role, not the technology behind it. If you want to understand how to build agents from scratch, the HBR piece goes deeper on the Salesforce setup and links to more technical resources.
The agent manager role assumes you already have agents running. If you don't, this article is a preview of a job that becomes relevant as soon as you deploy.
The Harvard framing also focuses on enterprise-scale deployments (Salesforce's Agentforce fleet). In a smaller business, the agent manager role is one part of someone's job, not a full-time position. The responsibilities are the same. The time commitment scales with the size of the fleet.
One Move This Week
Pick one agent you currently run. Write down three numbers that would tell you whether it's working. Check those numbers this week.
If you find drift, investigate it. That process, noticing, investigating, adjusting, is exactly what the agent manager does every day. You don't need the title to start doing the job.
The HBR article is linked below. It's behind a paywall for some readers, but the core argument is here: the person who understands the work is the right person to manage the machine doing the work.
Your best ops person is your agent manager. Give them the agents.