The next decade of customer operations will run on agents.
Not in theory, and not at some distant point in the future. It is already happening inside SaaS companies, where onboarding agents, churn-risk agents, renewal copilots, and LLM-powered workflows are being added to customer teams faster than anyone is redesigning the work around them.
What most companies have not fully absorbed is that agents do not manage themselves. Once they begin touching real accounts, surfacing real recommendations, and shaping real customer decisions, someone has to decide which ones to trust, which ones to tune, which ones to override, and which ones to turn off.
That person already exists in many organizations. The title does not.
The cleanest name we have for that role is the agent operator: the function-embedded, technically fluent operator responsible for making agents reliable, usable, and valuable in production.
Consider the CS lead at a Series C SaaS company who started her Monday this quarter with 42 accounts to manage, a QBR deck due Wednesday, and a stack of agentic systems she never explicitly signed up to run. One was an onboarding agent shipped by the platform team in February. Another was a churn-risk agent her VP had bought the year before and nobody ever disabled. A product manager had prototyped a renewal-prep agent over a weekend. Salesforce and Gainsight each had their own so-called AI copilots. Engineering had wired up an MCP connection to Claude so she could ask questions in plain language.
Nobody told her that supervising these systems was now part of her job. Nobody rewrote the role. But when the churn-risk agent flagged an account that product usage said was healthy, she was the one who had to decide what to trust, whether to dismiss the alert, and how that judgment should feed back into the system.
That is already a job. In a growing number of companies, it is becoming a critical one.
Why the Existing Titles Don’t Fit
Whenever this comes up, someone suggests the role already exists under another title. In practice, that usually makes the gap clearer.
It is not an AI engineer. The AI engineer builds the agent, or the infrastructure around it. The agent operator lives with the system after launch, when the hard question is no longer whether it works in a demo, but whether it should be trusted this week, on this account set, under real operating conditions.
It is not a prompt engineer either, assuming that label survives at all. The operator’s unit of work is not the prompt. It is the live behavior of an agent, or more often a portfolio of agents, across real workflows over time.
RevOps or CS Ops gets closer, and in many companies the role will emerge from one of those seats first. But traditional ops has mostly been about process, reporting, and human execution. Operating agents is different. Agents act semi-autonomously, fail in unfamiliar ways, and can compound bad logic quietly until a customer gets an email that makes it obvious nobody was really in control.
And it is not a Head of AI. That is usually a strategy role. The Head of AI decides what the company should invest in. The agent operator is doing daily operational work. They decide what stays live, what gets tuned, what gets escalated, and what gets turned off.
The reason existing titles do not quite fit is simple: this role sits in the gap between building an agent and trusting it in production. That gap is where a large share of operational work is about to move.
What the Agent Operator Actually Does
At a high level, the job has four responsibilities.
First, they run a portfolio. Every agent has a scope, a set of triggers, a set of accounts or workflows it can act on, and some threshold for escalation. Someone needs the portfolio view: which agents are firing, which have gone quiet, which are drifting, which are being acted on, and which are technically live but behaviorally dead because nobody trusts them anymore.
Second, they own the evals. An agent is not truly shipped when it is deployed. It is shipped when somebody is continuously measuring whether it is still making good calls. That means maintaining gold-standard examples, sampling outputs, reviewing edge cases, and asking whether the agent’s recommendations would hold up under the same scrutiny applied to a human decision.
Third, they arbitrate between systems. A churn-risk agent says the account is in trouble. A health model says the same account is fine. A copilot drafts an expansion email at exactly the wrong moment. Someone has to decide which signal carries weight, when uncertainty should be escalated, and how contradictory outputs get resolved before they become background noise.
Fourth, they design the handoff to humans. Every agent eventually reaches a person: a CSM, an AE, a support rep, a manager, a legal reviewer, a marketer. The operator shapes that interaction. What context gets passed through. What recommendation is made. How much confidence the system claims. What happens when the human disagrees. How that response feeds back into future behavior. Bad handoffs are why agents get muted. Good handoffs are why they become useful.
Why the Role Is More Technical Than It Sounds
It would be a mistake to hear the word operator and imagine a purely managerial role.
The agent operator is function-embedded, but they are also technically fluent. They do not need to be the best engineer in the company, but they do need to understand how agents are actually wired into work. They need to know how context moves between tools, where permissions break, why one workflow is reliable while another is brittle, and what it takes to turn a promising model into something a team can safely use.
In practice, that means familiarity with things like MCPs, CLIs, eval loops, workflow wiring, and the internal skills or lightweight tools that make agents usable inside a function.
Not because they are replacing engineering. And not because every operator needs to become a full-time builder. But because the person responsible for creating leverage from agents cannot be afraid of the machinery.
That is what makes the role important. This is the person who can walk into a marketing team, a legal team, an operations team, or a life sciences research group and make agents useful in context. They understand enough of the technical layer to shape the system, and enough of the functional layer to know where the leverage really is.
In that sense, the agent operator is not just supervising agents after deployment. They are the bridge between frontier model capability and day-to-day departmental execution.
The Skill Stack
The people who are good at this tend to share a specific mix of operating and technical judgment.
They have schema fluency. They know what is in the CRM or system of record, which fields are trustworthy, where the data breaks, and which source to believe when two systems disagree.
They have eval discipline. They care less about whether something looks impressive in a demo than whether it was right on a meaningful sample of real decisions. They are willing to spend an afternoon reviewing outputs by hand if that is what it takes to understand failure modes.
They have escalation judgment. Some errors are cheap. Others are catastrophic. Misclassifying a low-touch trial may not matter. Prompting a CSM to push expansion during a procurement dispute absolutely does. A surprising amount of the role comes down to knowing the difference.
They have technical fluency. They understand enough of the stack to work productively with it: MCPs, CLIs, workflow scaffolding, agent behavior, and the skills or tools that let agents operate inside a real team.
And they have voice in the room. An operator has to be able to tell a VP of CS that one agent is helping, another is noisy, and a third should not be live on strategic accounts this quarter. Without that credibility, the systems drift, trust erodes, and the program quietly dies.
Where the Role Comes From
In most companies, this role will not arrive through a clean job description. It will emerge.
Sometimes it is the RevOps lead who moved from dashboards into systems work and now spends part of the week tuning automations that have become something closer to non-human teammates. Sometimes it is the CS Ops manager who wrote playbooks, then workflows, then SQL, then LLM calls layered on top of the workflow stack, until they were effectively running the agents whether anyone had named it or not.
Sometimes it is a product manager who built an internal tool, watched it evolve into an agent, and kept operating it because nobody else could. That path is less common, but when it works, it often produces unusually strong operators because the person understands both the product logic and the organizational friction around adoption.
People with engineering backgrounds can absolutely do the job. But the common failure mode is predictable: they over-index on the model and under-index on the function. The strongest operators tend to understand both the machinery and the consequences of getting the workflow wrong.
Where the Role Goes Next
Customer operations may be where this role becomes obvious first. It will not be where it ends.
The same pattern will show up anywhere agents begin doing real work inside a function. Marketing teams will need someone who can operationalize agent workflows instead of collecting AI experiments. Legal teams will need someone who understands where model assistance is useful, where it is risky, and how review loops should work. Operations teams will need someone who can turn agents into leverage rather than noise. Research-heavy organizations, including life sciences teams, will need people who can translate model capability into domain-specific workflows without pretending the system is trustworthy by default.
The title may vary by department. The underlying job will not.
This is the person who makes agents real inside a function. Not by talking about transformation, but by wiring the systems, testing the outputs, earning trust, and making the workflow hold under live conditions.
As more work gets delegated to agents, companies will need people who are explicitly accountable for operating them. Those people will usually sit inside the function they serve, not inside engineering. They will be close enough to the work to understand what matters, and technical enough to make the systems useful.
The exact title may change. The need will not.
The Standing Ask
We are writing this down because the work is already here, even if the vocabulary is not.
If you are the person on your team who has quietly become responsible for deciding which agents to trust, how they should escalate, how they should be wired into real workflows, and when they should be corrected, you are already doing this job. If you are a founder or functional leader watching that responsibility accumulate under someone else’s title, it is worth naming it and giving it real authority. And if you are a vendor building agentic products without understanding who will actually operate them inside the customer, you are probably designing for the wrong buyer.
The next decade of customer operations will not be defined only by the agents companies deploy. It will be defined by the people who make those agents trustworthy enough to deserve a place in the workflow.
For now, the cleanest name we have for that person is the agent operator.
