Personalized AI infrastructure
Agent Operator Job Description: Role, Skills, Approvals and Metrics
A practical agent operator job description: duties, skills, approval limits and success metrics to write into a hiring post for your company's AI agents.

An agent operator is the person who runs a company's AI agents after launch: they watch performance, keep memory and tools accurate, review permissions, and decide which changes are safe to approve. The role is not only a prompt writer or a software developer, although either background can help. A strong job post describes the work as decisions, evidence and measurable results, not as a list of tools.
Key takeaways
- The agent operator owns how an agent behaves after launch, which is where most of the ongoing work sits.
- Write approval duties as explicit limits: what the operator may approve alone, what needs the business owner, and what stays frozen until a review.
- Hire for operational judgment and record discipline first, then check technical depth against the systems your agents actually touch.
- Measure the role with observable criteria such as task success, escalation quality, permission hygiene and cost per completed task.
- Document the operator's access and authority so the company can change the person without losing the agent's memory, history or approval trail.
What an agent operator does in a normal week
A business agent is not finished when it goes live. It answers real requests, meets edge cases nobody tested, and drifts as products, policies, prices and staff change. The operator keeps that drift visible and under control. In a normal week, the work usually includes:
- Reviewing a sample of completed tasks and escalations, with written notes on what the agent got wrong and why.
- Checking that each memory source (the documents, records and approved past conversations the agent relies on) is current and clearly labeled.
- Confirming that every connected tool still behaves as expected after a vendor update or an internal system change.
- Logging every change request with its reason, its approver and its rollback path, as described in how to request and approve changes to a managed AI agent.
- Preparing a short status note for the business owner that states what changed, what held steady and what needs a decision.
The operator is also the first responder when something breaks. If an agent sends a message it should not have sent, or a tool call uses the wrong account, the operator pauses the affected capability, collects the evidence and reports it. The rest of the company should not need to diagnose the agent to know what happens next. A written incident routine, with a named person and a stop procedure, is part of the job description, not an afterthought.
Which skills to require in the job post
Describe skills as tasks you can observe in an interview. Phrases like "passionate about AI" tell a candidate nothing and tell you little more. Ask for:
- Process discipline: the ability to keep a change log, follow a review checklist and reproduce a problem before attempting a fix.
- Working knowledge of how language models fail, including confident wrong answers, outdated context and instructions hidden inside user-supplied text. The OWASP guidance on prompt injection prevention is a practical reference for that last risk.
- Comfort with permissions: understanding who can read, write and send what, why shared logins are a risk, and how access should be granted and removed.
- Writing clarity: the operator writes tool descriptions, escalation rules and status notes that a non-specialist can act on without a follow-up call.
- Basic data literacy: reading a sample of records, spotting duplicates and recognizing when a source document has gone stale.
- Business judgment: knowing which mistakes are cheap, which can damage a customer relationship, and which carry legal or financial exposure.
Do not make deep machine learning research the default requirement. An operator who can read logs, write a clear rule and say no to an unsafe change is more useful to a mid-sized company than a researcher who has never run a service in production. Technical depth matters when the agent connects to booking, payment or records systems, so name those integrations in the post and ask candidates to walk through how they would verify one of them.
Approval duties: what the operator can decide alone
Approval duties are the part of most job posts that stays vague, and vagueness causes the most trouble later. Define three tiers before the first operator starts, and have the business owner sign them:
- The operator may approve alone: wording changes to non-binding responses, updates to reference documents that already passed review, and rollbacks to the last approved version.
- The operator proposes and the owner approves: new tools, new data sources, changes to what the agent may say about prices, refunds or contract terms, and any change that widens the agent's permissions.
- The change stays frozen until a formal review: anything touching sensitive personal data, regulated claims, or actions that cannot be reversed, such as sending payments or deleting records.
Write thresholds as examples, not abstractions. "Any change that affects money" is hard to apply on a Tuesday afternoon, while "any change that alters a refund amount or the conditions for one" is not. The NIST AI Risk Management Framework is a useful reference for matching these gates to risk. The gates should follow the risk of the action, not the seniority of the person asking for it. Review the tiers each quarter with the steps in the quarterly AI agent permission review guide, and raise a threshold only when the logs show the agent handled the current level reliably over a sustained period.
Metrics to write into the job post
A metric belongs in the post only if the operator can observe it from logs or records, and only if it changes a decision. Set baselines during the first month, then agree on targets with the business owner rather than copying figures from a vendor brochure. Useful measures include:
- Task success on a fixed sample of real requests, scored by a person who did not configure the agent.
- Escalation quality: whether handoffs to people arrive with enough context to act without reopening the conversation.
- Correction rate: how often a human has to fix, reverse or apologize for something the agent did.
- Permission hygiene: whether every connected tool has a named owner, a stated purpose and a scheduled review date.
- Memory freshness: whether the source documents and records feeding the agent were updated within the window the business agreed.
- Cost per completed task, tracked against usage so that growth in traffic does not hide growth in waste. The FinOps Foundation's material on unit economics gives a sound frame for this calculation.
Avoid measuring activity alone. The number of messages an agent sends says little about whether it helped anyone. Pair each activity measure with an outcome measure and report both in the same status note, so the owner sees the work and its effect together.
An illustrative scenario and a first-month plan
Illustrative scenario: a regional services company runs an agent that answers scheduling questions for its clients. Its operator notices that escalations about rescheduling increase after a policy change, while the agent's replies remain polite and fluent. Tracing the problem, the operator finds a policy document that still shows the old rules. She updates the source, records the change with a date and an approver, and reruns the same sample of requests. The fix came from record discipline, not from a new model. This example is fictional and included only to show the working pattern.
For a first month, the sequence matters more than the speed. A practical plan looks like this:
- Inventory the agent: its purpose, memory sources, tools, channels and named owners.
- Write the three approval tiers and get the business owner's signature on them.
- Build a fixed sample of real requests and record a baseline on each metric in the job post.
- Hold a weekly review of escalations and corrections, with written notes that feed the next status report.
- Run the first permission review and remove any access that nobody can justify in writing.
- Agree the evidence required before any approval threshold is raised, and wait until the logs support the change.
A fuller account of the early weeks is in what the first months with a managed AI agent look like.
Sharing the role and choosing the first task
Most companies cannot justify a full-time operator at the start, and that is workable if the duties are divided on paper. The business owner keeps the decisions that involve money, customers and policy. The operator, internal or external, carries monitoring, evaluations, integration upkeep and incident response. Be Human works this way: its managed operations cover monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements, while the client keeps approval over the thresholds. The first engagement defines the role, its memory and tools, its approval boundaries and its success criterion in writing, so the company knows which duties it must staff internally and which it has contracted out.
Before you post the job, decide which task the first agent should handle. A good first task is frequent, bounded, observable and low risk if it goes wrong. If you are not sure which task that is in your company, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.
Frequently asked questions
How many agent operators does a company need? In most small and mid-sized companies, one person with clear authority is enough for a small set of agents, and the monitoring work can be shared with an external team. Add capacity when agents touch several departments or when approval queues start to slip. The signal to watch is the backlog of unreviewed changes, not the number of agents.
Can an existing staff member take on the role? Often yes, especially someone who already owns a process the agent touches, such as customer support or operations. The main cost is time, so the company should free real hours for log review and permission checks instead of adding the duty on top of a full workload. Without that protected time, the role tends to disappear within a few weeks.
Should the job post include a salary range? Compensation is a decision for your own policy and local market review, and this article does not provide figures. What the post should state clearly is the scope, the reporting line, the approval tiers and the metrics, because those define the level of responsibility the hire is accepting.
What is the first sign that the operator role is missing? The agents keep running, but nobody can say why an answer changed, who approved a new tool, or when a source document was last updated. When answering those questions takes days of searching, the operator duties are not covered, and the next incident will be harder to contain.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


