A job description for an AI agent is the document that turns a vague automation intent into governed work. It defines mission, trigger, outcome, context, tools, permissions, evidence, metrics, and human review point. Without it, the company buys intelligence but does not build a role that can operate, learn, and be accountable.
The discussion gained current evidence. On 1September 2026, OpenAI published three cases of AI-native companies that brought agents for onboarding, account management, and product integration. The interesting point is not the brand of the model. It is the work pattern: teachable process, persistent context, tested action, and human at the decision point.
At XMACNA, we recognize this design because it starts with the role. Over +600 Digital Employees already operate in Brazil, and the practical lesson is consistent: model capability does not replace responsibility definition. A Digital Employee needs to know what starts the work, what outcome to deliver, where to seek context, what can be done, and when to call a person.
Why does the job description come before the model?
Companies often start the AI conversation by comparing models, prices, and demos. This order is comfortable but misplaces the main decision. A model can be excellent and still receive a poorly defined task, consult the wrong source, act with excessive access, or deliver something no one can verify.
Job descriptions follow the opposite path. First, the company chooses a process with value and consequence. Then defines what counts as completed. Only then decides which intelligence, memory, and tool are appropriate.
OpenAI's article shows this progression in three examples. Basis demonstrated an onboarding routine and turned it into reusable instruction, with trigger, known steps, correct tools, and a definition of “done”. Clay organized persistent context per account and kept evidence close to each recommendation. Exa took integration opportunities to tested artifacts, preserving human review before any submission.
These are vendor reports, not independent benchmarks. Still, they help make concrete what AI process automation requires: clear role, controlled context, and proof of execution.
What are the eight fields of an AI agent role?
A useful job description fits in eight fields. They can be written on one page, as long as they are specific.
1. Mission
The mission explains which operational problem the role solves. “Help sales” is too broad. “Qualify first contact, record needs, and forward the next action” describes observable work.
2. Trigger
The trigger defines when the role starts. It can be a new lead, a closed conversation, a received document, a completed step, or a detected exception. Without an explicit trigger, automation depends on improvisation or acts when it shouldn't.
3. Expected outcome
The outcome must be verifiable. It is not enough to “analyze well.” Better to define: updated record, created opportunity, summary with source, scheduled task, classified document, or case forwarded with full context.
4. Allowed context
Context is not full access. The role should consult only the data necessary to perform the work. PwC's recommendation on agent governance is clear: identity, role, specific permission per task, and auditable record must go together.
5. Tools and actions
List what the agent can consult, create, update, or only suggest. Reading a calendar is different from scheduling an appointment. Preparing a proposal differs from sending it. The job description needs to separate reading, writing, approval, and external communication.
6. Limits
Every role has boundaries. Discounts, commercial promises, financial decisions, sensitive data, conflicts, critical complaints, and regulated exceptions may require approval. A well-written limit is not a vague security phrase. It is a condition that changes the flow.
7. Evidence and metric
The agent must leave traces that allow decision review. Source used, action performed, prior state, final state, test, reason for escalation, and time form a trust base. The metric follows the outcome: cycle time, task accepted, rework, exception, quality, cost, or attributable revenue.
8. Human escalation
Define who receives the case, in what time, and with which context packet. Anthropic describes trustworthy agents as systems where a person can review before action and intervene during execution. Handoff is not automation failure. It is part of the role.
How to know if a process is ready to be delegated?
The process needs sufficient repetition to generate learning and manageable consequence to allow control. It also needs an owner capable of saying what is right.
A good candidate has recognizable input, repeatable steps, known sources, nameable exceptions, and an outcome that can be checked. Initial lead qualification, record updates, document triage, follow-up preparation, and request organization are common examples.
A bad candidate is “solve anything in the area.” A generic role creates generic permission, generic metric, and diffuse responsibility. IBM's research on agents in production reinforces the importance of observing real deployment, not just demo capability. Production is where exceptions, dependencies, review, and error cost appear.
What is the difference between autonomy and abandonment?
Autonomy is continuing to work within a contract. Abandonment is delivering a broad objective and hoping the system figures out limits the company never wrote.
An autonomous agent can research, compare, update, and test without asking for confirmation at every low-risk step. But it respects permissions, records evidence, and stops before a reserved decision. Humans don’t need to execute the whole routine; they need to keep authority over consequence, exception, and external commitment.
This separation is essential for AI agents in companies. The greater the consequence, the more explicit identity, authorization, monitoring, and review must be. Governance does not come after the flow has grown. It is born in the job description.
How to turn the job description into operation that improves?
The first version of the role will not be perfect. It must be testable.
Choose a small set of real cases. Record the human baseline. Run the flow with conservative limits. Review the result and mark where context was missing, where the tool did not respond, where permission was too broad, and where the human had to correct.
Then, update the function. Recurring exceptions become rules. Trusted sources become permitted context. Predictable failures become tests. Sensitive decisions become checkpoints. This is the foundation of the Intelligence Cycle: each execution produces evidence to improve the next, without confusing learning with unrestricted freedom.
A useful AI consultancy does not start by stacking tools. It starts by identifying the process, writing the job role, defining metrics, and designing collaboration between the carbon and silicon teams.
A practical template for the first Digital Employee
Before starting, answer on one page:
- What is the mission of the role?
- What event initiates the work?
- What artifact or change proves completion?
- Which sources can be consulted?
- Which actions can be performed without approval?
- What conditions interrupt or divert the flow?
- What evidence needs to be recorded?
- Who handles the exception and with what context?
If the team cannot respond, the priority is not yet to increase autonomy. It is to design the work.
In summary
- A job description transforms AI into an operational function.
- Eight fields connect mission, trigger, outcome, context, tool, limit, evidence, and responsible human.
- Current cases show that teachable processes, persistent context, and tests make execution reusable.
- A strong model does not compensate for poorly defined work.
- Governance starts before the first action, not after the first incident.
The next step is not to ask which agent to buy. It is to choose which function deserves to be clearly written. Take the AI Assessment and find where a Digital Employee can perform real work with limits, recording, and results.
Frequently asked questions about job descriptions for AI agents
What is a job description for an AI agent?
It is an operational document that defines mission, trigger, outcome, context, tools, permissions, evidence, metrics, and human escalation for a function executed with artificial intelligence.
How to choose the first AI agent function?
Choose a repetitive process with clear input, known sources, manageable consequences, responsible owner, and verifiable outcome. Avoid generic functions that try to help the entire company.
Can an AI agent act without human approval?
It can perform low-risk actions foreseen in the role. Sensitive decisions, external commitments, exceptions, and actions with significant impact must have checkpoints, stoppages, or approvals defined before deployment.
How to measure if the function is working?
Measure the process outcome: cycle time, task acceptance rate, rework, exceptions, required review, cost, and impact on business objectives. Response volume does not prove operational value.
What is XMACNA's role in this design?
XMACNA identifies the process, designs the function, defines limits, integrates context and tools, and operates the Digital Employee with evidence, supervision, and continuous improvement.