Role scope is the definition of which tasks, customers, data and systems belong to an AI agent's job, written in its brief and enforced at the moment of action. It is the first of the five bounds every action meets. An action outside the role scope is refused, however capable the agent and however reasonable the request.
Why does role scope matter?
Agents built on general models can do far more than the job they were given. A billing agent can write poetry, give investment opinions and draft legal letters. Each of those is a risk when nobody tested the agent for it, nobody owns it, and no rules were set for it. Role scope is the line that says: this agent does this job, and nothing else.
It also keeps accountability clear. When every agent has a defined role, every action has an owner, and every gap between roles is visible. Without it, agents drift into each other's work and nobody can say whose output was whose. That is why role scope comes first in agent governance is a bound, not a model.
How is role scope written and enforced?
A role scope usually states:
- Purpose: the job in one or two sentences.
- Tasks in scope: the specific things the agent does, as capabilities.
- Tasks out of scope: the nearby things it must not do, named explicitly.
- Data: which records and fields it may see.
- Systems: which tools and connections it may use.
- Customers or segments: whom it serves, if not everyone.
- Hand-offs: where out-of-scope requests go.
The agent role charter gives a layout, and the process of setting one is in how to set an AI agent's role scope. Enforcement happens at runtime: every proposed action is checked against the scope, and anything outside it is refused with role scope as the deciding bound.
What does that look like in an example?
A university's admissions agent has a role scope covering application status queries, document checklists and interview scheduling. Fee payment plans and visa advice are named as out of scope, with hand-offs to the finance office and the international team. An applicant asks whether they can pay their deposit in instalments. The agent recognises the request, refuses to answer the substance, and passes the query to the finance office with the applicant's details. Its role stays narrow, and the applicant still gets an answer from the right place.
How is role scope different from depth?
Role scope says what the agent may work on. Depth says how far it may act within that work: read, draft, execute, move money or coordinate. An agent can have a capability inside its scope at depth 1, meaning it may only look things up. Widening scope adds new kinds of work; raising depth lets the agent do more with work it already has.
Role scope is also different from the scope gate. The scope gate is a certification test that checks whether the agent stays within its role when tempted to step outside it. Role scope is the line; the scope gate tests whether the agent respects it before the runtime check has to.
How does role scope creep?
Usually through small, helpful additions: a new question type added to the brief, a new data source connected for convenience, a hand-off quietly removed because the agent seemed to cope. Each change widens the role without recertification. Controls against this are covered in how do you stop an AI agent's scope from creeping? and how do you stop an AI agent from going beyond its role?
Questions people ask
Who should define an agent's role scope?
The person who owns the work the agent does, such as the head of the team it supports, with risk and compliance reviewing the out-of-scope list. The team building the agent can draft it, but the owner signs it, because the owner answers for what the agent does.
Does changing the role scope require recertification?
Adding a capability does, because the agent has no evidence for it yet. Narrowing scope does not, though it should still be recorded. A changed scope is effectively a new role, and the certificate covers the role that was tested.