PrymeIntelligence

Agent governance·Playbook

How to set an AI agent's role scope

Set an AI agent's role scope by naming the work it does, for whom, through which channels, what is out of scope and where those requests go instead.

Describe the job before granting any access. An AI agent's role scope names the work it does, for which customers, products, channels and markets, what is explicitly out of scope, and where out-of-scope requests go instead. The process owner writes it, compliance checks the regulated edges, and the signed charter is encoded so every action is checked against it.

What is role scope, and who sets it?

Role scope is the outer edge of an agent's job. Depth says how far it may go within a capability; role scope says which capabilities and situations it has at all. A collections agent's scope might cover arrears on personal loans for retail customers in one market, through email and chat, and exclude business customers, disputes and bereavement cases. The idea is defined in role scope: what it means for AI agents.

The process owner sets the scope, because they own the work. Compliance reviews where the work touches regulated activity. The agent builder turns the scope into configuration. Role scope is not the same as role-based access control: access control decides what systems an identity can reach, while role scope decides what job the agent does. An agent needs both, and neither replaces the other.

What are the steps, and who acts at each one?

  1. Person, process owner: describes the job in plain words, as you would for a new hire: purpose, the work, who it serves.
  2. Person, process owner: lists in-scope capabilities as verbs on objects: read arrears history, draft reminder, agree plan within limits.
  3. Person, process owner: sets the boundaries: customer segments, products, channels, markets and languages the agent covers.
  4. Person, process owner: lists what is out of scope, explicitly, and names where each out-of-scope request goes: a team queue, a named role, a customer-facing message.
  5. Person, compliance: reviews regulated edges, such as where information becomes advice, or where a conversation could become a complaint or a sign of vulnerability.
  6. Agent, depth 1: is tested against the scope using past requests, including ones just outside each boundary, to confirm it refuses or hands over correctly.
  7. Person, process owner: signs the charter.
  8. Person, agent builder: encodes the scope, so every proposed action is checked against it before anything else happens.
  9. Person, process owner: reviews the scope at re-certification and whenever the process changes.
Step The agent does Depth Stays with a person
Describe the job Nothing none Writing the purpose
In-scope capabilities Nothing none Listing the work
Boundaries Nothing none Segments, products, channels, markets
Out of scope Nothing none Naming exclusions and hand-offs
Compliance review Nothing none Regulated edges
Test Handles requests at each boundary 1 Judging results
Sign Nothing none Approving the charter
Encode Nothing none Configuring the check
Review Nothing none Updating the scope

What evidence does each step leave?

  • Job description: a short purpose statement in the charter.
  • Capabilities: a numbered list, which later becomes the rows of the capability and depth matrix.
  • Boundaries: a table of segments, products, channels and markets in and out.
  • Out of scope: each exclusion with its hand-off destination.
  • Compliance review: notes on each regulated edge and the decision.
  • Test: cases at each boundary, the agent's response and the verdict.
  • Sign: the signed, versioned agent role charter.
  • Encode: a configuration export matching the charter.
  • Review: a dated review, even when nothing changed.

Which details are easy to miss?

Hand-offs, not just exclusions. "Disputes are out of scope" is half a rule. The agent needs to know where a dispute goes and what to tell the customer. Without that, out-of-scope requests stall.

Situations, not only topics. A request about arrears is in scope; the same request from someone who mentions a recent bereavement may not be. Name the situations that move a case out of scope.

Channels. An agent scoped for chat may behave differently by email, where messages are longer and carry attachments with their own instructions.

Markets and languages. Scope for one market does not carry to another, because rules differ. Name each market the agent may serve.

What else belongs in the description is covered in what should an AI agent's role description include?

What should you measure?

Count refusals at the role-scope bound by boundary, hand-offs by destination and whether the receiving team accepted them, requests that should have been handed off but were handled, and requests for new scope from the business. A steady flow of the last suggests the scope is too narrow or the process is changing.

Where does it usually go wrong?

The scope is the tool list. The agent's scope becomes whatever its connectors allow. Write the job first; connect only what the job needs.

The scope grows one exception at a time. Each small addition seems harmless. Together they produce an agent doing a different job from the one certified. Ways to resist this are in how do you stop an AI agent's scope from creeping?

Out-of-scope means silence. The agent refuses but does not route, and customers are left without an answer.

Scope is written but not enforced. If the charter lives in a document and nothing checks actions against it, it has no effect on the agent.

Where Pryme fits

In Pryme, role scope is the first of the five bounds every action meets, alongside the certificate, sources, your own limits and the platform guardrails. Agent Studio drafts the brief, including the role, from a plain-language description; the owner reviews and signs it. See agent governance is a bound, not a model.

Questions people ask

Can one agent have a wide role scope if its depth is low?

It can, but wide scope at low depth still exposes the agent to more situations, more data and more chances to mislead. Prefer several narrowly scoped agents over one broad one.

Who decides when a request sits on the boundary?

The scope should say where boundary cases go, usually to a person. Patterns in boundary cases go to the process owner, who may clarify the charter.

How is role scope different from your own limits?

Role scope defines the job. Your firm's own limits narrow what may happen within it, such as a refund ceiling or a product the agent may not discuss while it is under review. Both are checked on every action, and the tighter one wins.

Decide what your agents may do. Prove it. Enforce it.

Pryme Intelligence certifies each agent on your own cases, checks every action against five bounds at the moment it happens, and files every verdict.