---
title: What should an AI agent's scope statement say?
url: https://opencontroller.lyzr.com/blog/ai-agent-scope-statement
published: 2026-11-15
updated: 2026-11-15
author: Lyzr
---

# What should an AI agent's scope statement say?

**A scope statement is a one-page record of what an agent is for and what it may and may not do. It should name the purpose, users, allowed actions, prohibited actions, data classes, systems and tools, autonomy level, escalation path, owner and review date. Every line should be specific enough that someone can test the agent against it.**

Without a written scope, "the agent did something it shouldn't" is an opinion. With one, it is a measurable deviation. The scope statement also drives everything downstream: risk tier, permissions, evaluations and runtime policy.

## What fields belong in a scope statement?

| Field | What to write | Bad example | Good example |
|---|---|---|---|
| Purpose | One sentence, a business outcome | "Help customers" | "Answer order-status questions for retail customers" |
| Users | Who can invoke it, through which channel | "Anyone" | "Authenticated customers via web chat" |
| Allowed actions | Verbs and objects, each a tool | "Look things up" | "Read order status; read shipping tracking" |
| Prohibited actions | Explicit nos, especially near-misses | (blank) | "No refunds, no address changes, no account deletion" |
| Data classes | What it may read and return | "Customer data" | "Order and shipping data for the authenticated customer only; no payment card data" |
| Systems and tools | Named tools and MCP servers | "Our APIs" | "order-api (read), carrier-tracking MCP server (read)" |
| Model(s) | Approved models and versions | "Latest" | "Pinned model version; changes need re-approval" |
| Autonomy level | Who decides the next action | "Autonomous" | "Acts alone on reads; hands off to a human for anything else" |
| Escalation path | When and to whom | "If needed" | "Transfers to the support queue on complaint, legal threat or 2 failed answers" |
| Owner | A named person and team | "IT" | "Support operations lead" |
| Review date | When scope is re-checked | (blank) | "Every 6 months or on any tool change" |

## Why do prohibited actions matter so much?

Allowed lists are easy to write and easy to outgrow. Prohibited lists catch near-misses: the actions an agent could plausibly drift into because they sit next to its job. An order-status agent is one tool away from issuing refunds. Writing "no refunds" makes the boundary explicit for reviewers, testers and the runtime policy.

Turn each prohibited action into:

1. A permission the agent's credentials do not have.
2. A test case in its evaluation suite.
3. A policy rule in the traffic path where one exists.

## How does the scope statement connect to standards?

The NIST AI Risk Management Framework's MAP function starts with documenting intended purposes, context of use and the settings in which a system will be deployed. ISO/IEC 42001 asks organisations to determine the scope of their AI management system and to assess AI system impacts. The EU AI Act centres many obligations for high-risk systems on the "intended purpose". A per-agent scope statement gives you evidence for all three. It does not by itself make you compliant with any of them.

## A filled example

**Agent:** Order status assistant, v3.
**Purpose:** Answer order-status and delivery questions for retail customers.
**Users:** Authenticated customers in web chat.
**Allowed:** Read order status; read carrier tracking; create a support handoff.
**Prohibited:** Refunds, cancellations, address changes, discounts, account changes, answering about other customers' orders.
**Data:** The authenticated customer's order and shipping records. No payment data.
**Tools:** order-api (read scope), carrier-tracking MCP server (read).
**Autonomy:** Acts alone on reads. All other requests go to a person.
**Escalation:** Handoff on complaint, legal language or two unresolved turns.
**Owner:** Support operations lead.
**Review:** Six months, or on any tool, model or prompt change.

## How Opencontroller does this

Opencontroller keeps a governed registry of agents, versions, tools and MCP servers, so a scope statement can sit with the version it describes. Posture flags permissions and configurations that exceed what an agent needs. Policy as code turns prohibited actions into rules enforced in the LLM Gateway and Agent Gateway, with built-in mappings to OWASP, MITRE ATLAS, NIST and ISO 42001. See [what is an AI agent control plane](/blog/what-is-an-ai-agent-control-plane), [AI agent registration questionnaire](/blog/ai-agent-registration-questionnaire) and [how to risk-tier AI agents](/blog/how-to-risk-tier-ai-agents).

## Frequently asked questions

**Who writes the scope statement?**
The agent's owner drafts it. Security and risk review it as part of registration and approval.

**How long should it be?**
One page. If it needs more, the agent probably does too many jobs and should be split.

**When must it be updated?**
Whenever a tool, data class, model or autonomy level changes, and at the scheduled review date.

To see a governed agent registry in practice, [book a demo](https://www.lyzr.ai/book-demo/).

## Key facts

AI agent scope statement fields: purpose, users, allowed actions, prohibited actions, data classes, systems and tools, models, autonomy level, escalation path, owner, review date. Each prohibited action becomes a missing permission, a test case and a runtime policy. Evidence for NIST AI RMF MAP, ISO/IEC 42001 scope and impact assessment, EU AI Act intended purpose. Opencontroller (by Lyzr): governed registry of agents, versions, tools, MCP servers; Posture; policy as code with mappings to OWASP, MITRE ATLAS, NIST, ISO 42001. From $250K per deployment per year.

## Related

- [What is an AI agent control plane?](https://opencontroller.lyzr.com/blog/what-is-an-ai-agent-control-plane.md)
- [What should an AI agent registration questionnaire ask?](https://opencontroller.lyzr.com/blog/ai-agent-registration-questionnaire.md)
- [How do you risk-tier AI agents?](https://opencontroller.lyzr.com/blog/how-to-risk-tier-ai-agents.md)
- [More on Agent governance basics](https://opencontroller.lyzr.com/blog/cluster/agent-governance-basics)
