What is Scope Of Work?
Also called SOW, agent scope.
A scope of work is a written statement of what a worker or system is responsible for, what falls outside that responsibility, and what standard the output must meet. Applied to an AI agent, it bounds which tasks it may perform, which systems it may touch, and which decisions it must refer to a person.
The exclusions carry more weight than the inclusions. Inclusions can be inferred from observed behavior, but nothing signals a boundary unless it is written. A scope stating that an agent drafts customer replies but never issues refunds, never modifies a contract, and never contacts a customer flagged as in dispute is doing its real work in the second half.
A workable scope names four things: the tasks in bounds, the systems and data reachable, the decisions reserved for a person, and the standard the output is measured against. Vague standards defeat the purpose, so specifics such as every claim carries a source link, or no reply is sent without an order reference, are what make the document enforceable.
Scope is the natural place to record autonomy level per task rather than globally. The same agent might act unsupervised on internal research, produce drafts requiring approval for customer communication, and be excluded entirely from anything touching payments. Setting one autonomy level across all of an agent's work is a common and avoidable source of trouble.
The document is only useful if it is enforced somewhere. A scope existing solely as a description, with no corresponding permission configuration or approval requirement, describes intent rather than behavior. Practitioners generally treat the written scope and the technical permissions as two views of the same decision, and check periodically that they still match.
Key points
- States tasks in bounds, out of bounds, and the output standard
- Exclusions matter more than inclusions
- Names reserved decisions and reachable systems
- Autonomy level should be per task, not global
- Must be mirrored in actual permissions to mean anything
In practice
A scope for an outreach agent reads: researches prospect companies and drafts first contact emails; may read the CRM and the public web; may not send anything, may not add or edit CRM records, may not contact anyone marked as an existing customer or a former customer. Standard: every personalization claim cites a source URL. The CRM permissions are set to read only to match.