What is a standard operating procedure?

A standard operating procedure, or SOP, is the written, owned and version-controlled description of how a recurring task is carried out: who does what, in what order, with what checks and records. When an AI agent does part of the work, the SOP must also say what the agent does, at what depth, and exactly where a person takes over.

Why does an SOP matter when agents do the work?

Many SOPs were written for a process done entirely by people. Once an agent drafts, prepares or executes steps, the old SOP describes a process that no longer exists. Reviewers then work from habit, and auditors find a procedure that does not match the record.

An updated SOP is also the basis for training. You cannot test whether a reviewer follows the procedure for held payments if the procedure never mentions held payments. The SOP defines the standard; training brings people to it.

What does an SOP need when an agent is involved?

Beyond the usual scope, roles and steps, an SOP for a process with an agent should state:

1. The agent's part: which steps it performs, and its depth on each (read, draft, execute within bounds).
2. The handover points: where the agent stops and a named person acts.
3. What a held action requires: who may approve, what they must check, and what they must not approve.
4. Sources: which sources the agent relies on and which the person must verify.
5. Escalation: when to stop and who to call.
6. Records: what is filed automatically and what the person must add.

One distinction matters. The SOP tells people what to do. It does not control the agent. The agent's behaviour is set by its brief, sources and bounds, and enforced at the moment of action. If the SOP and the agent's bounds disagree, the bounds win at runtime and the SOP is wrong.

Templates help here: the human oversight procedure template and the two-person approval procedure template both slot into a wider SOP.

An example

A finance team rewrites its supplier payment SOP when an agent starts matching invoices and drafting payments. The new version says the agent matches at depth 3 within set tolerances, drafts payments at depth 2, and holds any payment above the team's limit or with changed bank details. It names the approver role, lists three checks before approval, and states that an emailed invoice is never on its own sufficient. The SOP links to the agent's record, which files every verdict, so an auditor can compare the procedure against what happened. See how an agentic operating system records decisions.

How is an SOP different from a policy?

A policy states what must be true and why: "payments above a set amount require two approvers". An SOP states how it is done: which screen, which checks, which record. Policies change rarely; SOPs change whenever the process does, including when an agent's capabilities or depth change.

An SOP also differs from a quick reference guide, which summarises it for use at the point of work and always defers to it.

Which terms are related?

• Gold standard: the agreed correct answer for a case, often used to test whether people follow the SOP; see what a gold standard is.
• Runbook: step-by-step instructions for a specific event.
• Management system standards: ISO/IEC 42001 expects documented processes; see what the standard requires of agent platforms.

For the full approach, read training teams for AI agents.