Short answer
An SOP for AI agents, or AI SOP, is a standard operating procedure written for an AI agent to carry out rather than a person to read. It covers one type of request: when it applies, what to check, the steps and how it ends.- SOP stands for
- Standard operating procedure
- Also called
- AI SOP, agent SOP, agent operating procedure, agentic SOP
- Scope
- One type of request, including what it is not for
- Before live
- Checked against real past conversations
SOP means the same thing in AI as anywhere else. What changes is the reader. A person following an SOP fills gaps with judgement and asks a colleague when something is unclear. An AI agent fills gaps by guessing, so an SOP written for one has to be narrower and more exact.
Two meanings of AI SOP
The phrase is used for two different things.
- An SOP written for an AI agent to follow. This is the meaning on this page: the procedure decides what the agent does when a customer asks for something.
- AI used for SOPs written for people. Tools that draft procedure documents from notes or recordings, and SOP chatbots or SOP assistants that answer staff questions about how a process works.
Both are useful. Only the first decides what a customer experiences.
What goes in an SOP for an AI agent
An agent SOP has five parts, and each one has to be explicit.
- When to use. The situation in plain words, and what it is not for.
- Conditions. Facts that must be true first: the order exists, it has not shipped, the message came in by email.
- Steps. One question or action at a time, in order.
- Limits. What the agent may not do or promise, and when it must hand off.
- Ending. Reply and stay open, hand off with a summary, or close.
The part most often missing is the "not for" line. Without it, a procedure written for damaged items also starts on late deliveries, because to a language model the two can sound alike.
In Aide, the "not for" line can be backed with examples: a procedure takes "Use when" and "Don't use when" messages from real customers. Each customer message in the thread then names the procedure Aide followed and the ones it considered but didn't, with the reason.
An example for a Shopify store
A damaged-item SOP, written for an agent:
When to use: a customer says an item arrived damaged or broken. Not for parcels that are only late, or items that are the wrong size.
Conditions: the order exists and has been delivered.
Steps: ask for a photo of the damage if none was sent; confirm which item on the order; explain the store's damaged-item policy; tag the conversation "damaged"; hand off to the team with the photo and order attached.
Ending: hand off.
Notice what it leaves out: a refund decision. That stays with a person until the team decides otherwise.
In Aide, this SOP is one procedure: a Collect step for which item arrived damaged, a step that adds the "damaged" tag, and a Hand off finish. Aide works through the steps one message at a time, skips any the customer has already answered, then hands off, so the team receives the order, the item and the photo up front instead of a bare "I've escalated this". Refunds on Shopify orders wait for a teammate by default, so the refund decision stays with a person even when Aide handles the rest.
How to test an AI SOP before it goes live
Writing the procedure is half the job. Checking it is the other half.
- Run its trigger against real past conversations. Which would it have picked up, and which would it have left alone?
- Read the near misses. A conversation it wrongly picked up usually means the "not for" line needs another sentence.
- Try it in a test conversation and read every reply.
- Set it live with the old version kept, so a bad edit can be undone.
- Read the conversations it handled in its first week.
The approach this page rejects is one long prompt that tries to cover every situation at once. Nobody can test that per situation, and nobody can tell which line caused a bad reply.
Where Aide fits
In Aide, these are Procedures, under Automations. Each one has a When to use in plain English, a choice of where it runs (whole channels or specific inboxes), optional conditions on the topic or the Shopify order, steps written as a short document, and a finish. Owners and admins check a procedure's When to use against recent real conversations and run the draft in a test chat before setting it live. Edits to a live procedure stay a draft until someone sets them live, and every version is kept.
A team with SOPs already written doesn't start over. Describe your process takes a pasted SOP, a macro or a sample chat and drafts the procedure using the workspace's own fields, macros, knowledge sources and teams, lists anything missing, and asks up to three questions where the SOP leaves something open. The draft stays switched off until someone sets it live.