Standard Operating Procedures an AI Can Follow

Most small business SOPs, where they exist at all, were written for a new hire. That reader brings judgment, context and the ability to ask someone. Software brings none of those, which is why a procedure that works fine for a person tends to fall over the first time an agent runs it. The useful part is that fixing it makes the document better for the humans too.

Where human SOPs lean on judgment

Written for a person	What has to be stated
"Check the invoice looks right"	Which fields, matched against what, tolerance for rounding
"Escalate if unusual"	What counts as unusual, to whom, by when
"Use your judgment on discounts"	The floor, and who may go below it
"File it appropriately"	The folder, the naming convention, the retention period

None of those left-hand phrases is bad writing. They are efficient shorthand between people who share context. The moment the reader is software, or a new hire in week one, the shorthand becomes a blank the reader fills in with a guess.

The four parts

The trigger. What starts this. An email arriving, a form submitted, a date, a threshold crossed. Vague triggers produce procedures nobody knows when to run.

The steps. Numbered, in order, each one a single action with a single subject. Where a step involves a system, name the system and the screen.

The exceptions. The part most SOPs omit entirely. What happens when the information is missing, the amount is outside the usual range, the customer replies with a question, or the system is down. Three or four named exceptions with a stated response covers most real cases.

The completion condition. How you know it finished properly. A record updated, a confirmation sent, a number matching. Without this, both people and agents stop at plausible-looking points, and the same failure applies to prompts, which we covered in OpenAI's advice on trimming instructions.

Write it in an hour

1. Record yourself doing the task. Screen recording with narration, or just voice while you work. Say what you are doing and why, including the small decisions.

2. Hand the transcript to an AI assistant. Ask for numbered steps with decision points separated from actions. This produces a structured draft in a minute from material you would never have sat down to type.

3. Fill in the exceptions yourself. The assistant cannot know these, because they live in your head and in the last two years of odd cases. Ask yourself what went wrong most recently and write the response down.

4. Have someone else follow it without help. This is the test that matters. Every question they ask is a gap in the document, and the version produced after their questions is the real SOP. Resist explaining out loud; write the answer into the page instead.

What changes when an agent reads it

Three additions turn a good human SOP into one an agent can run.

Name the systems and the access. Which application, which account, which permissions. That is the scoping question from three boxes for an AI agent, and the SOP is where it belongs in writing.

Mark the stop points. Which steps the agent prepares and a person approves. Anything that moves money, sends externally, signs or deletes belongs on that list regardless of how reliable the tool has been.

State the boundary in the positive. What the procedure covers rather than a list of things not to do. A list of prohibitions is only as complete as the imagination of whoever wrote it, which is the lesson from the denylist that missed three functions and, more pointedly, from the attacker whose own agents ignored his country exclusion list in the PaperCut campaign.

Keep them short and keep them current

One page per procedure. Long SOPs go unread, and an unread procedure is a document rather than a control. If a task genuinely needs eight pages, it is probably three procedures wearing a trench coat.

Put a review date on each one and check it twice a year. A stale SOP describing software you replaced does more damage than no SOP, because people find one wrong document and stop trusting the library. When an agent follows the procedure, staleness stops being an annoyance and becomes an operational fault, since the agent will follow the outdated instructions confidently and at volume.

Which ones to write first

Not all of them. Pick by risk and repetition: the tasks that happen often, and the ones where a mistake is expensive. A procedure for the monthly reconciliation earns its hour. A procedure for booking the boardroom does not.

Start with whatever only one person knows how to do, because that document is worth writing whether or not automation ever touches it. If you have not yet worked out which processes those are, the exercise in process mapping before you automate anything comes first.

Frequently Asked Questions

What is a standard operating procedure?

A written description of how a specific recurring task gets done, detailed enough that someone other than the usual person can complete it correctly. In a small business an SOP is typically one page: what triggers it, the steps in order, what to do when something is unusual, and how you know it finished properly. Long SOPs go unread, which makes them worse than short ones that people actually follow.

How do I write an SOP for my business?

Record yourself doing the task once, narrating each step, then have an AI assistant turn the transcript into numbered steps. Edit that draft rather than starting from a blank page, because the recording captures the small decisions you would forget to mention. Then have someone who does not do the task follow the draft without help. Every question they ask is a gap, and the version after their questions is the real SOP.

How is writing an SOP for AI different?

A person fills gaps with judgment and asks when stuck. Software does neither reliably. So the parts a human SOP leaves implicit have to become explicit: what counts as an exception, what happens on each exception, what the completion condition is, and where the boundary sits between deciding and doing. Writing those down improves the document for people too, which is why this is worth doing whether or not you deploy an agent.

Should I use AI to write my SOPs?

Yes, for the first draft, and the fastest route is a recording rather than a blank document. Narrate the task while you do it, hand the transcript to an assistant, and ask for numbered steps with decision points separated out. What AI cannot do is know your exceptions, so treat the draft as a starting structure that a person who does the work has to correct.

How often should SOPs be updated?

When the process changes, which in practice means putting a review date on each one and checking it twice a year. The failure mode is a stale SOP that describes software you replaced, because people stop trusting the whole library after finding one wrong document. If an agent follows the procedure, staleness stops being an annoyance and becomes an operational fault, since the agent will follow the old instructions confidently.