Most small businesses have very few written procedures. The way things get done lives in a few experienced people's heads. That works until someone's sick, someone quits, or you try to hand the job to AI.
Here's a useful discovery: a procedure written well enough for AI to follow is also the best training document you'll ever give a new employee. Same document, two uses. This post shows the format I use.
Why most SOPs don't work
The typical procedure document has one of two problems. Either it's a vague paragraph ("process invoices promptly and accurately") or it's a 40-page binder nobody reads. Neither tells someone what to actually do on a Tuesday morning with a stack of work in front of them.
AI exposes this fast. Give a model a vague instruction and you get a vague result. That's not the model failing. It's the procedure.
The five parts of a usable procedure
Every procedure I write, whether for a person or an agent, has the same five parts.
1. Purpose and trigger
One or two sentences: what this job produces and when it starts. "Each weekday at 10am, turn yesterday's service calls into invoices ready for approval."
2. Inputs
Exactly what you start with, and where it lives. "The job sheets in the Completed folder. The price list (Prices 2026 tab). The customer list in the accounting system."
3. Steps
Numbered, one action each, in order. Write them for someone who's smart but has never done this job. If a step involves a decision, spell out the rule: "If the job took over 4 hours, use the half-day rate."
4. Checks
How to know each output is right before it goes anywhere. "Total labour hours match the job sheet. Every part on the sheet appears as a line item. GST is 5% of the subtotal."
5. Output and handoff
What the finished thing looks like and where it goes. "A draft invoice in the accounting system, marked Ready for review, with a note listing anything unusual."
That's it. Most procedures fit on one page.
A worked example
Here's a shortened version of a procedure for a common job: answering a rental inquiry email.
Purpose: Reply to rental inquiry emails within the hour, with a clear answer and a booking link.
Inputs: The customer's email. Live availability from the booking system. The current price list. The policies page.
Steps: 1. Identify the dates, number of people and type of equipment requested. 2. If any of those are missing, ask for them in one short question. 3. Check availability for the exact dates. 4. If available, quote the price from the price list and include the booking link for those dates. 5. If not available, offer the nearest available dates or a similar option. 6. Answer any policy question using the policies page wording only.
Checks: The dates in the reply match the customer's request. The price matches the price list for that season. No promise is made that isn't in the policies.
Output: A reply under 150 words in a friendly, plain tone, signed by the shop. Anything involving a complaint, a refund or a group over 10 goes to the manager instead, with a one-line summary.
A new staff member could follow that on their first day. So can an AI agent, connected to the booking system and the price list.
Tips that make a big difference
Write rules, not hopes. "Be careful with large orders" is a hope. "Orders over $2,000 need manager approval" is a rule.
Name the exceptions. List the situations where the person or AI should stop and ask. This is where most real-world mistakes happen.
Include one good example. A sample of a finished output you were happy with does more than a page of description.
Use your real words. If your team says "work order", don't write "service ticket". Consistency matters more to AI than to people.
Keep it current. When the process changes, update the document the same day. An out-of-date procedure is worse than none, because it gets followed.
Start with the job that hurts most
You don't need to document the whole business. Pick the one job that takes the most hours, causes the most mistakes or depends on one person. Write it up using the five parts. Have someone who doesn't normally do the job follow it, and fix every place they got stuck.
Once it works for a person, it's ready for an agent. And even if you never use AI for it, you've made the business less fragile.