A standard operating procedure, or SOP, is a written description of how a task gets done in your business: what comes in, what decisions get made, what goes out, and who checks it. Most businesses have a few. Most are out of date, and most live in someone's head rather than on a page.
That matters now because an SOP is exactly what a large language model needs to do a task well. An LLM is very good at following clear instructions and very bad at guessing what you meant. When you hand it a vague request, you get a vague answer. When you hand it the same instructions you would give a new staff member, you get work you can use.
This guide covers what an SOP gives a model and an automation that a prompt alone does not, what a good LLM-ready SOP looks like, and how to turn one into a working automation without losing control of the result. By the end you should be able to pick one task in your business and write it up in a form that both a person and a model can follow.
What an SOP gives a model that a prompt does not
A prompt is a request. An SOP is the context behind the request. Ask a model to "reply to this enquiry" and it will write something polite and generic. Give it your enquiry SOP and it knows which services you offer, what you need to ask before quoting, which enquiries go to the owner, and what a reply from your business sounds like.
The same is true for automation that does not involve AI at all. A workflow tool cannot route an enquiry by service if nobody has written down what the services are and how to tell them apart. The SOP is the specification. The model or the workflow is just the thing that runs it.
- Inputs: what the task starts with, and where that information lives.
- Decisions: the rules a person applies, including the exceptions.
- Outputs: what a finished result looks like, with a real example.
- Checks: who reviews it, and what would make them send it back.
Why most AI pilots stall
The usual pattern: someone tries an AI tool on a real task, the first result is impressive, the tenth is wrong in a way that costs time to fix, and the tool gets quietly dropped. The problem was rarely the model. It was that every attempt started from a blank prompt and relied on whoever was typing to remember the rules.
An SOP removes that variable. The rules live in one place, the model reads the same rules every time, and when a result is wrong you fix the document rather than the prompt. Consistency is what makes AI output trustworthy enough to act on, and consistency comes from the procedure, not the model.
What an LLM-ready SOP looks like
You do not need a template or a tool. You need plain language, concrete examples and a clear line between what the model decides and what a person decides. A short SOP that is accurate beats a long one that is aspirational.
Keep in the SOP
- The goal of the task in one sentence.
- Where each input comes from: a form, an inbox, a spreadsheet, your website's database.
- The decision rules, written as "if this, then that", with the edge cases you have actually hit.
- Two or three real, finished examples of good output.
- What must never happen: prices quoted without approval, promises about delivery dates, personal data sent outside the business.
Leave out
- History and justification. The model does not need to know why the rule exists, only what it is.
- Steps that depend on judgement you cannot describe. Mark those as a human decision instead.
- Anything that changes weekly. Keep that as data the SOP points to, not text inside it.
Turning an SOP into automation
Once the SOP exists, automating it is a matter of connecting each part to the system it lives in. Inputs come from your website, your inbox or your CRM. The decision rules become either workflow logic, for anything that is a clean rule, or a model call with the SOP as its instructions, for anything that needs reading and writing. Outputs go back into the same systems. Checks become a review step before anything leaves the business.
This is why owning your website's data matters. If enquiries, services and prices are structured records rather than text inside page layouts, the SOP can point at them directly and the automation reads live information. If they are not, the first job is getting them there, which we cover in our six step guide to making a website agentic ready.
- Rule-based steps become workflow conditions. No AI needed.
- Reading and writing steps become a model call with the SOP as its standing instructions.
- Judgement steps stay with a person, with the model preparing a draft and the evidence.
Keep a person in the loop, on purpose
The goal is not to remove people from the process. It is to remove the retyping, the chasing and the first draft, so the person's time goes on the decision. A good automation drafts the reply, attaches the enquiry and the customer's history, and waits for a click. The SOP says who clicks and what they are checking for.
Start with approval on everything. As the results prove reliable, move the routine cases to automatic and keep approval for the exceptions the SOP already names. The document tells you where that line is, because you wrote it down.
Where to start
Pick the task that is repetitive, well understood and slightly annoying. Enquiry triage, quote preparation, appointment confirmations and follow-ups after a job are the usual candidates. Write the SOP first, run it by hand for a week using the model as an assistant, then automate the parts that held up.
One task, one page, one connection. When that works, the second one is faster, because you now have a way of describing work that both your staff and your systems can follow.
Key takeaways
- 01
An SOP is the context a model needs. A prompt is just the request.
- 02
Inconsistent AI results usually mean the rules were never written down.
- 03
Keep SOPs short, concrete and split between model decisions and human decisions.
- 04
Automate rule-based steps with logic, reading and writing steps with a model, and keep judgement with a person.
Frequently asked questions
Do I need an SOP before I can use AI in my business?
No, but you need one before you can rely on the result. Without a written procedure, every use of the model starts from scratch and depends on whoever is typing. The SOP is what makes the output repeatable.
How long should an LLM-ready SOP be?
One page is the target. The goal, the inputs, the decision rules with edge cases, two or three real examples, and the hard limits. Anything longer usually means history or judgement calls have crept in.
Can an SOP be automated without any AI?
Often, yes. If every decision in the procedure is a clean rule, a workflow tool can run it end to end. AI earns its place when a step involves reading, summarising or writing.
What if the SOP changes?
Change the document and the automation follows, because the automation reads the SOP rather than having the rules hard coded. Keep a version number on it so you can tell which version produced any given result.