Your best employee has a process. It’s just not written anywhere.
It lives in their head, in a few sticky notes, and in the way they’ve learned to handle the weird cases. When they’re out sick, things slow down. When they quit, things break. You’ve felt this.
Now try handing that same job to AI. It doesn’t have the sticky notes or the years of experience. It has whatever you give it. If you give it nothing, it fills the gaps with confident guesses.
Written process is the fuel for AI. That’s the reframe. SOPs for AI aren’t paperwork. They’re the instructions your agents run on.
Why AI needs written process more than people do
A new hire can ask you a question. They can watch how you handle something and copy it. They can sense when a customer is upset and slow down.
An AI agent can’t do any of that unless you tell it to. It reads your instructions literally. Vague instructions in, vague work out. That’s why two owners using the same AI tool get completely different results: one wrote down how the work gets done, and the other just said “handle my email.”
There’s a bonus most people miss. Once a process is written well enough for AI, it’s also written well enough for a new hire, a contractor or your future self. You fix two problems at once.
How to record a process once
You don’t need to sit down and type a manual. Here’s how I do it.
- Do the task for real — not in theory, the next time it actually comes up.
- Talk out loud while you do it. Say what you’re looking at, why you’re choosing what you choose, and what would make you do something different.
- Record it. A screen recording with voice, or a voice memo on your phone. Either works.
- Transcribe it. Most AI tools can turn audio into text.
- Have AI draft the SOP. Paste the transcript and ask for the structure below.
- Edit it yourself. Fix what’s wrong, add the edge cases you forgot to mention, cut what doesn’t matter.
- Test it. Let AI follow the SOP on real work while you review every output.
Here’s a prompt you can use for the drafting step:
You are an operations manager writing a standard operating procedure. Below is a transcript of me doing [TASK NAME] and explaining it out loud. Turn it into an SOP with these sections: purpose, trigger, inputs and where they live, step-by-step instructions, decision rules, definition of done, when to stop and escalate to [OWNER NAME], things to never do, and one example of good output. Use short numbered steps. Flag anything in the transcript that’s unclear or contradictory instead of guessing. Transcript: [PASTE TRANSCRIPT]
The SOP structure an AI agent can follow
Every SOP I write for AI follows the same skeleton. Consistency matters — agents perform better when every document is laid out the same way.
| Section | What it answers |
|---|---|
| Purpose | Why this job exists and what it protects |
| Trigger | What starts it |
| Inputs | What information it needs and exactly where that lives |
| Steps | Numbered, one action per step |
| Decision rules | If this, do that — the judgment calls written as rules |
| Definition of done | What a correct finished output looks like |
| Escalation | When to stop and hand off to a named human |
| Never do | The hard lines — sending without review, promising things, touching certain data |
| Example | One real sample of good output |
The two sections that matter most are decision rules and escalation. Decision rules capture the judgment you’ve built up, written as plain if-then statements. Escalation tells the agent where its job ends. An agent that knows when to stop is far more useful than one that guesses.
What changes when you write for AI
A few habits make an SOP work for an agent instead of just a person:
- Name the location of everything. Not “check the customer file” — say which folder, which system, which field.
- Define good. Show a finished example. AI copies examples better than it follows adjectives.
- Write the edge cases. The weird ones are where AI gets in trouble. “If the customer mentions a lawyer, stop and escalate.”
- One job per SOP. If a document covers three jobs, split it.
- Date it and give it an owner. Someone reviews it on a schedule and updates it when the work changes.
Build a second brain for your agents
SOPs work better when they live with everything else your AI needs: templates, past examples, key client details, pricing rules written as rules rather than numbers, your voice and tone guide. That organized collection is what people call a second brain — a knowledge base your AI tools and agents read for context.
Mine runs on a notes app and a shared drive, and it feeds my AI chief of staff. When the agent drafts a reply, books an appointment or updates a lead, it’s working from how we actually do things, not from generic internet advice. That’s the difference between an agent that sounds like you and one that sounds like everybody.
Keep sensitive client data out of tools that aren’t set up to handle it, and know which settings control where your data goes before you load anything in.
How this fits the bigger picture
SOPs are the Strategy piece of my BOSS framework — the written playbook for how the business wins. In an AI-first operations model, that playbook gets a second reader: the agent. Once the SOP exists, you can turn it into a workflow with a trigger, AI step, human check and action, and give it a seat on your human and agent ownership chart.
If you’re building your first agent from these documents, my guide on how to build an AI agent picks up where this page ends. In real estate, the same idea runs inside REI Boss, where the acquisitions and dispositions processes are written out so both people and AI can follow them.
Objections I hear
“My business is too custom for SOPs.”
Every business feels that way. The custom part is usually a handful of judgment calls. Write those as decision rules and the rest is routine.
“Things change too fast to document.”
That’s why each SOP gets an owner and a review date. A living document is still better than one that only exists in your head.
Your next step
Next time you do a recurring task, hit record and talk through it. Use the prompt above to draft the SOP, then edit it until a stranger could follow it.
If you want help turning your processes into a running AI chief of staff, see how Chief works.
Questions people ask
How do I write SOPs for AI?
Talk the process through out loud while you do it, record and transcribe it, then have AI draft the SOP from the transcript. Edit it into a fixed structure: purpose, trigger, inputs, steps, decision rules, definition of done, escalation and an example. Test it by letting AI follow it under review.
What's the difference between an SOP for people and an SOP for AI?
An AI SOP has to spell out what a person would fill in from experience: exactly where the information lives, what good output looks like, when to stop and ask, and what never to do. People infer. AI follows what’s written, and guesses where it isn’t.
What is a second brain for AI?
It’s an organized knowledge base — notes, SOPs, templates, past examples and key documents — that your AI tools and agents can read for context. It keeps AI working from your actual way of doing things instead of generic answers.