Most SOPs are a wish list written after the last incident. The real procedure lives in one person's head and three Slack threads. If you paste the wiki into a model and ask it to "automate this," you will automate the fiction.
This playbook turns a messy SOP into a small AI-assisted workflow: human still owns the decision; the model does the mechanical middle.
Step 0: Capture the real SOP, not the posted one
Pick one procedure that already happens at least weekly. Good first targets: intake triage, meeting-to-action notes, vendor email sorting, first-line customer replies, weekly report assembly.
Sit with the person who actually does it. Watch one real run or reconstruct the last one. Write what happened, including the ugly parts:
- What triggered the work
- Which tabs they opened
- What they copied
- Where they hesitated
- Who they asked
- What they sent, and to whom
- What they skipped because "we never do that part"
That list is the SOP. The wiki is optional commentary.
Step 1: Mark each step: mechanical, judgment, or forbidden
Go line by line. Tag every step with exactly one label:
- M — Mechanical: rewrite, extract, format, compare two texts, fill a template
- J — Judgment: approve, price, hire, fire, legal call, "is this the same customer," "do we make an exception"
- F — Forbidden for a model: anything that needs a live system of record you have not connected, secrets, or a decision you would not let a new hire make on day one
Rule: the model may draft M. A human must do J. F stays out of the prompt.
If more than half the steps are J or F, this is not an AI workflow yet. It is a checklist. Write the checklist. Stop.
Step 2: Write the AI job in one sentence
Bad job: "Help with the SOP."
Useful job: "From the ticket and the approved reply library, draft a first-line response and list any missing facts. Do not send."
The job must name:
- Inputs (what you will paste or attach)
- Output (the artifact)
- The human gate (who reviews, and what "done" means)
If you cannot name the gate, you are planning to auto-send. That is a different, later project.
Step 3: Build the working packet
Use the operator prompt system (Role, Constraints, Examples, Eval). The INPUT is only:
- The real SOP excerpt for this job
- The source material for this run (ticket, notes, email)
- The allowed facts (policy snippet, price list, template)
Do not paste the entire company handbook. Paste the two pages that apply.
Step 4: Insert human gates on a single page
Write the path as a straight line. No architecture diagram required.
- Trigger (email, form, Monday, a person drops a file)
- Human prepares INPUT (redact, attach the right policy snippet)
- Model drafts the M steps
- Human scores the eval and edits
- Human sends, files, or rejects
- If rejected: save the FAIL into the example list
That is the workflow. Tools come after the path is boring.
Worked example: "New vendor onboarding" that nobody follows
Posted SOP says: collect W-9, insurance cert, bank form, security questionnaire, legal review, then add to the AP system.
Real run from last Tuesday: someone forwarded a PDF, asked in Slack if we had used them before, paid a deposit on a card, and promised to "do the paperwork later."
Tagged steps:
- Search "have we paid this vendor" — J (identity). Human.
- Extract legal name, address, and requested payment method from the PDF — M. Model may draft.
- Decide whether insurance is required for this job type — J. Human, using a one-page rule.
- Draft the missing-docs email from a template — M. Model may draft.
- Add vendor to AP and pay — F until you have a real integration and an approver.
AI job: "From the vendor PDF and the missing-docs template, list fields we have, fields we lack, and draft the email asking for the gaps. Do not say we are approved. Do not mention payment."
Copy-paste: SOP stripper
You are helping me turn a procedure into an AI-assisted workflow. You are not allowed to invent steps we do not do.
>
From INPUT, produce:
1. Trigger (one line)
2. Steps as a numbered list. After each step tag M, J, or F and a five-word reason.
3. The single best AI job (one sentence: input, output, human gate).
4. Risks if we let the model do a J or F step.
5. A first prompt packet (role, hard constraints, eval) for the M steps only.
>
If INPUT is incomplete, list the questions you need and stop. Do not complete a fictional SOP.
>
INPUT:
[watched run notes + any wiki text, labeled as WIKI or REAL]
Five-case dry run
Before anyone else uses this:
- Two typical runs
- One missing-information run
- One hostile or exception run ("just pay them, I already promised")
- One case where the posted SOP and the real run disagree
Score each draft: would you send it, and did the model stay on M steps? If it starts making J calls, your constraints are too soft.
Failure modes
- Wiki worship: you automated the document, not the work.
- Silent send: no named human gate. The first error goes to a customer.
- Kitchen-sink input: the whole handbook in the prompt. The model picks a decorative paragraph and ignores the rule that matters.
- One-man workflow: only the author can run it because the packet lives in their chat history. Put the packet in a shared doc on day one.
- Tool first: buying n8n or an "agent" before the path is a straight line. A tool will not find the SOP you refused to write.
Decision rule: ship the assisted version
Ship when a second person can run the packet on a new case and pass eval without Slack-ing you. If they cannot, the workflow is still a personal habit.
Download
SOP stripper and gate sheet: /downloads/sop-to-workflow-template.md
You are done when
You have one real procedure tagged M/J/F, one AI job sentence with a named reviewer, and five scored runs.