Standard operating procedure (SOP) generator: write it, then find out if an agent can run it

What a standard operating procedure is, and what it is not

A standard operating procedure is the document that lets a competent person who has never done a task produce the same result as the person who normally does it. That is the whole test. Not length, not formatting, not whether it lives in the wiki. Hand it to someone new and watch where they stop.

The term comes from regulated physical work. OSHA's process safety management standard requires employers to “develop and implement written operating procedures that provide clear instructions for safely conducting activities involved in each covered process”, and the EPA publishes a whole guidance document on how to write one. Look at that guidance and you will see it is built for laboratories and field work — instrument calibration, sample handling, health and safety warnings. Useful, and not the shape of a procedure for answering an email or raising an invoice. The eight sections below are the office version.

What an SOP is not: a policy, which says what is allowed; a process map, which shows how work moves between teams; or a work instruction, which is the keystrokes for one step. Those three have their uses. The SOP is the level you can automate from.

The standard operating procedure template we use, free

Here it is, in plain text, no email and no download. Select it, copy it, paste it into whatever your team already uses. The generator above fills the same skeleton in for you and adds the readiness verdict, but if you would rather just take the template and go, take it.

STANDARD OPERATING PROCEDURE

Title:            [TASK NAME]
Version:          1.0
Owner:            [NAME, ROLE]
Backup:           [NAME, ROLE]
Last reviewed:    [DATE]
Next review:      [DATE]

1. OUTCOME — WHAT DONE LOOKS LIKE
[One sentence a third person could mark yes or no against.]

2. TRIGGER — WHEN THIS RUNS
[The event or the time that starts it. Be specific enough that software could detect it.]

3. OWNER AND BACKUP
Owner:  [NAME] — accountable for this procedure and for reviewing it.
Backup: [NAME] — covers holidays, illness and overflow.

4. SYSTEMS THIS TOUCHES
- [System 1 — what it is used for here]
- [System 2 — what it is used for here]

5. THE STEPS
1. [Action] — [who does it] — [in which system]
2. [Action] — [who does it] — [in which system]
3. [Action] — [who does it] — [in which system]

6. DECISIONS, EXCEPTIONS AND ESCALATION
- If [condition], then [what to do instead].
- If [condition], then [what to do instead].
- Anything that matches none of the above: stop and tell [NAME] the same day.

7. THE KPI THIS SERVES
KPI:          [The number this task moves.]
Measured by:  [Where the number is read from, and how often.]
Target:       [The number you are aiming at.]

8. REVIEW
Reviewed every [CADENCE] by [NAME].
Version note: [What changed this time, and why.]

One warning about templates in general. A blank form makes it feel like the work is filling in boxes, and it is not. The work is watching the task happen and writing down what actually occurred, including the bit where somebody checks a spreadsheet nobody has mentioned in two years. The template only saves you the argument about structure.

The eight sections every usable standard operating procedure needs

Each section earns its place by being the thing that breaks when it is missing. We arrived at these eight by writing procedures for clients and then watching which ones a new hire, or an agent, could actually follow.

A complete worked example: answering a new inbound lead inside one business hour

This is an illustration, not a client's document. We picked a task almost every company runs, wrote it the way we would write it, and left the company-specific things in brackets. Read it for the shape — especially step 3, which is the judgment call that stops this procedure from being fully automatable, and section 6, which is the part most people leave in one person's head.

STANDARD OPERATING PROCEDURE

Title:            Responding to a new inbound lead within one business hour
Version:          1.0
Owner:            [NAME], Head of Sales
Backup:           [NAME], Account Executive
Last reviewed:    [DATE]
Next review:      [DATE + 3 months]

1. OUTCOME — WHAT DONE LOOKS LIKE
The lead has a written reply from a named person, the enquiry is recorded in the CRM
against a contact and a company, and either a meeting is booked or the lead is marked
Not a fit with a one-line reason.

2. TRIGGER — WHEN THIS RUNS
A new submission on the website contact form, or a first-time email into the shared
sales inbox. Business hours only; anything arriving outside them starts the clock at
the next opening time.

3. OWNER AND BACKUP
Owner:  Head of Sales — accountable for the response-time number and for this document.
Backup: The named account executive on the weekly rota.

4. SYSTEMS THIS TOUCHES
- Website form (submissions land in the shared inbox and in the CRM)
- Shared sales inbox
- CRM (contacts, companies, deals, deal stages)
- Calendar and the booking link
- The team chat channel used for handoffs

5. THE STEPS
1. Read the enquiry and pull out company name, role, what they asked for, and any
   deadline they mentioned — account executive — shared inbox.
2. Search the CRM for the company and the email domain. If a record exists, attach the
   enquiry to it. If not, create a contact and a company — account executive — CRM.
3. Check the enquiry against the three fit criteria written at the bottom of this
   document — account executive — judgment call, no system.
4. If it is a fit, reply with a short answer to the actual question asked, plus the
   booking link. Do not send a template that ignores what they wrote — account
   executive — shared inbox.
5. If it is not a fit, reply saying so in two sentences and point them somewhere useful
   — account executive — shared inbox.
6. Set the deal stage: New enquiry, Meeting booked, or Not a fit with the reason in the
   note field — account executive — CRM.
7. Post the enquiry and what you did in the sales channel, so the next person sees it
   without asking — account executive — team chat.

6. DECISIONS, EXCEPTIONS AND ESCALATION
- If the enquiry is from an existing customer, it is a support matter, not a lead. Send
  it to the support inbox and stop.
- If the enquiry names a deadline inside 48 hours, tell the owner directly rather than
  waiting for the channel post to be read.
- If the enquiry is clearly automated or a sales pitch at us, mark it Not a fit and do
  not reply.
- Anything that matches none of the above: stop and ask the owner the same day.

7. THE KPI THIS SERVES
KPI:          Percentage of new inbound leads that receive a first human reply within
              one business hour.
Measured by:  Time between the CRM created-at timestamp and the first outbound email on
              the record, reviewed weekly.
Target:       Set your own. Pick a number you are currently missing, not one you are
              already hitting.

8. REVIEW
Reviewed every quarter by the owner.
Version note: First version. Fit criteria at the bottom were agreed with the founder and
should be re-read whenever the offer changes.

Run that example through the rubric and you get: five of seven steps mechanical and in a named system, one judgment call at step 3, a machine-detectable trigger, and an escalation rule. The verdict is mostly agent-ready. An agent could read the enquiry, find or create the CRM record, draft the reply, set the stage and post to the channel. A person keeps the fit decision and presses send. That split is the normal answer, and it is worth far more than waiting until the whole thing is automatable.

What makes a standard operating procedure ready for an AI agent

Here is the test, in the words we use on calls: if a role has clearly determined KPIs and clearly determined SOPs, and the work is mostly digital, an agent can very likely do that work much better — or at least take a real part of it off the person. Three conditions. Miss any one and you are guessing.

The most upvoted version of this belief in the places where operators actually talk to each other is blunter: there is no automation without documentation. We agree with it, with one correction. Writing the procedure is not a tax you pay before the interesting part starts. It is most of the work, and the build afterwards is comparatively cheap.

Now the honest counter, because that belief has a failure mode of its own. Documenting a bad process does not make it good — it makes it repeatable, which means automating it means being wrong faster and in more places. Before you hand a procedure over, read it and ask whether the task should exist at all. We have deleted more steps for clients than we have automated.

The rubric, in full

This is exactly what the generator applies. No model, no black box — four gates, four step rules, four bands.

Gates. Any one of these missing and the answer is not yet, whatever the steps look like.

• No KPI written down → not ready. Write the KPI first. This is the one we get argued with about, and we hold it, because a task that moves no number has no defensible reason to be automated before the tasks that do.
• A KPI with no measurement method → not ready. "Faster response times" is a hope. "Time between the created-at timestamp and the first outbound email, read weekly from the CRM" is a KPI.
• No definition of done → not ready. The finish line is the stop condition for the agent.
• No steps → not ready. There is nothing to assess.

How each step is classified.

• A step with a named system and no judgment call is counted as runnable. This is the raw material of automation.
• A step marked as a judgment call is flagged individually and stays with a person, or becomes draft-then-approve.
• A step that happens off the screen — a phone call, a visit, a signature on paper — is flagged as a boundary. The automation stops there.
• A step with no system named and no judgment flag is counted as vague, not as runnable. It is a sentence somebody still has to interpret, and interpretation is exactly what you are trying to remove.

The bands.

• Every step runnable, a machine-detectable trigger, and an escalation rule → agent-ready. Build this one.
• 70% or more of steps runnable → mostly agent-ready. Hand over the mechanical part, leave the judgment with a person.
• 35% to 70% → a co-pilot. Drafting and prep, not end-to-end running.
• Under 35% → keep it human. Write the procedure anyway for onboarding, and automate a different task first.

Flagged separately, because each one changes what the worst day looks like.

• A trigger that depends on someone noticing. An agent cannot notice.
• A missing escalation rule. This line decides how bad the worst day gets.
• No named owner, or no backup.
• No review cadence on a procedure that touches a tool you do not control.

The one rule we will not bend: no written KPI, no automation. Not because we like paperwork. Because the KPI is how you find out in week two that the agent is quietly failing, instead of finding out in month three from a customer.

What we find when companies open their own documentation

No more than about a quarter of the companies we assess arrive with documentation an agent could run from. That is our own number, from our own assessments, and it is lower than anyone selling you an AI platform will tell you.

Three things stall deployments, and none of them is the model. First, the answers are not in the wiki — they are in Slack threads, in two senior people's heads, and in three documents that contradict each other. We have had to interview the senior people and write it down before there was anything to index at all. Second, a flat rollout. One client in the 20 to 50 person range gave every employee a personal agent with its own token budget; spend hit roughly $3,000 to $5,000 a month and they shut the whole programme down inside about two months. Idle agents were burning tokens on heartbeats and cron loops with nobody asking them anything, and most people never used the one they were given. Nobody had written down what any of those agents was for. That is the same missing document as an SOP, at company scale. What we would do now: shared departmental agents before personal ones, a token budget per role, idle loops killed, and cheap tasks routed to cheap models.

Third, generic beats nothing but loses to specific. Nobody touched the general-purpose assistant at one deployment until we built skills per department. Adoption followed specificity — and specificity is what a written procedure is.

For a company-wide rollout, expect roughly 30% of staff to be active users at three months. That is a normal, healthy number, not a shortfall.

Standard operating procedure vs process documentation vs work instructions

These get used interchangeably and they are not interchangeable. The difference matters most when you are deciding what to automate, because only one of the three is at the right altitude to build from.

A policy is a fourth thing again — it says what people are allowed to do rather than how the work gets done. If that is what you actually need, the AI acceptable use policy generator is the other tool on this site.

How to write a standard operating procedure in about twenty minutes

Do the task once with a notepad open. Write what you did, not what you believe you do . Those are different documents, and the second one is fiction. Then fill the eight sections in order and stop. A first version that exists beats a perfect one that is still being drafted next quarter.

Then the only quality check worth running: give it to someone who has never done the task and watch, silently, while they try. Every place they hesitate is a step you wrote for yourself. Fix those, version it, and put a review date at the top. Twenty minutes to draft, ten to watch someone fail at it, ten to fix. That is the whole method.

One thing to avoid: writing procedures for tasks that run twice a year. The document will be wrong by the time it is needed and nobody will trust it. Write a note instead and spend the effort on the task that runs forty times a week.

Who should not use this, and where a screen recorder wins

If the task is a click-by-click walk through a piece of software, do not type it out. Record it. Scribe and MagicHow capture your screen while you work and turn the clicks into an illustrated walkthrough, which is faster than writing and much easier to keep current when a button moves. Both are paid products. We have not run procurement on either and this is not a review — it is an honest pointer, because for UI sequences they are the better tool.

If your real problem is that nobody reads the procedures, a generator does not fix that. Waybook and Whale wrap procedures in onboarding: assignments, tests, read receipts, a library that new starters are walked through. Also paid, also a different job from this page.

And the bigger one. If you have a single repeatable workflow and one motivated operations person, you do not need an agency at all. Write the SOP, build it in n8n, Make or Zapier, and keep the money. We say this on the first call more often than people expect. The point at which outside help earns its cost is when several procedures come back agent-ready at once, when the systems involved do not have clean interfaces, or when nobody internally can own the thing after it is built.

The limit of this page, stated plainly: it can only score what you type. If you write vague steps, it will tell you they are vague, but it cannot know that your CRM holds two records for the same company or that one client is contractually owed a phone call. Those are the details that decide whether the build actually works, and they only come out of watching the work.

What to do once the SOP exists

Fill in every square bracket, put it where your team already looks rather than in a new tool, and give it to the next new starter as their first task. Then write the second one. Three or four procedures in, you will start seeing which of them share the same systems, and that overlap is usually where the first agent should go.

To test a whole role rather than one task, the AI readiness assessment runs the same logic across ten questions. If you are still deciding what an agent even is versus a workflow or a bot, read AI agents vs workflows vs RPA, which is the honest version of that comparison. For what outside help costs and how to pick someone, our guide to AI automation consulting publishes our own prices and the cheaper alternatives. Procedures are also the raw material a company second brain indexes. The reason it works is that somebody wrote the answers down first. And when a stack of agent-ready procedures needs building, owning and managing, that is Agentic OS.

Frequently asked questions

Sources

• OSHA, 29 CFR 1910.119 Process safety management of highly hazardous chemicals — paragraph (f) requires employers to “develop and implement written operating procedures that provide clear instructions for safely conducting activities involved in each covered process”. Cited only for where the term comes from: written operating procedures started as a legal requirement in physical operations.
• US EPA, Guidance for Preparing Standard Operating Procedures (EPA QA/G-6) — cited only as evidence that the canonical SOP guidance is written for laboratory and field work, with sections on calibration, sample handling and safety warnings. That is why it does not transfer cleanly to a digital business task, and why the eight sections on this page are different.
• The readiness rubric, the quarter-of-companies figure, the token-cost story, the adoption numbers and the three stall causes are our own first-hand observations from Cognio Labs deployments, anonymised. They are experience, not research, and we have not published a study behind them.