One SOP, written four ways
What breaks when the one person who knows how it’s done is out for a week? An SOP is supposed to answer that. In my experience most never get written, and time is why: writing one by hand takes more than a full working day, by Anthropic’s estimate.
So here’s the same SOP four ways. Each level takes more of the work off your plate. But the goal isn’t the last level. It’s the one you’re comfortable running.
The example: onboarding a new volunteer
It’s the kind of process that lives in one person’s head, and it has forms, screens and a few steps nobody remembers until they’re skipped.
Every SOP goes through the same seven steps: capture how it’s done, transcribe it, write it, add the visuals, verify it with the person who does the work, check it, and approve it. What changes from level to level is who does each step.
- YouA person does it by hand.
- You + AIA person does it in a chat with AI.
- AIAI does it with your setup, and a person checks it.
- AgentAn AI agent does it on its own and flags what it can’t pass.
- OwnerThe person who owns the SOP signs off.
What the public data says about the time
Anthropic estimates how long writing a procedure document takes a person by hand, and with AI.
About 13 times faster, by Anthropic’s estimate, for the work activity “Prepare procedural documents,” May 2026.
That’s one number for all of “with AI.” It mixes plain chat with Cowork, Claude’s more hands-off tool, so it can’t say which level gets you there, and it only counts the conversation itself.
In my experience the gain comes in steps. A chat still means back and forth, and it starts from zero every time. Setting AI up for the work is where it gets a lot better. So below, each level is measured the way I see it change: by how many of the seven steps you still do yourself.
Level 1 By hand
7 of 7
steps you do yourself.
You sit with the person who does it, take notes, type them up and build the document one screenshot at a time. Then they mark it up and you do it again.
It works. It’s also why most SOPs are out of date: nobody has a spare day to redo one.
You do all 7 steps
- CaptureYouYou sit with the person who does it and take notes.
- TranscribeYouYou type up your notes.
- WriteYouYou write each step into a blank template.
- VisualsYouYou take, crop and paste every screenshot.
- VerifyYouThe person who does the work reads it and marks it up.
- QAYouYou compare it with your other SOPs by eye.
- ApproveOwnerThe owner signs off.
Level 2 Asking AI in a chat
5 of 7
steps you still do yourself. The chat helps with 2.
This is where most people are. You paste your notes into a chat, it drafts the steps, and you fix what it got wrong. Two steps get faster. The other five are still yours.
And it’s still back and forth. In 46 of every 100 of these conversations, people built the document turn by turn. In the other 54 they handed it over, at most passing back what went wrong. Anthropic calls that kind automation, but it’s still a person in a chat, so it’s level 2, not level 4.
Every chat also starts from zero. Ten SOPs come out ten slightly different ways, and you still take every screenshot.
You do 5 of 7 steps
- CaptureYouYou still take notes, or record a voice memo.
- TranscribeMoved at this levelYou + AIYou paste your notes in, and the chat cleans them up.
- WriteMoved at this levelYou + AIThe chat drafts the steps. You fix what it got wrong.
- VisualsYouYou still take every screenshot.
- VerifyYouThe person who does the work still reads it.
- QAYouEvery chat starts from zero, so you check it matches the others.
- ApproveOwnerThe owner signs off.
Level 3 AI set up for your SOPs
3 of 7
steps you still do yourself. AI does 4, and you check them.
This is where it gets a lot better. The AI already has your template, your words for things and a few of your best SOPs. A skill holds the steps for writing one. Plugins connect it to your files and your recordings.
The person who does the work shows it once, on a recording, and every SOP comes out the same way.
It’s still working with AI, not handing it off. You start every run and check every step. This is where my own SOP work sits today.
You do 3 of 7 steps
- CaptureYouYou record the person doing the work, once.
- TranscribeMoved at this levelAIA plugin transcribes the recording.
- WriteMoved at this levelAIA skill writes it in your template, with your words for things.
- VisualsMoved at this levelAIIt pulls stills from the recording and draws the flowchart.
- VerifyYouThe person who does the work reads the draft.
- QAMoved at this levelAIA skill checks it against your SOP checklist.
- ApproveOwnerThe owner signs off.
Level 4 AI agents, each with one job
1 of 7
step you do yourself: the sign-off. Six agents do the rest.
This is the most hands-off it gets. The person who does the work walks through it once. Six agents take it from there, each with one job: one records and files the walkthrough, and the others transcribe it, write it, add the visuals, verify it and check it.
Each checks the work of the one before it. That’s what makes it safe enough to run, and it still ends with a person: the owner reads what got flagged and signs off. Nothing lasts that nobody owns.
It’s the best this gets, and most teams don’t need it. I haven’t built level 4 for a client yet, and I wouldn’t start anyone here.
You do 1 of 7: the sign-off
- CaptureMoved at this levelAgentA recording agent captures the walkthrough and files it. That starts the run.
- TranscribeMoved at this levelAgentA transcription agent turns it into text.
- WriteMoved at this levelAgentA writing agent drafts it from the transcript, your template and your other SOPs.
- VisualsMoved at this levelAgentA visual agent pulls the screenshots and draws the flowchart.
- VerifyMoved at this levelAgentA verifier agent checks every step against the recording and flags anything nobody said.
- QAMoved at this levelAgentA QA agent checks format, terms and missing steps, and sends what it can’t pass to a person.
- ApproveOwnerThe owner reads what got flagged and signs off.
Where the work moved
All four levels side by side. Each level takes more off your plate: you do 7 steps, then 5, then 3, then 1. The owner’s row never changes.
| Step | 1. By hand | 2. Chat | 3. Set up | 4. Agents |
|---|---|---|---|---|
| Capture | You | You | You | Agent |
| Transcribe | You | You + AI | AI | Agent |
| Write | You | You + AI | AI | Agent |
| Visuals | You | You | AI | Agent |
| Verify | You | You | You | Agent |
| QA | You | You | AI | Agent |
| Approve | Owner | Owner | Owner | Owner |
Which level does your SOP need?
Level 4 isn’t the goal. The right level is the one you’ll actually use, and for most people that’s somewhere in the middle. Answer four questions and I’ll show you my rule of thumb.
My rules of thumb
- A few SOPs a year: level 2. A good chat and your notes are enough. Setting up tools for four documents a year costs more time than it saves.
- No shared template yet: write one first. A chat can help you draft it. Until there’s one way to write an SOP, there’s nothing for level 3 to follow.
- A steady flow of SOPs and a template: level 3. Set the AI up once with your template and your words for things, and every SOP after that comes out the same.
- Several a month and nobody owns them: level 3, and name an owner. Level 4 with nobody responsible for it breaks quietly.
- Several a month, one template, one owner: level 4 is worth a look. Start with one agent, the transcriber or the checker, not all six.
Then stop where you’re comfortable. A chat you know how to steer beats agents you don’t trust. Most people get the most from the middle.
What the numbers leave out
There’s one public number on this page, and the levels are mine. Here’s what that number is, and what it isn’t.
- It’s an estimate. An AI read each sampled conversation and estimated how long the task would take a person alone and with AI. Nobody ran a stopwatch.
- It’s writing procedures in general, not your SOP. The number is for the work activity “Prepare procedural documents.” For the narrower task of reviewing and updating SOPs, the estimate is 5 h 31 min alone and 31 min with AI, from far fewer conversations.
- It ends when the conversation ends. It doesn’t count getting time with the person who does the work, the screenshots or the review rounds.
- It can’t tell the levels apart. The data mixes plain chat with Cowork, Claude’s more hands-off tool, and doesn’t split out skills, plugins or agents. I checked the whole file. The step counts come from my own model of the work, not from the data.
- It’s steady, which isn’t the same as right. April’s estimate was nearly the same: 9 h 26 min alone and 42 min with AI.
The biggest thing the numbers leave out is whether anyone follows the SOP. That part is training and ownership, not writing.
Bring one SOP
Show me one SOP you keep meaning to write, or one that’s out of date. We’ll work out together which level it actually needs, with the people who do the work and the tools you already have.
That’s Better Way: one recurring task, $500, online or in person when we can. If you don’t leave with a clear, practical better way forward, I refund the $500.
Or call Sarah, my AI assistant, at (920) 645-0459. She can answer questions and book the same call.
Keep going
- How much time does AI save in your job?Search 717 jobs in the same public data.
- Prepare the HumansWhere I write about the people side of AI at work.
- The free process assessmentTwelve questions about one process, and a one-page brief.
- Try what I builtAn AI phone assistant, a quote builder and a working dashboard.
Sources and credits
Time estimates. Anthropic Economic Index, data release of June 26, 2026: Claude chat and Cowork conversations, May 2026 (April for the comparison). Maxim Massenkoff, Eva Lyubich, Szymon Sacher, Zoe Hitzig, Shaoyi Zhang, Ryan Heller and Peter McCrory, “Anthropic Economic Index report: Cadences,” Anthropic, June 26, 2026. The data is released by Anthropic under a CC BY license.
What I changed. I used two rows: the work activity “Prepare procedural documents” and the task “Review and update standard operating procedures or quality assurance manuals.” I rounded the times for reading and worked out “times faster” as the average time alone divided by the average time with AI. The four levels, the seven steps and the volunteer example are mine, not Anthropic’s.
Work activity and task names. This page includes information from the O*NET® database by the U.S. Department of Labor, Employment and Training Administration (USDOL/ETA), as used in the Anthropic Economic Index. Used under the CC BY 4.0 license. O*NET® is a trademark of USDOL/ETA. Build With You AI has modified all or some of this information. USDOL/ETA has not approved, endorsed, or tested these modifications.
No affiliation. Build With You AI isn’t affiliated with Anthropic or the U.S. Department of Labor, and neither reviewed or endorsed this page. Anthropic, Claude and Cowork are trademarks of Anthropic.
Updated October 31, 2026. Spot a mistake? Email me and I’ll fix it.