Lesson 15

SOPs, AI and Automation — Make the System Work

The Sequence That Most People Get Wrong

This lesson is not about learning Zapier or Make. It is about something more fundamental:

Do it manually → Document the steps (SOP) → Build a template ↓ AI assists the repetitive parts ↓ Automate what is fully stable ↓ You handle only exceptions

The sequence matters. Automating a process you do not yet understand produces automated mistakes at scale. First do it right. Then write it down. Then let the machine do the repeating.

Why SOPs Are Not Optional Beyond the First Site

With one website, you can hold everything in your head: where GA4 is, how the sitemap is set up, what to check before launch, what the prompts are.

With six websites, your memory becomes a single point of failure. With ten, it breaks entirely.

The deeper problem is not remembering what to do. It is the decision cost of figuring it out again from scratch every single time. Every re-decided question was already answered the last time — and the answer was the same. SOP captures that answer so you never pay that cost twice.

A Good SOP Is a Short Checklist

Most corporate SOPs fail because they are 40-page documents that no one reads. An OPC SOP should be a checklist you can act on in real time.

Example — New Website Launch:

That is a complete launch SOP. No document to hunt for. No wondering if you missed something.

The Four SOPs Every OPC Should Build First

SOPWhat it captures
Content productionFrom idea to published post — keyword check, intent, draft, human edit, fact check, internal links, SEO, publish, post-publish review
Product launchFrom requirement to live — design, build, test, launch checklist, first analytics review
Customer supportHow to classify and respond to each message type: bug, refund, how-to, feature request, spam
Weekly data reviewExactly which numbers to check, what questions to ask, what decisions to make

Rule: the third time you do something the same way, write the SOP. First time is exploration. Second time is confirmation. Third time is a pattern — and patterns belong in a document, not your head.

Content Production: Quality Over Volume

The worst version of AI-assisted content: "AI finds keywords → AI writes 100 articles → auto-publish." That is automating garbage at scale.

A useful content SOP goes through real gates:

Real question or keyword found ↓ Is it worth writing? (search intent, competition, relevance) ↓ Gather reliable sources ↓ AI structures outline ↓ AI writes first draft ↓ Human edits for accuracy and voice ↓ Fact check ↓ Internal links added ↓ SEO basics checked ↓ Publish ↓ Review at 30 / 60 / 90 days ↓ Update / keep / merge / remove

That last step matters as much as the first. Publishing is not the end of the process. It is the start of the measurement loop.

Where AI Fits — and Where It Does Not

AI is good at

  • Structuring outlines
  • First drafts from a brief
  • Summarising sources
  • Format conversions
  • Translation drafts
  • Code boilerplate
  • Classifying support emails
  • Spotting anomalies in data
  • Generating meta descriptions

You must keep

  • What to build next
  • Which projects to stop
  • Whether to trust output
  • Product direction
  • Final quality call
  • Spend decisions
  • Whether to launch or not
  • Founder relationships
  • Brand voice and values

AI removes low-value cognitive load. You keep judgment.

Three Questions Before Automating Anything

  1. 1

    Is it genuinely repetitive?

    If every instance is different, AI adds limited value. Automation requires predictable inputs and outputs.
  2. 2

    Is "good vs bad" objectively clear?

    Checking if a canonical tag exists: yes. Deciding whether this article is worth writing: no.
  3. 3

    How bad is an AI mistake here?

    Wrong button label: low cost. Auto-processing a payment dispute, publishing medical information, or sending bulk email: high cost. Higher cost = more human review required, not less.

The Four Levels of Automation Maturity

Level 1

Checklist

New site launch SOP. You do the work; the list prevents skipped steps.

Level 2

Template

One-Person Site Starter. The structure is pre-built; you fill in the content.

Level 3

AI Assisted

AI writes the draft, generates the code, or classifies the email. You review and approve.

Level 4

Automated

Git push → Cloudflare deploys. Payment received → confirmation email sent. No human in the loop for routine cases.

A functioning OPC typically runs all four levels simultaneously. You do not graduate from one to the next — you apply the right level to each specific task.

Human in the Loop: The Right Model for Early OPC

Human in the Loop means: AI does 80% of the mechanical work, you handle the final 20% that requires judgment.


Example — customer support:

Incoming email ↓ AI classifies: Bug / Refund / How-to / Suggestion / Spam ↓ AI drafts a response ↓ You review (30 seconds) ↓ You send

That is very different from AI deciding unilaterally whether to issue a refund. The judgment stays with you. The grunt work does not.

When Zapier and Make Are Worth It

A simple formula for deciding whether automation is worth building:

Frequency × Time per instance × Error cost > Setup cost + Maintenance cost

A task that happens weekly and takes two minutes: do it manually. A task that happens 50 times per day and takes three minutes each: that is 2.5 hours per day — automate it immediately.

Early OPC trap: spending three days configuring Zapier + PayPal + Sheets + webhooks to handle two orders per month. That is negative ROI. Automate genuine bottlenecks, not low-frequency tasks that feel inefficient.

Payment Automation: The One Place to Be Careful

Once real orders arrive, manual fulfillment (read PayPal email → manually send PDF) will not scale. An automated order flow looks like:

Payment completed (PayPal/Stripe webhook) ↓ Server validates: real payment? correct amount? correct currency? ↓ Order created ↓ Product access granted / download link sent ↓ Confirmation email sent ↓ Transaction logged

Critical rule: always validate on the server side using the official webhook from the payment provider. Never trust the frontend to report "payment succeeded" — that signal is trivially spoofed. The closer the automation is to money, identity, or permissions, the more important server-side verification becomes.

Email Automation Done Right

A simple post-purchase sequence for a $4.99 digital product:

Day 0 Purchase confirmation + product delivery Day 1 Quick tip for getting the most from it Day 7 "Any questions?" — genuine, not promotional Day 14 Invitation to leave an honest review

The danger of email automation is the same as any automation: it scales whatever you already have. A helpful sequence helps at scale. A pushy promotional sequence becomes harassment at scale. Design the sequence as if you were sending each email yourself.

Data Review Automation

Managing multiple websites manually means logging into GA4, GSC, revenue dashboards, and cost trackers one by one. With five sites this is tedious. With twenty it is impossible.

The target state:

GA4 + GSC + Revenue + Costs ↓ Unified dashboard ↓ Weekly digest ↓ AI flags: what went up? what went down? which page is anomalous? ↓ You ask: why? and what next?

You move from reading every number to responding only to signals that matter.

Management by Exception

A mature operating system should not require the founder to monitor normal operations. Normal is not news.

Normal payment processed → no notification needed Payment failed → notify founder Normal deployment → no notification needed CI check failed → notify founder Normal traffic → no notification needed Traffic drops 60% overnight → notify founder

If your dashboard is full of green, you should be able to close it and go work on something that matters. The system's job is to tell you when something breaks — not to demand daily attention to confirm nothing is wrong.

The Three Roles in a Functioning OPC

System / Automation

  • Repetitive execution
  • Monitoring & alerts
  • Checks & validation
  • Deploy & delivery
  • Data collection

AI

  • First drafts
  • Classification
  • Anomaly detection
  • Summarising inputs
  • Code generation

You (founder)

  • Direction & priorities
  • Judgment calls
  • Product decisions
  • Risk & money
  • Final approval

The goal is not a company with no humans. It is a company where the human's time is spent on decisions — not on tasks the system can handle.

What to SOP First: A Priority Order

SOPPriorityWhen to build
New website startupCriticalNow
Pre-launch checklistCriticalNow
Weekly portfolio reviewHighNow
Content productionHighNow
New tool / feature releaseHighNow
Bug handlingHighNow
Customer support categoriesWhen first emails arriveAt first users
Payment & refund handlingBefore first chargeBefore launch
Email lifecycleAfter first customersAfter first customers
Automated data digestWhen managing 5+ sitesWhen managing 5+ sites

Currently active automation worth your attention: GitHub → Cloudflare deploy, CI checks, and a simple dashboard data pull. You do not need a SaaS stack to get started.

SOP → Template → Software: The Leverage Ladder

The highest-leverage version of an SOP is when it becomes code.

SOP: "every new site needs canonical, sitemap, GA4, robots.txt..." ↓ Template: One-Person Site Starter already includes all of it ↓ Software: Claude builds new sites from the template automatically

You no longer tell Claude how to set up canonical tags on every new project. The template already knows. That is what it means to not solve the same infrastructure problem twice.

Eight Things to Carry from This Lesson

  1. SOPs save decisions, not just time. Every re-decided question was already answered before.
  2. Do it right manually first, then document it, then automate it. Never in the reverse order.
  3. The third time you do something the same way, write the SOP.
  4. AI handles repetitive cognitive work. You keep direction and judgment.
  5. SOP becomes template; template becomes software. Each step multiplies leverage.
  6. Automate genuine bottlenecks, not low-frequency tasks that feel inefficient.
  7. Payment automation requires server-side webhook validation. Never trust the frontend to confirm a payment.
  8. The target is Management by Exception: normal operations run silently; you respond only when something breaks.

"The goal is not full automation. It's getting the system to run, AI to assist, and you to decide — instead of rebuilding the same thing and fighting the same fires every week."

Discuss this lesson on GitHub →