Workflow automation: decide what to automate (and what not to)

• You feel “death by a thousand cuts” from small repetitive actions.
• You want a repeatable way to choose automation candidates.

Classic automations (Zapier/n8n/Google)

• Best for stable workflows with predictable inputs/outputs.
• Usually lower risk, easier to maintain, less “magic.”

AI-assisted automations (Claude Code/Cursor/etc.)

• Best when the automation requires judgment, text manipulation, or code changes.
• Higher leverage, but higher variance and requires guardrails.

The 3V test: Volume, Variance, Vulnerability

Automate when:

• Volume is high (many runs/week)
• Variance is low (inputs don’t change much)
• Vulnerability is low (mistakes are cheap or easily reversible)

If variance or vulnerability is high, prefer:

• A simpler partial automation (reduce steps, not full autonomy), or
• A checklist/playbook (standardize the human process), or
• Delegation/training rather than automation.

Step 1: Capture one “automation candidate” exactly

Write down:

• Trigger (what starts it)
• Inputs
• Steps (5–15 bullet points)
• Output
• What could go wrong

Step 2: Score it quickly

• Volume: runs per week?
• Variance: are inputs structured?
• Vulnerability: what’s the cost of a wrong run?

Step 3: Choose the simplest tool that works

• If it’s mostly moving data between systems → n8n/Zapier/Google automation
• If it’s editing text/code with rules → AI coding assistant with guardrails

Step 4: Add guardrails

• Logging/notifications
• Dry-run mode when possible
• Clear rollback path

• Automating something you don’t understand well yet.
• Automating a broken process (you just make mistakes faster).
• Building a “house of cards” integration without monitoring.

• “AI for software engineers: Codex vs Cursor vs Claude Code” (library page)