Which tasks are actually worth automating? Take the minimum, not the average

Most "automate your workflow with AI" advice gives you the same instruction: automate your repetitive tasks. That advice is how people spend a weekend automating invoicing — the task that feels worst — and get 40 minutes a week back.

The problem is that repetitive is necessary but nowhere near sufficient. Half of your repetitive tasks will burn a weekend and give you nothing. Here's the screen I use to tell which half is which.

Score four axes, then take the minimum

For each recurring task, score 1–5 on:

• A — Input availability. Can a machine reach the raw material? 5 = it's in a queryable place (a sheet, Slack, Stripe). 1 = it lives in your head or a phone call.
• B — Output tolerance. What's the cost if it's wrong (not whether you can review it — almost everything is reviewable). 5 = a bad draft is a 90-second fix. 1 = a wrong invoice amount goes to a real client.
• C — Decision density. 5 = fixed shape, content is lookup. 1 = every instance is a new judgment call.
• D — Frequency. 5 = daily, 1 = monthly.

Then the one move that matters:

automatability = MIN(A, B, C, D)

Do not average. A task that scores 5-5-5-1 averages 4.0 and looks like a green light — and you'll build it beautifully and run it once a month. Automation is a chain; the weakest link is the real strength.

Why invoicing is a trap

Invoicing usually scores A5 (the amount is in the contract), C5 (fixed format) — it feels like the perfect first automation. But B1 (a wrong number to a client is real money) and often D2 (twice a week). MIN = 1. The most automatable-feeling task on your list is one of the worst actual candidates. Every ranking method that averages, or just counts repetition, gets this exactly backwards.

Then price it

real_rate = monthly_revenue / hours_actually_worked (including admin — this is your capacity rate, not your quoted rate). For each task, cost/year = times_per_week × minutes/60 × real_rate × 46. Rank by MIN first, then hours/week. The tasks with MIN ≥ 3 are your pool; sum them and you have a real number to act on.

Run it in 30 seconds

I packaged this as a free browser tool (no signup, EN/日本語/ES): Automation Audit. There's also a Claude Code / agent skill if you'd rather run it in your terminal — both are MIT-licensed, fork away.

If you sell automation to clients, sending a scored diagnostic like this before you quote is how you stop competing on a price list. That method comes from The 8-Hours-a-Week-Back AI Automation Audit — the full 90-minute audit and the three builds that recover the time.