What to automate first

Goal. This template helps you choose the one task in your business to automate first, instead of chasing every piece of AI news. You count how much time each repeating task takes and what happens if the automation gets it wrong. At the end you have one decision: what, who, by when, how you will know it works, and when you stop it.

In 15 minutes. After this you know which task to start with and when to stop. The time is our estimate, not measured. You take the numbers from memory, so this picks a candidate, it is not a calculation: before you build anything, check them against your calendar and email from the last two weeks.

1. The list (part A), 3 minutes. Write five tasks that repeat in your business, each as a verb and a thing (“I retype orders from emails”). Cross out any that could simply stop existing.
2. Three numbers (part A), 4 minutes. For each task, write: how many times a month, how many minutes each time, how many errors or corrections in the last three months. Work out the hours a month: times × minutes ÷ 60.
3. Gates (part B), 3 minutes. Drop a task if it is done differently each time or has no written steps, if a mistake reaches a client or the money and nobody checks it, or if it takes less than about 2 hours a month.
4. Choice (part C), 3 minutes. From the tasks left, take the one with the most hours. If it’s a tie, take the one in your bottleneck (see the template “A walk from inquiry to payment”). Write the “before” number: hours a month and errors.
5. Stop rule (part E), 1 minute. Finish: “I stop and go back to doing it by hand if by … there are more than … errors, or checking one result takes more than … minutes.”
6. Decision, 1 minute. Write one line: “I decide that … Who … By when … I will check …”. For example: “I decide that we automate answers about order status first. Who: Marta in the office. By when: a first version in 2 weeks. I will check: in 30 days, how many status questions we still answer by hand (today 60 a month, target under 20).”

The full version below adds five checks before you start and a table that shows why the other tasks dropped out.

Instructions (full version) (about 45 minutes, an estimate, plus time to check the numbers in your calendar and email)

1. Look at your calendar and email from the last two weeks and write 5 to 10 tasks that repeat. Cross out the ones that could simply stop existing.
2. For each task, write the three numbers and work out the hours a month (part A).
3. Go through the three gates (part B). A task that fails a gate isn’t lost: it gets the result “fix first” or “stays with a person.”
4. From the tasks that passed, choose one, and choose the cheapest kind of first version (part C).
5. Go through the five checks (part D). Before you build the automation, run the process by hand, following written steps. Write the decision and the stop rule (part E).

Template fields

A. List of tasks (5 to 10)

Hours a month = times × minutes ÷ 60. If a task could simply stop existing, cross it out before you count it.

B. Three gates

If a mistake goes to a client, to money or to an authority and can’t be undone, the result is “stays with a person” or “AI prepares, a person approves,” never fully automatic. The 2-hour line is our rule of thumb, not a research result.

C. Choice

• Chosen task: ______ · hours a month: ___ · errors in 3 months: ___
• Why this one and not the next: ______
• Is it in my bottleneck? ☐ yes ☐ no
• First version (take the cheapest that is enough): ☐ a rule or template with no AI (for example an automatic email using data from your system) ☐ AI prepares, a person approves ☐ fully automatic (rare: only when a mistake is cheap and reversible)

D. Five checks before any automation

1. Data: where does the automation get its data, and is it complete and in one place, not in someone’s head? ______
2. Owner: who, by name, answers for the result and fixes things when the automation gets it wrong? ______
3. Rollback: how do you undo a mistake, what does it cost, and who switches the automation off? ______
4. Human check: who checks the result before it goes to a client, into money or into records, and how many minutes does it take? If checking takes as long as the work itself, drop it. ______
5. Privacy: does the data include personal data of clients or employees, and are you allowed to use this tool for it (contract, terms)? This is not legal advice; for personal data, ask a lawyer. ______

E. Decision and stop rule

Stop rule: “We stop and go back to doing it by hand if by ______ there are more than ___ errors, or checking one result takes more than ___ minutes, or nobody has time to check.”

F. One number every week

• The “before” number, counted the same way: ___
• Paired quality measure (errors, corrections, complaints): ___
• Check: ___ (in 30 days) · next task on the list: ___ (only after the check)

Fictional example (the firm and the numbers are fictional, for this template)

An office supplies wholesaler, 9 people.

A.

Crossed out before we started counting: retyping the daily warehouse report into a spreadsheet. Nobody reads it, because the weekly report shows the same thing.

B.

The firm’s bottleneck (from the walk from inquiry to payment): entering orders.

C. Chosen task: answering “where is my parcel,” 4.0 hours a month, 1 error in 3 months. Why this one and not the next: the most hours among the candidates (4.0 against 3.0 and 2.5), a small cost of a mistake, and the answer always comes from the system. Task 1 had the most hours (8.0) but failed the first gate, so we sort out the orders first. Is it in the bottleneck? No, but the hours freed here go to the office entering orders, which is our bottleneck. First version: a rule, no AI. When an order is packed, the system sends the client an email with the status and the tracking number. AI stays for questions outside the status, as a draft reply for approval.

D.

1. Data: the order status is in the warehouse system, and the picker enters the tracking number. For a week we check that it is always entered.
2. Owner: Marta in the office. She answers for the result and fixes wrong statuses.
3. Rollback: we fix a wrong status with a follow-up email. Marta switches the automation off in the system settings, which takes a few minutes.
4. Human check: for the first two weeks Marta reads every email sent, in the morning (about 10 minutes a day), then every tenth one.
5. Privacy: the email contains the order number, the status and the tracking number, nothing beyond what the client already knows. No client data goes into any AI tool.

E.

Stop rule: “We stop and go back to answering by hand if by the end of the first month more than 3 clients got a wrong status, or Marta has to correct more than every tenth email, or nobody has time to read the emails sent.”

F. Number: status questions answered by hand in a week (today about 15). Paired quality measure: wrong statuses and client complaints. Check in 30 days. Next task on the list: the weekly report, only after the check.

From us. This template rests on four ideas from books: Ethan Mollick (Co-Intelligence, ch. 3 and 6) writes that the line between what AI does well and badly is uneven and invisible, so you test it on your own tasks, that a person stays “in the loop,” and that fully automated tasks are for now a very small group; Lidwell, Holden and Butler (Universal Principles of Design, the entry “Gates’ Rule of Automation”) write that automation magnifies both efficiency and mess, so run the process by hand first; Eliyahu Goldratt (The Goal, ch. 19 and 28) shows that an hour saved outside the bottleneck does not speed up the whole; Tim Ferriss (The 4-Hour Workweek) says not to automate what you can eliminate. Goldratt wrote about a factory, so carrying this over to a service business is our step. What we added is the three gates, three numbers for each task, the five checks and the stop rule.

When to call us. Call us if:

• after 15 minutes two people in your company describe the same task with different steps, and that task is meant to be first;
• an automation mistake would reach a client, money or an authority, and you have nobody to check the result before it goes out;
• the chosen task touches personal data of clients or employees and you don’t know whether you may use the tool for it.

The first call with us is 30 minutes and free. Bring your list of tasks.

Prompt for your AI. This prompt works in your own ChatGPT, Claude or Gemini account and knows only what you tell it; our own method, built on what your company knows, is what we do in a Diagnosis.

You are a calm coach. Take me through the “What to automate first” card, the 15-minute version.
Rules:
1. Ask one question at a time and wait for my answer.
2. If an answer is vague, ask for a number or example.
3. Never invent anything about my company. Write unknowns as “to check”.
4. Don't choose the task for me. Show which drop out and why; I choose.
At the start, tell me: “Do not paste passwords, account numbers or personal data.”
Questions: (1) five repeating tasks, each a verb and a thing; (2) for each: times a month, minutes, errors in 3 months, then the hours; (3) three gates: same way every time, cost of a mistake, at least about 2 hours a month; (4) one task and the “before” number; (5) a stop rule.
At the end, write one line: “I decide that … Who … By when … I will check …”.