Don’t automate a mess: three questions before you automate

Before you automate a process, check three things. Does it run the same way every time, and is it written down? Does it repeat often enough for the automation to pay back the time it takes to build and maintain? What happens when the automation makes a mistake, and who will catch it? If you answer “I don’t know” to any of these, put the process in order first.

Automation copies whatever it’s given

Two rules are attributed to Bill Gates (the authors of Universal Principles of Design cite his 1995 book The Road Ahead as the source): if you automate a process that works well, you multiply what works; if you automate one that works badly, you multiply what doesn’t. The authors of Universal Principles of Design (Lidwell, Holden, Butler) note that the first rule is a simplification. Most processes have parts that work well and parts that don’t, so automation can speed up one part and clog the next. Their example: automated order intake speeds up ordering, then fulfillment can’t keep up and the backlog grows. They advise running the process by hand first.

Michael Hammer wrote about the same thing in Harvard Business Review as early as 1990. In his view, outdated processes shouldn’t be locked into software; they should be scrapped and redesigned from scratch, or else we are “paving the cow paths.” Tim Ferriss adds the order: don’t automate anything you could simply cross off. Whether a task needs to exist at all is something I cover in the article on first steps with AI. Here I assume it does.

Take a hypothetical digital print shop with eight people. Quote requests come in by email, by phone and through a form. Two people do the quotes, and each calculates in their own way: one adds a rush surcharge, the other rounds down for regular clients. The owner wants AI to answer the requests on its own. It would answer faster, but by two price lists at once.

Question 1: does the process run the same way every time, and is it written down?

The simplest test: ask two people, separately, to list the steps of the process. If the lists differ, first agree on one version, write it down and run it by hand for a few weeks. AI can help write it up from a recording.

The content can vary, because quoting posters looks different from quoting business cards. The rules have to be shared. In the print shop, that means one price list on one page: surcharges, discounts, rounding, minimum order.

Question 2: does the process repeat often enough?

An automation has to be built, tested, and then fixed every time the form, the price list or the tool changes. So count how many times a month the process repeats and how many minutes it takes. A task that comes up once a month and takes fifteen minutes adds up to three hours a year. In my view, a checklist handles a process like that better than an automation. Quote requests at the print shop come in every day, so they meet this condition.

Question 3: what happens when the automation makes a mistake, and who checks it?

Sort the possible errors by cost and by whether they can be undone. A wrong label on an email you fix with one click. A wrong price in a quote you’ve already sent means lost margin or an awkward conversation with the client. A wrong invoice needs a correction. The more expensive the error and the harder it is to undo, the closer to a person the automation should work.

Before you switch it on, decide who checks the results, where errors go and how to roll back a change. The person who approves should see what the automation will do, on what data, and how to reverse it. Run the rules on old cases first. I describe how I did this with my email filters before turning them on on the Proof page.

So the print shop can start with an in-between version. AI prepares the quote from the written price list, and a person approves it before it goes out. Full automation waits until later, for standard products, once the quotes have needed no corrections for a few weeks.

When it doesn’t work

in a young business where the process changes every month: it’s too early to write it down, so do it by hand and note what repeats;
when one person keeps the process in their head and doesn’t want to hand it over: start with a conversation, and the writing down will come later; I cover how to talk to your team about AI in a separate article;
when the process is clumsy but gets you where you need to go: the authors of Universal Principles of Design concede that with a good strategy, automating even a poorly designed process can speed up the work; in that case, tidy it up in parallel, but have your answer to the third question before you start;
the three questions won’t tell you which process to start with; that’s a separate choice, covered in the article on first steps with AI.

For tomorrow

Ask two people, separately, to list in bullet points the steps of the process you want to automate. Compare the two lists.

Business X-ray in 5 minutes

How this text was written: the research, the choice of sources and the final version are mine. AI helped me prepare a draft; I checked every fact, and I am responsible for this text. — Maciej Chmura

Sources

William Lidwell, Kritina Holden, Jill Butler, Universal Principles of Design, 3rd updated and expanded ed. (Rockport 2023), entry 081 “Gates’ Rule of Automation” (citing Bill Gates, The Road Ahead, Viking 1995).
Michael Hammer, “Reengineering Work: Don’t Automate, Obliterate,” Harvard Business Review, July–August 1990: https://hbr.org/1990/07/reengineering-work-dont-automate-obliterate (accessed October 16, 2026).
Tim Ferriss, The 4-Hour Workweek (expanded ed., 2009), Step III “A is for Automation,” ch. “Outsourcing Life.”