What to automate, and what never to

Most people who conclude that automation "wasn't worth it" automated something that happens four times a year. The tooling wasn't the problem. The selection was.

The four conditions

Automate when all four hold:

It recurs. Weekly or more. A twice-yearly task will take longer to automate than to do.

The steps are stable. If the process changes every time, you'll spend more time maintaining the automation than running it.

The inputs are structured enough. Not perfectly — an AI step handles messy input well — but predictable in kind.

Failure is detectable and recoverable. You'd notice it broke, and fixing it costs an apology rather than a lawsuit.

Miss any one and reconsider.

Never automate

Anything where a silent failure causes real harm. Payments, safety, legal deadlines, medical, anything regulated. Not because it's technically impossible, but because "it ran and quietly did the wrong thing for six weeks" is the characteristic failure mode of automation — and for these, that's unacceptable.

Relationships. Personalised outreach at volume is spam with better grammar. Recipients can tell, and it costs you the relationship you were trying to build.

Judgement with consequences. Hiring, firing, credit decisions, pricing exceptions. Automate the preparation. Never the decision.

Anything you don't understand. If you can't do it manually and explain each step, you cannot design its checkpoints and you cannot debug it.

Do the maths before you build

Automation isn't free. Budget honestly:

• Build: two to eight hours for something real
• Maintenance: roughly an hour a month, forever
• Debugging: a bad afternoon, once or twice a year

A task taking 10 minutes weekly costs about 8.5 hours a year. Worth automating, but only just, and only if the process is stable. Two hours a week is an obvious yes.

Note what this implies: frequency beats duration. The 10-minute thing you do daily costs 43 hours a year. The two-hour monthly report costs 24. Intuition consistently gets this backwards, which is why the audit is worth doing on paper rather than from memory, or in the automation ROI calculator, which shows its sum.

The manual dry-run

Before you open any tool, do the whole thing by hand once and write down every step.

Everyone skips this. It's the difference between an automation that works and three days of confused debugging.

You'll find steps you didn't know you were doing. You check the exclusion list, you skip the internal ones, you reformat the date because the other system wants it differently. You'll find decisions you make without noticing: "if it's under £500 I just approve it" is a branch, and it has to be in the automation. And you'll find the edge cases: no new records, two records for the same person, a record with a missing field. Every one of those will happen.

Your numbered list is the build spec. (The first-workflow implementation checklist walks through turning it into one.) Every step maps to a step in the tool, every branch to a condition, every edge case to an error path.

Building becomes transcription. People who skip this aren't building, they're discovering their own process through trial and error inside a tool, which is the slowest possible way to learn it.