First Principles: What Work Is Worth Automating

Every founder has an automation wish list. Most of those items are not worth automating. The cost of building the automation, maintaining it, and debugging its edge cases often exceeds the cost of doing the task by hand. Here is the first principles test I use before I build anything, and a recent example where it saved me a month of wasted work.

The Three Questions

Does the task repeat at least weekly?
Will the task exist in twelve months?
Does the failure mode cost more than the task itself?

If any answer is no, do not automate. Do it by hand. The questions are ordered by how often they kill automation projects: most often it is the first one, then the second, then the third.

The Hidden Costs

Automations look free after they ship. They are not. They carry:

Dependency upgrades every quarter
Monitoring overhead every week
Silent failures that surface weeks later and cost trust
Lock-in to specific tools or APIs that age out
Cognitive load of remembering the automation exists at all

Amortize those over the task lifetime before you start. A ten-hour build that saves one minute a week and lives for six months is a net loss even if the math looks good on day one.

A Recent Example

I almost built a scheduled scraper for a report I read twice a month. Running the numbers, the hand effort was five minutes and the automation would have cost forty hours to build and four hours a year to maintain. I kept it manual. Saved a day of work and a month of maintenance. The report still happens. I still read it. Nothing broke.

When to Automate Anyway

One exception: when the automation is also the learning. Building a small agent pipeline to understand how agents work is valuable even if the task itself does not justify it. Be honest about which project you are actually doing. If the learning is the real goal, scope it like a learning project and throw away the code afterwards.

The Counter-Test

Before I do not automate, I ask the opposite question: would I regret in six months not having built this? If the answer is no, the decision is easy. If the answer is maybe, I revisit in three months rather than committing now. Delayed decisions on automation rarely go bad.

The Final Guardrail

No automation without a retire-by date. If a scheduled script has no plan for retirement, it lives forever and slowly rots. Every automation I ship gets a sunset comment in the header.