SpunkArt
Long reads and site policies

Choosing tasks worth automating: a practical method

Automation pays off when the time it saves is larger than the time it takes to build and look after. That sounds obvious, yet many automations never earn back their setup. This guide gives a way to pick tasks and a worked example of the maths.

Start with an inventory, not a tool

For one normal week, write down every repeated task you do on a computer. Note three things for each: how long it takes, how often you do it, and what goes wrong when you rush it. Do not judge yet. Most people find a handful of tasks that feel small but add up, such as renaming files, copying figures between spreadsheets, or sorting the same kind of email.

Score each task

Give each task a quick score from 1 to 3 on five questions:

  1. Frequency. Daily scores 3, weekly 2, monthly 1.
  2. Stability. Are the steps the same every time? The more stable, the higher the score.
  3. Clear inputs. Does the data arrive in a predictable format, such as a CSV export, rather than a scanned letter?
  4. Low cost of error. If the automation gets it wrong, is that easy to notice and fix?
  5. Annoyance. Is it dull enough that you make mistakes doing it by hand?

Tasks scoring 12 or more out of 15 are good candidates. Tasks that score low on stability or cost of error are better left manual, or automated only partway with a human check at the end.

Rules first, AI second

If a task can be described as exact rules (“move every invoice PDF into the folder for its month”), use a plain script, a spreadsheet formula or a scheduled job. Rules are fast and predictable. Add a language model only for the step that genuinely needs judgement, such as reading a free-text message and deciding which category it belongs to. Keep a human review on that step until you trust it.

Worked example: the time-saved maths

Example task. Each order confirmation email has to be copied into a tracking sheet. It takes about 6 minutes, and it happens 15 times a week. Assume 48 working weeks in a year.

  • Time per week: 6 × 15 = 90 minutes, or 1.5 hours.
  • Time per year: 90 × 48 = 4,320 minutes = 72 hours.
  • Estimated build time for the automation: 10 hours.
  • Estimated upkeep (fixing it when the email layout changes): 1 hour per month, or 12 hours a year.
  • First-year saving: 72 − 10 − 12 = 50 hours.
payback weeks = build hours ÷ (hours saved per week − upkeep hours per week)

Upkeep of 12 hours a year is 12 ÷ 52 ≈ 0.23 hours a week. So payback is 10 ÷ (1.5 − 0.23) ≈ 7.9 weeks, about two months. That is a clear yes.

Now change one number. If the same task happened only 15 times a month, it would save about 18 hours a year (90 minutes × 12 months), and 10 hours of building plus 12 hours of upkeep would wipe that out. Frequency usually decides the answer.

Costs people forget

Hidden costWhy it matters
UpkeepOther systems change their export format, login flow or layout, and the automation breaks.
Silent failureA broken job that nobody notices can do more harm than doing the task by hand.
Running costPaid services, API calls and model usage cost money per run. Multiply by frequency.
Knowledge lossAfter a year, nobody remembers how to do the task manually when the automation stops.
Access riskAn automation often needs a password or token. That access must be stored safely.

Build it safely

  • Run the automation on a copy of the data first.
  • Make it write a short log: what it did, how many items, any errors.
  • Send yourself an alert when it fails, not just when it succeeds.
  • Write a five-line note explaining how it works and how to do the task by hand.

Questions

Should I automate a task I do once a year?

Usually not. A clear written checklist is better. You will forget how the automation works by next year anyway.

Is AI automation reliable enough for important work?

For drafting, sorting and summarising, often yes with review. For anything where an error is costly or hard to spot, keep a person checking the output.

What if I cannot code?

No-code tools and spreadsheet features cover many simple cases. The same maths applies: estimate the build and upkeep honestly.