The problem, stated precisely
Most automation fails before any code is written, in the choice of what to automate. Someone picks the task that is most annoying rather than the one that is most expensive, spends a fortnight on it, and saves twenty minutes a month. The tooling worked perfectly. The decision was wrong.
The second failure is subtler and more common. A process gets automated as someone described it, rather than as it actually runs, so the automation handles the clean case and a human quietly handles everything else — which was most of the work. Now there are two systems, and the expensive one is still you.
This method is an audit first and a build second, and it is deliberately front-loaded. Steps one to three take an afternoon and decide whether the rest is worth doing at all. Most of the value is in that afternoon.
What you will have built A costed inventory of the repetitive work, one automation on the most expensive item, honest handling of the three failures that actually occur, a single human checkpoint, and a measurement against the number you wrote down before you started.
Who this is for You run something, or you are fixing it for someone who does. You do not need to write code — several of the best outcomes here are a shared inbox rule and a form that posts somewhere.
1. Count the work before you automate any of it
For one week, log every task that happens more than once. Not a time-and-motion study. Five columns, filled in as you go:
| What | Who | How often | Minutes each | What breaks it |
|---|---|---|---|---|
| Copy invoice totals into the sheet | Ops | 60× | 4 min | Supplier sends a PDF |
| Chase unpaid invoices | Owner | 20× | 6 min | Some already paid |
| Answer a new enquiry | Owner | 90× | 5 min | Out of hours; duplicates |
The last column is the one people leave blank, and it is the one that decides whether an automation survives. "The supplier sends a PDF instead of a spreadsheet." "Sometimes the name has a middle initial." Write it down when it happens, because in a month you will swear it never does.
A week of honest logging beats a month of remembering. Nobody's recollection of their own week survives contact with the log.
2. Price every line
Now put a number on each row. Minutes, times frequency per month, times a realistic hourly cost.
The cost to use is what an hour of that person's time is worth to the business, not their wage. The hour someone spends copying invoices is an hour not spent on what you actually hired them for, and that is the real price.
Add a second number: how much it hurts when it goes wrong. A task costing an hour a month that occasionally loses an order is worth more attention than one costing four hours that nobody notices. Multiply the risk by roughly how often that bad outcome occurs, and add it in.
You now have a ranked list. It will not match your intuition. The thing that annoys you most is usually a few hours a year, and the thing at the top is usually something nobody complains about because they stopped noticing.
3. Pick one, by cost and by variance
Take the top of the list, then apply one filter: how many exceptions does column five have?
High cost and low variance is your first automation. High cost and high variance is a later one, or a partial one, and you should know that going in.
Automate one thing. Finish it. The fortnight where you have three half-built automations and are personally the glue between them is worse than the week before you started.
The test that saves fortnights If you cannot describe the task's rule in two sentences without using "usually" or "it depends", it is not ready to automate. Either narrow it until you can, or go down the list.
The one that is always worth automating
If the audit produces nothing obvious, look at how long it takes to respond to a new enquiry.
Speed of first response predicts conversion more strongly than almost anything else a small business controls, and the gap between a form submission and a human seeing it is usually hours of nobody's fault. Closing that gap to under a minute is a day's work, needs no judgement, and is measurable in the only way that matters.