What to automate first in a small company (and what to leave alone)
The four-question rule for deciding whether a task is worth automating, the five that almost always pay off and the three that should stay in a person's hands.
· Automation
The commonest mistake when automating is not choosing the wrong tool: it is starting with the wrong task. People almost always start with the most visible or the most annoying one, which is rarely the one that eats the most time. The result is an automatic flow that impresses in a demo and saves twenty minutes a month.
The four questions that decide it
A task is worth automating if the answer to all four is yes:
- Does it repeat? At least once a week, always the same. What happens twice a year does not get automated: it gets documented.
- Can the rules be written down? If you can explain to somebody exactly what to do in each case, it can be automated. If the answer starts with «it depends how you look at it», not yet.
- Are mistakes detectable? An error you notice straight away is an incident. A silent error discovered six months later is a problem. You only automate the second kind if there is a way of checking it.
- How much time is it per month? Multiply the minutes by the frequency. The figure tends to surprise in both directions.
The five that almost always pay off
1. Entering invoices into the accounts
It is number one in almost any business. Receive the invoice, read it, type in supplier, base, VAT and date, file it. Between one and three minutes per document, times a hundred or two hundred documents a month. Today it is read automatically and you only review what the machine flags as doubtful.
2. The alerts done by hand
Remembering a due date, warning that documentation is missing, confirming an appointment. These are also tasks nobody does consistently well, because they depend on somebody remembering.
3. Moving data from one place to another
From the shop to the warehouse, from the form to the sheet, from the bank to the report. Every manual copy is time and a chance to make a mistake.
4. The usual report
The one somebody prepares on the first Monday of each month by joining three sources. It does itself, and it also stays up to date daily instead of once a month.
5. The first sorting of the inbox
Not replying for you: sorting. Separating what is an invoice from what is an enquiry and from what is advertising, and labelling it. It removes a good deal of noise from the day.
The three best left untouched
- Anything that decides about a person. Hiring, dismissing, granting or refusing. As well as being a bad idea, it is precisely what the AI rules classify as high risk when automated.
- Anything with direct legal effect. Filing a tax return, signing, accepting a condition. It can be prepared automatically; it is sent with a person watching.
- Dealing with an angry customer. We cover that separately in chatbot or person.
The prior step everybody skips
Automating a bad process just gets it done badly, faster. Before automating, the right question is whether the step is needed at all.
A real and very common example: a company wants to automate sending a weekly report to five people. Asking around, it turns out three of those five never open it. The correct automation was to stop sending it.
Out of every ten processes somebody wants to automate, two or three can simply be dropped. That discovery is free and is usually the biggest saving of all.
What it is done with
There are three levels, and it is best not to jump to the third by default:
- What you already have. Rules in your email, templates, automatic fields in your invoicing program. Free and underused.
- A connector between applications. Tools that join two systems without programming. They solve most cases.
- Custom development. When the process is yours and looks like nobody else's, or when the volume means paying per operation stops making sense.
And a warning about generative AI inside an automatic flow: it works very well for classifying, summarising and extracting data, and it works badly for anything where a plausible but false answer would go unnoticed. If you use it to read invoices, reviewing the doubtful items is not optional.
If this sounds like you
The first question — what is worth it in your specific case — we answer in one session, with your real processes and a list ordered by what it saves: what you can automate with AI. If you already know your problem is invoices, that goes straight to your accounts without touching an invoice. And what lies behind it, in full, in AI and automation.
We are Mindset & Code: automation, data and development for small businesses. We automated our own invoicing before anybody else's, so what we describe here we paid for with our own hours first. You can see what we do and what it costs.
General guidance. The reference to high-risk systems comes from Regulation (EU) 2024/1689 on artificial intelligence, as amended by Regulation (EU) 2026/1744; its Annex III high-risk block applies from 2 December 2027.