Which tasks to automate with AI: the four-question filter

The scene repeats itself. A company pays for an AI subscription in January, uses it for two weeks, and by March nobody opens it. It was not bad luck or a bad tool: it was bought without knowing which task it was meant to solve. Deciding which tasks to automate with AI is preliminary work, dull and cheap, and it separates the businesses that save hours from the ones collecting subscriptions.

What happens when you buy first and think later

The symptoms are always similar. One person on the team uses the tool and nobody else knows the password. The text it produces would suit any company in the sector. And nobody can say how many hours have been saved, because nobody measured how long it took before.

The underlying problem is that AI does not fix a process: it speeds it up. If the process is badly set up, what you get is the same thing done badly, only faster and with a monthly invoice on top.

The preliminary work you cannot skip

Before looking at tools you need to know where the time goes. No software or consultancy required: a notebook and a week.

How to write it down, in three columns

For five working days, every time you finish a repetitive task, note three things: what you did, how long it took and how many times a month you do it. No embellishment. By the end of the week you will have between fifteen and thirty lines, and the list sorts itself: whatever eats the most hours goes on top.

This is the part almost everyone skips, and the only one that produces data of your own. Without it, any decision about AI is an opinion.

The four-question filter

With the list in front of you, each task goes through four questions. Four yeses make it a candidate. One no and it waits; two and it is out.

1. Does it repeat often?

Automating something you do twice a year does not pay off even if it takes you a morning. The practical threshold is several times a week: there, saving ten minutes each time adds up to hours a month.

2. Does it have clear rules?

If you can explain how it works to a new starter in five sentences, it has rules. If the answer begins with "it depends" and goes on for ten minutes, not yet: the process needs sorting out first, and AI will not do that for you.

3. Is a mistake cheap?

A badly worded draft email takes a minute to fix. A quote with the wrong price sent to a customer costs considerably more. The more expensive the mistake, the more human review you need, and the smaller the saving gets.

4. Is the information stored somewhere tidy?

AI can only work with what you give it. If your prices live in three different spreadsheets and one person's head, no tool is going to write reliable quotes. Tidying the data comes first, and it is usually the real work.

How to rank the list

For the candidates that passed the filter, add two columns: hours a month you would save and how hard it looks to set up, from one to three. Start with whatever saves the most hours at the lowest difficulty, even if it is not the flashiest.

It is the least glamorous part and the one that decides the outcome. A small company's first automation is almost always something as unremarkable as moving form data into a spreadsheet, or drafting a reply somebody then reviews.

Tasks that almost always pass

• Drafting repetitive text: answers to frequent questions, product descriptions, first versions of something that gets reviewed afterwards.
• Summarising and sorting: reading incoming emails or forms and splitting them by type, urgency or area.
• Moving information between places: from the form to the spreadsheet, from the spreadsheet to the invoicing software, from the order to the delivery note.
• Preparing template documents: quotes, standard contracts, monthly reports that always follow the same structure.
• Transcribing and organising notes from site visits, calls or meetings so they do not get lost in a notebook.

Tasks that almost never pass

Worth saying just as plainly, because knowing where not to go saves more money than the automation itself.

• Decisions with nuance: handling a delicate complaint, negotiating a price, deciding whether to take on an unusual job.
• Anything requiring technical judgement and responsibility: diagnoses, signed-off calculations, anything where a mistake has legal consequences.
• Processes that change every month and have not settled yet.
• Tasks involving sensitive personal data pasted into free tools without checking where that data ends up. The Spanish Data Protection Agency publishes guides for small businesses worth reading beforehand, not afterwards.

The two-week test before paying for anything

Once you have picked the first task, the test is simple and free: do it by hand five times with a stopwatch, then do it five times with the candidate tool, counting the time spent reviewing and correcting too. If the saving is not obvious across those five, it will not be across fifty.

This step kills half the ideas and avoids the subscription nobody opens in March. We cover it with more cases in the article on process automation.

The rules to put in writing

Before the tool enters daily use, three decisions on one page: what information it may be given and what it may not, who reviews the output before it leaves the office, and what happens when it gets something wrong.

This is not bureaucracy. In Europe the legal framework is already in motion with the European Commission's AI regulation, and for the housekeeping —passwords, access, backups— INCIBE has clear material for small companies. Half an hour of rules prevents the fright of somebody pasting customer data into a public tool.

What to measure afterwards

Three numbers and no more: hours saved per month on that task, how many times the output had to be corrected, and what the tool costs. If the saving is two hours a month and the subscription is 40 €, the maths works. If it is twenty minutes and everything needs checking, switch it off without drama and move to the next item on the list.

Measuring works the other way too: when something works, it shows quickly and can be extended to other tasks in the team.

What an AI audit is and what comes out of it

You can do all of the above yourself with a notebook. When you hand it to us, the same thing gets done in an orderly way and with the technical side solved: we sit down and go through processes and tools, point out where AI fits and where it adds nothing, and you come away with a document anyone can read, with priorities and rough costs. That is the AI audit, and it is useful whether you implement it with us or on your own.

Two roads come out of it: automating processes with off-the-shelf tools properly connected, or building something of your own when nothing fits, which is what we cover in when custom software pays off.

What it costs

AI audits, process automation and custom software are quoted per project, because the price depends on what has to be reviewed or built; publishing a fixed rate would make no sense. What is published are the web and marketing plans: 100 € a month for the landing, 200 € for the site with up to ten pages, and 400, 750 and 1,500 € for the plans with articles, Google profile, campaigns and social media. Excluding VAT and with no lock-in.

If you want the general context of AI in a small company, we have it in artificial intelligence in your company.

Start with the notebook

If you take one thing from all this, take this: write down one week. Without that list, any conversation about AI ends up being about tools and trends; with it, it ends up being about hours and money, which is the point.

And if you would rather we looked at it together, message us on WhatsApp telling us what your company does and which task eats the most time. We will tell you whether it looks automatable or not, and where we would start. No strings, same day.