Skip to content
Practical technology, explained simply

Artificial intelligence · automation

Which tasks are worth automating (and which not)

How to choose what to automate with AI without losing control: the right task, five questions first, minimum permissions and when not to do it.

Ibrahin · Published on 2 September 2026 · 3 min read

Text drafted with artificial intelligence. The topic, the angle and the final review are Ibrahin's. How AI is used here

Someone arranging repetitive tasks into a simple flow at their desk
Illustration, not a real screenshot.

What has happened

Connecting your email, documents or calendar to a tool that acts on its own is now within anyone's reach, without coding. The new AI tools don't just answer: they can sort messages, move files, draft and send. Spain's data protection authority (AEPD) has already published an analysis of this AI that acts on its own, which it calls “agentic”.

Why it matters to you

Because a mistake you used to make once, once automated, repeats every time the workflow runs, and sometimes nobody notices until the damage is done. A tool allowed to delete, pay or publish can do it on a wrong interpretation. And a fast workflow that keeps making mistakes doesn't save time: it just moves it elsewhere.

What you can do

1. Choose a task that is repetitive, measurable and reversible

For a week, note one task you repeat: sorting emails, copying data between documents, preparing a summary or flagging a change. Note how long it takes, how many exceptions come up and which mistake would be serious. If it only happens once a month or needs different judgement each time, automating it will cost more upkeep than it saves.

2. Answer five questions before you start

  1. What makes it start?
  2. What information does it need?
  3. What result does it produce?
  4. What happens if it fails or gets an odd piece of data?
  5. How do I stop it, and how do I undo the last change?

If you can’t answer the fifth, don’t connect it to real data yet.

3. Start with a version that does one thing

For example, sorting messages into a folder without deleting or replying. And separate preparing from acting: the tool can draft an email, but you approve sending it; it can find duplicate files, but you review the deletion; it can prepare a payment, but not make it without your confirmation.

4. Give it only the permissions it needs

A task that sorts an inbox doesn't need permission to delete messages or read all your documents. Use a test account if you can. And don't put passwords or access keys in a spreadsheet, an AI prompt or a shared document: they go in the place the tool provides for that. When you stop using a connection, remove it.

5. Make it keep a record of every run

Without including sensitive data, have it keep:

  • when it started;
  • what result it gave;
  • whether it asked for your approval;
  • what error it found;
  • which version of the rules was active.

That record is what tells you whether the automation helps or just moves the mistakes around.

6. Test it with what can go wrong

Don't just test the easy case. Try an empty message, a file in another format, a duplicate, incomplete data and an interruption. Check that it stops and leaves an explanation you can understand. Knowing when to stop is a feature too.

The recommendation

Choose a reversible task, test it with made-up data, watch it for ten runs and keep your approval before any action others will see, such as sending or publishing. Only expand it once you've seen its mistakes, not the first time it works.

And don’t automate it yet if:

  • it decides on someone else’s health, money or rights;
  • you can’t review what it produces;
  • a mistake would be hard to undo;
  • the process changes faster than you can keep it up to date;
  • nobody knows who answers when something goes wrong.

You can automate the safe part and keep the final decision in your hands.

Sources

Review

Responsible for publication: Ibrahin. Date: 2026-09-02. Last checked: 2026-09-02.