GlossaryAutomation
Automation
Automation hands a repetitive task to software that carries it out alone, following rules written in advance, each time a trigger occurs.
An automation is described in three parts: a trigger, such as an incoming email, a date or a submitted form, some conditions, and then actions in one or more pieces of software. When a quote is signed, copying its lines into the tracking sheet and creating the invoicing task is a typical example. It is built in a workflow tool such as Make, n8n or Zapier, or with the rules built into a piece of software, and it has two qualities that are easy to underestimate: it costs little, and when a case falls outside what was planned, it stops and the failure shows. The France Num 2026 barometer measures how much room there is to grow: 40% of French small and mid-sized businesses say they use an artificial intelligence solution, but only 11% use it to automate tasks, against 34% to generate text, voice or images.
Automation or AI agent
The criterion is not how complex the task is but whether you can write out all its branches in advance. Try to write its full procedure: if you write “it depends” more than twice, the task calls for judgement and belongs to an AI agent, which decides as it runs instead of following a script. An automation fails loudly, an agent fails by producing something plausible that has to be checked, and the two combine well, with deterministic plumbing on one side and unforeseen cases on the other. Our comparison of agents and automation details the rule and the costs.
The European AI Act draws the same line. Its definition of an AI system does not cover systems based on rules defined solely by natural persons to execute operations automatically (recital 12), so a classic automation, with no model inferring anything, falls outside its scope.
What is better left unautomated
Three tests prevent most disappointments: can the result be checked faster than it is produced, is a mistake reversible, and does the time saved have a named destination? Time saved that nobody has assigned to something else dissolves into the same tasks, and setting up one automation at a time is how you learn which one pays off, as our article on automations that return nothing shows.
Two limits come from the nature of what is automated. What leaves the company, such as a message to a customer, a reminder or a reply to a candidate, deserves human approval, because a mistake sent to a hundred people cannot be taken back. And a decision about a person falls under the GDPR: everyone has the right not to be subject to a decision based solely on automated processing when it produces legal effects or similarly significantly affects them, except in the cases the text lists (article 22), and rejecting an application that nobody has actually reviewed is the most common example.
Balt turns recurring work into routines requested in a sentence, runs them at the right time and keeps the steps he has learned as scripts, waiting for your approval for anything that leaves the company.
Sources
- France Num, France Num 2026 barometer (Crédoc for the DGE, 9,655 businesses, published 17 September 2026 and updated 12 October 2026, in French)francenum.gouv.fr
- EUR-Lex, Regulation (EU) 2024/1689 on artificial intelligence, recital 12 and article 3 (read 1 November 2026)eur-lex.europa.eu
- EUR-Lex, Regulation (EU) 2016/679 (GDPR), article 22 (read 1 November 2026)eur-lex.europa.eu

