The World Economic Forum’s Future of Jobs Report 2025 does a piece of arithmetic that needs no economics degree: if the world’s workforce were 100 people, 59 of them would need training before 2030. Eleven will not get it. The report gathers answers from more than 1,000 employers representing over 14 million workers across 55 economies, and it puts skills gaps at the top of the barriers to transforming a business — 63% of employers name them.

Eleven in a hundred is not a fact about technology. It is the accumulated result of a great many small management decisions, taken quickly and almost always in good faith.

Nobody automates a job

You automate tasks, and a job is a list of very unequal ones. Drafting a first version, sorting a request, summarising a long meeting, spotting a pattern across a thousand rows: a machine does that, and does it well. Working out why a customer is actually angry, deciding on half the information, standing behind a decision that may go wrong in front of the person it goes wrong for: that has no substitute yet, and does not look like getting one soon.

So the useful question is not which jobs disappear. It is what happens to each task once you put a system on top of it.

  • The ones the system can close on its own. Fewer than any demo promises. They share one trait: being wrong is cheap and the mistake shows up immediately.
  • The ones it speeds up but does not close. The saving is real here, though always smaller than advertised, because somebody checks. If that checking is not in the budget, the saving is an accounting entry rather than a fact.
  • The ones that must stay in human hands. Not out of sentiment: because someone will have to face a customer, an auditor or a judge, and machines do not face anyone.
  • The ones automation itself creates. Supervising, correcting outputs, keeping clean the data the system feeds on. They appear in no plan and always land on the same team.

Two rules we apply before switching anything on

First: we do not start an automation unless we know how it stops. Written down, with a named person who can stop it without asking anyone’s permission. It sounds obvious until the day you need it and the system turns out to be tangled into six other processes. That switching off costs more than switching on is not something we learned from artificial intelligence. We learned it from networks.

Second: we measure rework, not just speed. Before touching anything we write down how long the task takes today, how many errors it carries and how much time the team spends redoing someone else’s output. If a month later the time is down and the rework is up, we have gained nothing: we have moved the effort somewhere else and made it invisible. In our experience that shift explains a good share of the pilots that never become a process.

Training stopped being a goodwill gesture

The AI literacy obligations in the European regulation have applied since 2 February 2025, as the European Commission states, and the bulk of the act applies from 2 August 2026. The wording is modest; the consequence is not. Whoever deploys a system has to make sure the people using it understand what it does and where it breaks. We have written separately about who signs off on that inside a company.

There is no case for going slowly, because adoption is running ahead of understanding. Eurostat counts 19.95% of EU enterprises with ten or more employees using artificial intelligence in 2025, against 13.48% the year before. In Spain the jump was from 11.31% to 20.27%. Among large European companies it is already 55%.

Training does not mean turning the payroll into engineers. It means a salesperson knows what information must never be pasted into a chat, a manager knows when a recommendation needs a second look, and the board knows what risk it is accepting when it says yes.

Automating without redesigning does not remove work. It moves it towards whoever has the least room to argue about it.