You often hear that manual trades would be protected and office jobs threatened. That's a misleading simplification in both directions. A plumber won't be replaced any time soon, that's true, but neither will a lawyer, and not for the reasons people imagine. Three criteria allow a far finer assessment of an activity's real exposure.
Over the past few weeks, we've covered AI-driven job cuts at Meta and Allianz. The question everyone is asking is legitimate: what about my job? Here's a framework for answering it beyond gut instinct.
What the data actually shows
Let's start with a figure that contradicts the cliché. Google's ATLAS study, which we analysed in a dedicated article, shows that non-routine cognitive tasks, such as creative design or hypothesis formulation, account for 65% of AI usage, versus 35% of the broader economy.
In other words, AI is being used massively precisely where it was thought to be weakest: on creative and complex tasks. Conversely, the same study finds that less than 10% of interactions fully automate a task, and that usage also reaches predominantly manual occupations, on their administrative side.
The manual-versus-intellectual divide therefore doesn't accurately describe reality. Here are three criteria that work better.
Criterion 1: the cost of an error
This is the most discriminating criterion. An AI produces plausible results, with a low but non-zero error rate, and never signals its uncertainty, as we explained in our article on calibration.
In an activity where an error costs little and is quickly corrected, automation advances rapidly. In an activity where an error affects a life, a liberty, a fortune or a reputation, someone must remain accountable. Not because humans make fewer mistakes, but because a system cannot answer for an error, in the legal and moral sense, as we detailed in our article on responsibility.
This criterion protects the anaesthetist, the pilot, the trial lawyer, the crane operator. It protects the product-description writer far less.
Criterion 2: the share of relationship in the value
Second criterion: to what extent is what you sell inseparable from the fact that it's you?
An AI can write a technically flawless letter of condolence. It cannot sign it as someone who knew the deceased. Software can produce an optimal training plan; it cannot be the person who waits for you on Tuesday morning and notices you're not doing well.
This criterion protects care work, education, negotiation, management, complex sales. It also protects, more prosaically, all those whose clients want someone they can shout at.
An AI manipulates language. As soon as a task requires perceiving an unforeseen material situation and reacting to it with the body, it hits a wall. That's what we noted about its clumsiness with even elementary arithmetic: these systems have never touched anything. A plumber facing a non-compliant installation, in a cramped cupboard, with an unobtainable part, solves a problem no corpus describes. Robotics is progressing, but far more slowly than software, for physical reasons.
What these criteria imply
Cross them and you get a different reading of the initial cliché.
Highly exposed are activities where error is cheap, relationship is weak and the work is purely document-based: generic content production, repetitive administrative processing, routine translation, first-level support. Many of these roles are skilled office jobs, which explains why the wave is hitting populations who thought they were protected.
Little exposed are activities that combine at least two criteria: care, skilled craft, intervention trades, teaching, anything that involves deciding under uncertainty while bearing responsibility.
In between, most jobs won't disappear but will be recomposed: the document-based share will be absorbed, and the share of judgement and relationship will take proportionally more room. An accountant will spend less time on data entry and more on advising. That's a demanding transformation, because it often requires skills that initial training didn't provide.
The nuance that matters
We should restate what we highlighted about Meta: AI doesn't need to replace a job to eliminate positions. If it makes a team 20% more productive, management can decide to do the same work with fewer people. Your job can be structurally protected and your position cut anyway.
That's why the useful question may not be will my job disappear, but what share of my work falls within what the machine does well. If the answer exceeds half, recomposition concerns you, whatever your job title.
One final remark, more optimistic. The three criteria we've just described—responsibility, relationship and anchoring in the real—point to exactly what modern societies have tended to underpay. If value shifts toward these qualities, the consequence could be less bleak than the prevailing narrative suggests. Provided, of course, that pay follows, which is by no means automatic.