About

The study

An assessment of how exposed 778 French occupations are to two technologies that do not follow the same calendar: automation and generative AI.

Three stages since 2018

2018

First study on automation

Emeric Lebreton, doctor of psychology and researcher in human sciences, rates 200 occupations on four factors predicting physical automation. The work leads to the book Révolution Robot.

2021

Extension to nearly 800 occupations

Expert judge rating is extended to every major occupational family. The assessments are standardised, then passed to OPTIMA for statistical processing.

2026

Generative AI added

A second dimension is added, built by task-level breakdown rather than overall rating. This is the study WorkmetriQ makes available, with its two independent scores and three horizons.

Why two separate dimensions

Most studies published since 2013 reduce automation to a single figure. That aggregation hides what matters, because the two technologies affect different tasks, different occupations and at different speeds. A bricklayer is highly exposed to automation and almost untouched by generative AI. A legal writer is in the opposite position. A single figure would have placed both at the same level of risk, which helps nobody make a decision.

The study therefore produces two independent scores for each occupation, each between 0 and 100. They are read separately and are not meant to be added together.

What a score measures

The score expresses the share of the occupation's tasks that the technology can technically automate. It describes a technical possibility, not an employment forecast. Moving from technical feasibility to the actual removal of a job depends on equipment costs, regulation, industrial relations and customer acceptance, none of which is modelled here.

The four exposure categories

• Preserved. The core of the occupation rests on capabilities the technology does not reproduce. Work organisation changes little.
• Preserved with productivity gains. The occupation remains, the technology takes over peripheral tasks and frees up time.
• Transformed. A significant share of tasks changes hands. The occupation survives but its content shifts, which calls for reskilling.
• High substitution. Most tasks are technically automatable. The occupation does not mechanically disappear, but its content empties of its present substance and retraining becomes a serious question.

Three horizons

Each occupation is projected at three points: the current situation in {{Y}}, a short horizon of 3 to 5 years covering {{S}}, and a horizon of 8 to 10 years covering {{L}}. The two dimensions progress at different rates, set out in the methodology.

Who it is for

People preparing a career change or a first career choice, the consultants who support them, companies building their workforce planning, and teachers and guidance staff. The results are there to open a discussion, not to close it: a high score flags an occupation to watch, never an occupation to flee.

Methodology

How the scores are built

The two dimensions are not measured the same way. Automation rests on expert judge ratings collected since 2018. Generative AI exposure rests on a task-by-task breakdown. The statistical processing of both was entrusted to OPTIMA, an independent research firm.

First method: expert judge rating, for automation

The study began in 2018 on 200 occupations, under the direction of Emeric Lebreton, researcher in human sciences, and was extended in 2021 to nearly 800 occupations. Each occupation was rated on four factors held to predict physical automation: physical demand, attractiveness, complexity and the need for human interaction.

Rating used the judges method. Each occupation was assessed independently by three complementary profiles: an HR director, an occupational psychologist and a recruitment expert. Using three evaluators from different backgrounds limits each one's individual bias. The assessments were then standardised and passed to OPTIMA for statistical processing.

An important reading note. The percentages shown correspond to the occupation's rank in the list, not to a probability in the strict sense. A score of 84 does not mean an 84% chance of being automated, but that the occupation sits very high in the ranking of 778.

Second method: task analysis, for generative AI

The second phase could not rest on an overall rating: generative AI does not affect an occupation, it affects tasks within an occupation. Each occupation was therefore broken down using the major international activity frameworks, notably the detailed O*NET task lists.

Tasks are split into three blocks summing to 100%:

• Manual or physical: out of reach of a generative model.
• Substitutable cognitive: producing and processing text, structured analysis, synthesis, visualisation, code. This block gives the base score.
• Non-substitutable human: personal liability, decision under uncertainty, management, public credibility, arbitration.

The base score is then adjusted, within a maximum range of ten points, to account for liability, the impact of an error on safety or the public, and the need for final human validation.

The variables used in the calculation

Physical demand

Physical, postural and environmental constraints. Rated 1 to 9.

Attractiveness

Interest, social standing and working conditions. Rated 1 to 9.

Complexity

Diversity and cognitive sophistication of tasks. Rated 1 to 9.

Human interaction

Importance of social interaction in the work. Rated 1 to 9.

AI risk

Exposure of cognitive tasks to generative AI. External rating, 0 to 100.

AI risk is treated separately from the other four. It measures a different kind of phenomenon and already has its own scale. Merging it into the same calculation would have blurred two distinct things. The data support that choice: the correlation between physical demand and AI risk is -0.55, meaning physically demanding occupations are precisely the ones generative AI affects least.

The automation index, built by principal component analysis

The first four variables each measure, from a different angle, one latent factor: the degree of physical automation. They are standardised, then submitted to a principal component analysis. The first component explains 58.8% of total variance, the first two 79.2%. The first is therefore enough to carry the index.

The weights are not chosen by hand, they come out of the calculation. On that first component, physical demand loads positively (0.66) while attractiveness (-0.92), complexity (-0.87) and human interaction (-0.57) load negatively. In other words: a demanding, undervalued, simple and low-contact occupation is exposed; a complex, valued and relational one is protected.

The resulting score is rescaled linearly from 0 to 100, where 0 is the least exposed occupation in the sample and 100 the most exposed. Mean 44.7, median 41.3.

The generative AI exposure index

The external score is rescaled to the same 0 to 100 range by min-max transformation. Mean 31.8, median 29.2. The correlation between the two indices is -0.45: the dimensions are related but clearly distinct, which is why they are read separately.

The four categories, defined by quartiles

The thresholds are not arbitrary, they are the quartiles observed on the actual distribution of each index. Each category therefore holds roughly a quarter of the occupations.

• Automation: preserved below 28.4; productivity gains from 28.4 to 41.3; transformed from 41.3 to 59.1; likely to disappear above 59.1.
• Generative AI: preserved below 14.2; productivity gains from 14.2 to 29.2; transformed from 29.2 to 49.2; likely to disappear above 49.2.

Crossing the two classifications produces sixteen risk profiles, each with its own operational conclusion. That conclusion is what appears on every occupation profile.

Margin of uncertainty

The initial ratings involve a degree of judgement. To measure it, a thousand perturbed score sets were generated per occupation, adding Gaussian noise of 5% then 10% around each value, with the index recomputed each time using fixed weights. This is a Monte-Carlo sensitivity analysis, preferred to classic resampling which would have destabilised the weights from one iteration to the next.

At 5% noise, the 95% confidence interval averages 7.5 points wide for automation and 6.2 points for AI, on a 100-point scale. In practice, a gap of three or four points between two occupations is not significant; a gap of twenty points is.

Why two trajectories and not a single score

A weighted composite score was tested, then rejected. It produced an artefact: an occupation highly exposed to AI and little to automation appeared to improve between the 3-5 year and 8-10 year horizons, purely because automation regained weight in the formula. The two risks do not offset each other, they add up. They are therefore projected separately.

• Automation: already-deployed technology, slow and predictable progression. The score rises by 5% at 3-5 years, then 8% at 8-10 years.
• Generative AI: emerging technology with fast deployment. The score rises by 25% at 3-5 years, then 35% at 8-10 years.
• Scores are capped at 100. At the far horizon, automation gains 3.5 points on average while generative AI gains 11.

Points to bear in mind

• The ratings rest on expert assessment rather than field measurement.
• The analysis works at the level of job titles. Two people holding the same job in different companies may have very different exposures.
• The diffusion rates are working assumptions, identical for every occupation. A technological or regulatory break would invalidate them.
• The study measures technical exposure, not a net employment balance. Job creation driven by the same technologies is not modelled.

Anticipate

How to anticipate the way your occupation will change

A high score does not mean an occupation will disappear tomorrow. It signals work content that is about to shift, and therefore a window to act before you are forced to.

What a high score really means

Exposure is measured task by task, not job by job. In an occupation rated "transformed", part of the work moves to the machine and the rest reorganises around what the machine cannot do: deciding under uncertainty, carrying personal liability, sustaining a relationship over time, coordinating people. Anticipating means identifying those areas and moving into them deliberately, while the choice is still yours.

Three possible paths

• Reposition within the same occupation. Shift your activity towards low-exposure tasks and become indispensable where the technology stops. This is the fastest and least costly route.
• Move up in skills. Add an expertise that reshapes the role: running the tools, managing people, compliance, complex client relationships. The occupation stays, its level changes.
• Change occupation. When exposure is high on both dimensions and at every horizon, retraining stops being a distant plan and becomes the defining decision of the next three years.

Choosing between these three requires combining three things: what the market will ask for, what you already know how to do, and what suits you over the long run. That is precisely what a career assessment is for.

The ORIENTACTION career assessment

In France, the bilan de compétences is a scheme governed by the Labour Code. It runs for up to twenty-four hours spread over several weeks, in three phases: a preliminary phase that frames the request, an investigation phase that explores the options, and a conclusions phase that produces an action plan. The summary document belongs to you and stays confidential: your employer has no access to it without your agreement.

At ORIENTACTION, the support is one-to-one, delivered by a consultant from the network, remotely or in person depending on your schedule. It draws on psychometric tests designed in-house: professional values, needs at work, personality preferences and soft skills. Those results, combined with market analysis and the WorkmetriQ projections, are used to build a workable plan rather than a list of ideas.

What we anticipate

An exposure score says what is happening to the occupation. It does not say what is happening to you. Two people in the same post do not lose the same thing, because they do not enjoy the same tasks, are not good at the same ones, and do not build their professional identity on the same ones.

ORIENTACTION's support builds this reading in. The consultant takes the tasks identified by WorkmetriQ, those technology can take over and those that remain, then cross-references them with three dimensions only you know: what you enjoy doing, what you do best, and what you could not lose without ceasing to feel you are practising your occupation.

That cross-reference reveals four very different situations. Exposed tasks you dislike: their automation is a relief, it gives you time back. Exposed tasks you enjoy and that define you: this is where the real risk lies, a felt loss of standing long before any job is cut. Protected tasks you master without realising it: the reserve to make provable. Protected tasks that are enjoyed and identity-forming: the ground to build on.

We therefore separate three levels that are often confused: an activity can be threatened without your value being threatened, and your value can be affected without your job being at risk. An accountant may lose half of their repetitive entries without losing any of what clients actually pay for. A professional may keep their post while watching AI commoditise the very expertise their standing rested on.

This distinction changes the conclusion of the assessment. The question is no longer "will my job disappear?", but "in work where AI does part of the task, what will my value be, and which part of my occupation do I want to preserve?"

Funding

The scheme can be funded through the personal training account, through the employer's skills development plan, or through France Travail depending on your situation. It can be undertaken without informing your employer if it takes place outside working hours.

When to start

The right moment is while you are still in a position to choose: in post, with visibility on your sector and no financial pressure. A path prepared two years ahead can be negotiated; the same path forced on you after a redundancy has to be salvaged.

Discover the ORIENTACTION career assessment

Author of the study

Dr. Emeric Lebreton

Doctor of psychology, founder and president of ORIENTACTION.

Background

Emeric Lebreton completed a doctorate in psychology, then taught for almost ten years at university and in French grandes écoles, notably Audencia and the École nationale des ponts et chaussées. His work focuses on the psychological drivers of career choices: what makes someone stay, move up or leave.

He left teaching to found ORIENTACTION, on the premise that serious career support requires measurable tools rather than interviews based on intuition alone.

ORIENTACTION

The network now brings together more than a thousand consultants specialising in career assessment, accreditation of prior learning, outplacement, business creation and educational guidance. More than one hundred thousand people have been supported using his methods. He chairs the company and still designs the tools the network uses.

The psychometric tests

He designed the test battery used in career assessment: Sens® on professional values, Harmony® on soft skills, Preferences® on personality preferences and Needs® pro on needs at work. Each rests on an identified theoretical model, respectively Schwartz's values, behavioural competencies, Jung's psychological types and Maslow's hierarchy of needs.

Publications

Nearly fifty books, more than 500,000 copies sold and translations into several languages. Among his essays: Révolution Robot (2019), on artificial intelligence and work, La Psy'Action (2018) and Une Histoire de la conscience (2026). He also writes novels.

How WorkmetriQ came about

The question raised in Révolution Robot in 2019 has become, within a few years, the most frequent question in career assessment: "will my job still exist in ten years?". Until now it was answered by intuition. This study answers it occupation by occupation, with a stated method and acknowledged limitations.

Full biography and bibliography