Will AI eliminate 5 million jobs in France? What the shocking study reveals

On March 18, 2026, a study jointly published by Coface and the Observatoire des Emplois Menacés (OEM) sent shockwaves through the French economic press.

Its matter-of-fact title conceals a figure that demands serious thought: 5 million jobs could be at risk from AI in France within the next 2 to 5 years.

Should we be alarmed or keep things in perspective?

This study deserves better than a quick summary or a swift dismissal: it deserves a critical reading, grounded in its actual methodology, its blind spots, and the data from other studies that help put it in context.

The study in 60 seconds: key figures

Coface, a global leader in trade credit insurance, and the OEM, founded by economist Axelle Arquié of CEPII, published in March 2026 the most granular mapping to date of French employment’s exposure to AI.

The methodology deliberately moves away from classic sector-level approaches: each profession is broken down into elementary tasks, and each task is assigned an automatability score based on AI’s current and projected capabilities.

The result is a two-figure picture that frames the entire debate: 3.8% of jobs are currently vulnerable to generative AI, and that rate jumps to 16.3% within a 2-to-5-year horizon, roughly 5 million positions out of a workforce of 30 million.

As Aurélien Duthoit, senior economist at Coface and co-author of the study, explains: “This estimate is based on what French companies are doing today: deploying chatbots and using language models in various ways.”

These are not two contradictory figures: they are two stages of the same acceleration, separated by the arrival of agentic AI.

3.8% today: a deceptive calm

The first figure, 3.8%, reflects the reality of generative AI as it is deployed today in French companies.

Chatbots, content generation tools, and LLMs embedded in productivity software. These make up the bulk of France’s operational AI landscape in 2025.

That figure may seem reassuring, but it masks adoption rates that are still very low: according to BSI IA4Business data from October 2025, only 7% of French employees use AI daily at work, compared to 26% globally and 61% in China.

The current calm does not reflect structural resilience in French employment against automation: it reflects an adoption gap that will close, one that is already compressing the time available to respond.

16.3% within 2 to 5 years: agentic AI changes everything

The jump from 3.8% to 16.3% cannot be explained by a simple linear extrapolation of today’s generative AI.

It rests on a specific technological bet: the rise of agentic AI.

Where generative AI assists humans with discrete tasks, an AI agent can receive a high-level objective and execute it end-to-end: analyzing, planning, deciding, acting, and evaluating results.

Axelle Arquié offers a concrete example: asking an agent to find the best product to buy, then completing the purchase autonomously, without any further human input.

In employment terms, this means that one professional in eight holds a role where at least 30% of tasks are fully automatable by AI agents.

Adoption remains embryonic for now: by the end of 2025, only one third of large companies had launched agentic AI projects.

But the momentum is there, and organizational barriers are gradually giving way as the first return-on-investment results come in.

Agentic AI does not replace a single tool: it replaces an entire workflow. That qualitative leap alone explains the staggering gap between today’s 3.8% and tomorrow’s 16.3%.

The surprise: high earners face the greatest risk

This is the Coface/OEM study’s most striking reversal: unlike previous waves of automation, which targeted routine and low-skilled jobs first, AI poses the greatest threat to high earners.

The top 10% of salaries in France faces an exposure rate of 22.1%, compared to significantly lower levels for the bottom deciles.

All three top salary deciles exceed the 20% risk threshold.

The logic is counterintuitive but coherent: generative AI excels precisely at the cognitive, analytical, and communication-based tasks that sit at the heart of the work done by executives, lawyers, architects, engineers, and consultants.

Economist Gregory Derville captures the generational stakes of this reversal: “In the coming years, hundreds of thousands of young people will face deep frustration, having invested time, energy, and money to earn a master’s degree or a grande école diploma, only to discover there are virtually no career openings at that level of qualification.”

The sectors on the front line

The study’s sector-by-sector breakdown maps out exactly which professions are under the most pressure.

Architecture and engineering top the list with 26.9% exposure, driven by AI’s aptitude for spatial analysis, design optimization, and technical document generation.

Computing and mathematics follow with 24.9% overall exposure, and as high as 31% for IT alone: the sector that builds AI is seeing part of its own workforce threatened by the tools it creates.

Administrative and clerical roles reach 23.8%: document management, data entry, calendar management, and email processing are squarely in the crosshairs of AI agents.

Creative and cultural professions hit the same threshold of 23.8%, a reality already playing out in the field: freelance translators report losing a third of their revenue between 2023 and 2025.

Legal professions face 21.6% exposure, and what Aurélien Duthoit compares to “replacing a junior lawyer”: legal research, contract drafting, and regulatory monitoring are all directly in scope.

The relatively protected sectors

The study also identifies professions whose task structure holds up better against automation.

Cleaning and maintenance show just 5.4% exposure: physical robotics in non-standardized environments remains an unsolved technical challenge.

Agriculture, fishing, and forestry reach 7.9%, shielded by biological diversity and weather-dependent decision-making that complicates automation.

Hospitality and food service limits exposure to 7.5%: the human experience, real-time reading of preferences, and cultural adaptation remain difficult to automate at an economically viable cost.

The irony of AI-driven automation: so-called “unskilled” jobs are often the hardest to replace, while the degrees long held up as shields have become targets.

What the study doesn’t say

At risk does not mean eliminated

The Coface/OEM study’s central methodological limitation is also the one most consistently overlooked in media coverage.

“At risk” refers to partial or total exposure of a role’s tasks to automation, not the elimination of the role itself.

A lawyer whose tasks are 30% automatable becomes more productive, not necessarily replaceable.

The line between partial automation and job destruction depends on organizational, regulatory, and cultural factors that the study makes no claim to model.

Agentic AI: between promise and practical reality

The 16.3% scenario assumes rapid and widespread adoption of agentic AI.

Axelle Arquié herself points to real-world obstacles: “You can’t hand over temperature regulation in a chemical plant to an AI without being able to verify its reasoning. Right now, AI is a black box that even labs are still trying to understand.”

Constraints around governance, legal liability, and explainability are slowing the adoption of AI agents in high-regulatory-impact processes, which pushes the timeline of the pessimistic scenario further out.

AI washing distorts the data

Announcements of layoffs “caused by AI” deserve a critical eye.

The October 2025 Challenger Report counted 31,000 layoffs attributed to AI in the United States in a single month, a striking figure that often conceals restructurings planned for entirely different reasons.

Companies like Capgemini and IBM have publicized AI-generated savings while keeping headcount nearly unchanged: this is what is known as AI washing: dressing up ordinary business decisions in an AI narrative.

Economist Laura Verdugo of the OFCE notes that AI advocates tend to overstate its impact on employment, often to attract investors. A rigorous reading of the Coface/OEM study cannot ignore this structural bias.

Comparison with other studies

The Coface/OEM study sits within a body of international research whose conclusions vary depending on methodology and scope.

Study	Key figure	Scope
Coface/OEM (2026)	16.3% of French jobs at risk within 2-5 years	France, elementary tasks
OECD	27% of jobs highly automatable	OECD countries, by profession
Goldman Sachs	300 million jobs exposed	Global, generative AI
PwC AI Jobs Barometer	+252% productivity in AI-intensive sectors	International, productivity gains

The gap between OECD (27%) and Coface/OEM (16.3%) is partly explained by differing weights assigned to economic and organizational feasibility criteria.

The OECD measures theoretical technical automatability; Coface/OEM factors in the reality of how AI tools are actually deployed in French companies.

As for the PwC figure of +252%, it points to the other side of the picture: sectors with heavy AI adoption are seeing productivity surge, which can sustain redefined jobs rather than eliminate them.

To understand the long-term dynamics between AI and human work, economists’ projections diverge significantly depending on the public policy assumptions involved.

What to do if your profession is on the list

For employees

The first concrete step is to map your own task portfolio: identify which of your daily responsibilities can already be handled by a quality LLM, and which still require an irreplaceable human presence.

Then invest in skills that work alongside AI: critical thinking, complex client relationship management, AI system oversight, and deep sector expertise.

The French government’s Osez l’IA program aims to train 15 million people in AI usage: explore the available resources, but don’t stop at generalist training.

For executives and HR leaders

Your teams’ exposure is not inevitable: it depends largely on your AI implementation choices.

The question to ask is not “which roles should we automate?” but “which tasks can we free up so our teams create more value?“

The AI Act, taking effect in August 2026, imposes regulatory constraints on high-risk AI systems, including in HR contexts. Get ahead of compliance to avoid unpleasant surprises.

The impact on social protection

One dimension underplayed in media coverage of the Coface/OEM study deserves attention: if high earners face the greatest risk, they are also the largest contributors to social security, and potentially the first to exit the labor market.

France’s social protection system, funded primarily through employee and employer contributions, depends on a base of workers in stable employment.

Mass unemployment concentrated in the upper income brackets would create a double fiscal shock: a rise in unemployment benefits on one side, and a collapse in contribution revenues on the other.

Axelle Arquié raises an additional dimension: if young graduates anticipate the devaluation of their qualifications, personal investment in higher education could lose its individual economic rationale, with profound consequences for the funding and appeal of France’s grandes écoles.

The 5 million at-risk jobs are not an apocalyptic projection: they are 5 million individual situations that will require collective responses we have not yet built.

Conclusion

The Coface/OEM study brings something rare to the AI/employment debate: methodological rigor that resists both catastrophism and denial.

Its figures, 3.8% today and 16.3% tomorrow, are neither verdicts nor certainties. They are conditional scenarios that hinge on the actual pace of agentic AI adoption, on corporate organizational choices, and on public policy support measures.

What the study says without ambiguity is that response windows are shrinking, and that the most exposed profiles are not the ones anyone expected.

The detailed study data and official resources are available on the Coface France website.

FAQ

What is the Coface/OEM study on AI and employment?

It is a study published in March 2026 by Coface and the Observatoire des Emplois Menacés, which analyzes the exposure of French jobs to AI by breaking each profession down into elementary tasks.

What are the key figures from the Coface/OEM study?

3.8% of French jobs are currently exposed to generative AI, and that rate could reach 16.3% within 2 to 5 years with the rise of agentic AI, roughly 5 million positions.

Why are high earners more at risk than low earners?

Generative AI excels at the cognitive, analytical, and writing-based tasks that form the core of skilled professions, while manual roles in variable physical environments remain technically difficult to automate.

Which sectors are most exposed according to the Coface/OEM study?

Architecture and engineering (26.9%), computing (31% for IT alone), administrative and clerical roles (23.8%), creative professions (23.8%), and legal professions (21.6%).

Which sectors are least at risk from AI?

Cleaning and maintenance (5.4%), hospitality and food service (7.5%), and agriculture, fishing, and forestry (7.9%).

What is the difference between a job being “at risk” and a job being “eliminated”?

“At risk” means that a significant share of a role’s tasks are automatable: it does not mean the role will disappear. An employee’s responsibilities can be redefined and their productivity increased without them losing their job.

What is agentic AI and why does it accelerate the threat to employment?

An AI agent can receive a high-level objective and execute it autonomously, from analysis through to decision, without human intervention at each step. That shift from point-in-time assistance to full-workflow automation is what multiplies its impact on employment.

How does the Coface/OEM study compare to the OECD study?

The OECD estimates 27% of jobs in France are highly automatable, compared to Coface/OEM’s 16.3% within 2-5 years. The gap reflects more conservative assumptions about the actual pace of AI tool deployment in French companies.

What does France’s Osez l’IA program involve?

The Osez l’IA initiative aims to train 15 million people in AI usage; it is presented as the government’s primary response to AI’s impact on French employment, though its scope remains largely generalist.

What impact could AI have on France’s social protection system?

If the threat is concentrated among high earners, the largest contributors to social security, their exit from the labor market would trigger a double fiscal shock: rising unemployment benefit costs alongside a collapse in contribution revenues.