A first numerical signal is shown by the AIA and the recruitments in France: around 25,000 fewer job vacancies in the third quarter of 2026, according to a Coface study published on 17 November. This result relates to advertisements in France Travail. It does not constitute a count of dismissals or evidence that 25,000 employees have been replaced by software.
Economists Axelle Arquié and Aurélien Duthoit bring the exhibition of trades closer to the automation of the evolution of advertisements. Their estimate, relayed by a news agency, puts the effect on the recruitment needs expressed by companies. This is a different measure of the number of people already on duty.
The distinction changes the reading of the figure. A company may waive a planned hiring without removing an existing job. An advertisement may also be withdrawn, renewed or met multiple needs. Moving directly from advertisements to jobless would confuse several stages of recruitment.
IA and recruitment: what do the 25,000 offers measure?
The AI-LERT indicator, developed by Coface and the Observatory for Threatened and Emerging Jobs, uses the France Travail offers from the first quarter of 2021 to the third quarter of 2026. The method described in the French press compares the trajectories of professions according to their exposure, around the launch of ChatGPT in November 2022.
It is therefore not a question of finding 25,000 letters from employers assigning cancellation to artificial intelligence. The approach seeks a statistical gap between observed developments. The figure refers to an estimate of fewer advertised recruitments within the scope under review, not an administrative register of abolitions.
The detailed result relayed by a news agency combines a 10% exposure of tasks with a 12.4% decrease in third quarter announcements. This rate is not a 12.4% decrease in all French jobs. Nor does it mean that each company would have reduced its workforce in that proportion.
| Indicator | What it measures | What it doesn’t measure |
|---|---|---|
| 25 000 less estimated offers | A lack of T3 2026 announcements attributed to AI by the authors | 25,000 redundancies recorded |
| Task exposure | Automation potential | Effective replacement |
| Hires | Employment | Total number of employees |
A technical exhibition is not a convicted post
The mapping published earlier by Coface and OEM helps to understand this approach. It breaks down 923 occupations into tasks and then basic actions. It evaluates their automation potential, taking into account in particular their importance and frequency.
This study explicitly distinguishes between technical capacity and net job destruction. It does not take into account, to its gross extent, all changes in demand, new tasks or obstacles to deployment. Computer science, administrative activities, finance and certain creative professions are among the families at risk. There are other constraints to physical or highly relational activities.
Automation of part of the work can produce several effects. At constant activity, a company can reduce its labour needs. If it lowers its prices or sells more, it can also use the time saved to increase its production. The potential of technology is therefore not sufficient to determine the final balance of jobs.
The Bank of France also observed a slowdown in
A separate research, presented by the Bank of France on 8 October, provides additional insights. It follows the hirings and purposes of CDI between January 2019 and March 2026, rather than announcements. Among the 21-29 year-olds, hirings for the most exposed occupations were about 28% below their November 2022 level. The decline is 8% in the least exposed occupations.
This comparison does not show that the AI alone explains the 20 points of difference. Researchers point out that user companies are not randomly selected. Their survey measures the use declared at the beginning of 2026, while the differences in recruitment begin before.
Contract purposes are also decreasing. In firms reporting intensive use of AI, the authors did not observe a statistically clear difference in total employment. The signal is read more in staff turnover and new hires than in an immediate downsizing.
The figure is not enough to isolate all causality
The two works must not be added together. Advertising and hiring are not two separate jobs. Their periods, populations and exposure measures also differ. Their rapprochement makes it possible to examine the recruitment channel, not to make a national total of jobs destroyed.
To measure a causal effect, the comparison must distinguish technology from other changes in occupations. A drop in orders, a sector reversal or a change in recruitment channels can also reduce advertisements. Access to detailed protocol, robustness tests and statistical uncertainties remains necessary to accurately assess the Coface estimate.
Nor should the quarterly result be mechanically transformed into an annual balance sheet. Multiply 25,000 by four describes a constant rate of extrapolation, not four actually observed quarters. Aggravation, stabilization or rebound of recruitment would alter this calculation.
The next point to follow will therefore be the persistence of this gap in offers, then its translation into contracts actually signed. Measurement of ads can detect early change. Employment and workforce data will be used to verify how far it is transmitted to the labour market.





Comments