10 Key AI Deployment Metrics at France Travail

Ten years already since “The Right Apprenticeship” emerged at France Travail, following in the wake of “The Right Job.”

These services each deploy an algorithm that analyzes past hires to forecast future ones. They were among the first AI use cases within the public institution.

Since then, two programs have taken shape. First, “Intelligence Employment,” rolled out in 2019 and 2022. It led to the creation of a technological platform based on open-source technology and a dedicated team within the IT department. Next, “Data AI,” launched in 2024. It is driven by generative AI, stronger data governance, and the broadening of France Travail’s missions under the full-employment law.

The Court of Accounts has looked into how AI was deployed by the public employment agency over 2017-2025. Here are some elements drawn from its report.

87 Use Cases

The Court of Accounts highlighted 87 use cases deployed or tested during that period.

Among them, 27 are used at a large scale. 16 are in testing. 25 are in the design phase. 17 were abandoned during testing or after deployment. Several are variations of the same tool, notably the ChatFT (generalist chatbot).

The bulk of these 87 use cases — 61, including 26 deployed — have direct beneficiaries only among France Travail staff. According to the Court of Accounts, this illustrates a deliberate strategy to first foster an internal AI culture.

The main use cases where job seekers are direct beneficiaries aim to:

• Make it easier to complete their profile in the personal area
• Recommend occupations based on skill sets
• Automatically read uploaded documents and extract information

These use cases show relatively positive results. Among them, the automatic CV analysis, used by 75% of job seekers.

Of the six use cases benefiting employers, three were abandoned in 2017. One addressed the predictive analysis of the attractiveness of job postings.
The two currently deployed are:

• Forecast the time to fill a posting within 30 days (little used)
• Present, on a public site, general employment data at the territorial level (little added value for recruitment projects)

9% Daily AI Users

France Travail employs about 54,000 staff.

Of the 34,945 respondents to an internal survey conducted in March 2025, 9% reported using the AI tools every day. 18% said they use them several times a week. 39% stated they do not use them.

ChatGPT has been accessible to all personnel since November 2024. By the end of June 2025, 37,600 agents had used it at least once. That month, 17,400 had used it on at least three separate days.
Towards the end of the testing phase, an internal study of around 700 conversations showed that the main use was composing email responses (51%), far ahead of seeking general information (10%).

€108 Million in Costs

Taking a conservative estimate of development costs, France Travail mobilized €93 million for AI between 2017 and 2024.

The “other unallocated expenditures” correspond to the budgets for a project on “Emotion Recognition.”

Adding these figures yields the €93 million cited above. When adding the projected €15 million for 2025, the total cost of AI development over the period in question reaches €108 million. The Court of Accounts compares this amount to the €66 million the Ministry of Economy and Finance invested in 2015-2023 and explains the difference by a greater number of direct user-facing relationships.

€120 Million in Efficiency Gains

These gains are calculated using a high-end estimate of benefits accrued since 2017. They include three direct, expected gains:

• Use-case from the “Intelligence Employment” program: 205 FTEs per year starting in 2023
• “Upload Simplifié” service (document recognition): 350 FTEs/year starting in 2024
• MatchFT service (prequalification of profiles via SMS exchange with an AI): 100 FTE in 2025

In total, over the entire studied period, 1,415 FTEs are considered potentially value-adding, with an estimated value of €85 million.

Indirect gains tied to avoided workloads for counselors would amount to 375 FTEs. On one hand, for automatic CV analysis, 27 FTEs in 2023 and 48 per year from 2024. On the other, 84 FTEs per year from 2023 for the “Lego” service, which identifies illegal job offers and prevents their dissemination. Taken together, these would be worth €23 million.

Additionally, there are avoided costs from replacing software for certain use cases, foremost among them the automatic document recognition with “Upload Simplifié.”

€14.4 Million Overrun for “Intelligence Employment”

The planned budget for this program was €49.5 million.
Actual expenditure amounted to €63.9 million (€29 million in payroll, €33.9 million in operating expenses, and €1 million in investments).

France Travail justifies this overrun by a 10-month extension of the project duration, mainly due to:

• The Covid crisis
• The development of initially unforeseen use cases (Lego, CV analysis, assistance with information searches on Pôle emploi sites, etc.)
• The accounting of expenses tied to common technical capabilities shared with other AI solution develop­ments not integrated into the program

As early as January 2020, before the Covid episode, the Dinum had warned about the amount of intellectual services expenditures contemplated (€22.7 M). It deemed this “overdimensioned for an exploratory approach whose ROI is not guaranteed.”
The external services ultimately cost €33.9 M.

The Dinum had also anticipated, rightly, the uncertain potential for adoption by advisers of the use case “Automated email management and virtual assistant.” Pôle emploi had resisted halting it, arguing that efficiency gains and user satisfaction could be observed in the short term.

205 FTE Gained with “Intelligence Employment”

In 2018, Pôle emploi forecasted an annual gain of 164 FTE by 2022 (87 through “Contact via mail” and 77 via using a chatbot).

The gain ultimately reached 205 FTE (107 via “Contact via mail,” the rest via Lego). The target was therefore exceeded; but in hindsight, it appears not particularly ambitious. The Court of Accounts points to a ROI notably weaker than that of a selection of AI projects from the Ministry of Economy and Finance.

These gains did not translate into a net reduction in staff but rather into intra-post redeployments.

From 18 to 4 Months to Deploy a Use Case

Before the first program, deploying a use case took 18 months. It now takes 4. At the same time, reliance on external contractors fell from 60% of total expenditures in 2019-2020 to 40% in 2021-2022.

Of the six main use cases implemented during the “Intelligence Employment” program, three were mainly aimed at easing the work of advisers. A misalignment with the audiences targeted would widen with the “Data AI” program.

Before the launch of the “Intelligence Employment” program, France Travail identified more than 80 potential use cases. 90% proved unsuitable, largely due to data issues, integration with the information system, or overestimation of value.

30 Minutes Less to Fill a Profile

The use cases implemented under the first program yielded relatively modest efficiency gains. A few minutes per day for email management, for instance.

The value added comes primarily from the service delivered. The automatic CV analysis reduces the time a job seeker needs to fill out their skills profile from 45 to 15 minutes.

779 FTE to be Freed Between 2025 and 2027

The transformation work to meet France Travail’s broadened missions is formalized in a program, extended into a three-year efficiency plan (2025-2027).

Last reports indicate this plan aims to secure at least €3,192 FTE in gains. The AI must contribute 779 FTE.
In April 2025, the plan envisioned 822 FTE, of which 78% would come from three use cases:

• Writing assistance via ChatGPT (226 FTE)
• Interview preparation via Neo, a dossier information search engine for job seekers (241 FTE)
• More effortless feeding of the GED via Panoptes (157 FTE), of which “Upload Simplifié” is the online version of France Travail (in agencies, this is called “Scanner,” while the government site uses “Scanlab”).

Ethical Analysis for 18 Use Cases

Respect for commitments in the ethics charter published in April 2022 is not guaranteed, notes the Court of Accounts. Only 18 use cases have had a formal initial ethical analysis conducted.

The use of AI takes place within the broader digitization of the relationship with France Travail. Between 2017 and 2024, the number of visits to agencies fell by 42%, while the volume of emails rose by 72%.