An AI agent for HR handles the high volume tasks of recruitment and people administration. It reads CVs, calls candidates for pre-qualification, writes into the ATS and answers new joiners' internal questions. The hiring decision itself stays with a recruiter.
Artificial intelligence entered French HR teams through payroll and time tracking, not through recruitment. Agents move that starting point. They no longer suggest a piece of text, they act inside the tools. Here are the four use cases that genuinely work in 2026, what they cost, and what the law requires before you connect anything.
What an AI agent for HR actually is
An AI agent for HR is a program that receives a goal, picks its tools and acts inside your systems. The difference with an HR chatbot lies in that ability to act. A chatbot answers the employee, an agent creates the candidate record, sends the invitation and updates the file.
Three levels of automation coexist in an HR team, and they are not equivalent.
- The workflow. A trigger, then fixed steps. An application arrives, its fields get extracted, a row appears in the ATS. Robust and cheap, but rigid.
- The assistant. It drafts a job ad, summarises an interview, rewrites a report. It touches no tool and leaves the handling to the recruiter.
- The agent. It receives a goal and decides what follows. It chases a candidate who has not replied, picks the relevant tool and stops when the exit condition is met.
Most HR teams gain more from two solid workflows than from one ambitious agent. Tandem builds those blocks under the name autopilot, and the level gets chosen on the real volume of the task, not on the demo effect. For the build mechanics, the tutorial on building an AI agent with n8n covers the blocks step by step.
Where French HR teams really stand
French HR teams are investing, but in administration. The SD Worx Research Institute surveyed 5,936 HR decision makers across sixteen European countries between 27 January and 20 February 2026, including 300 in France. The uses cited most often by French respondents are time tracking, at 34%, and payroll, at 32%. Reporting follows at 26%, compliance at 25%, and workforce planning at 24%.
Recruitment does not top that list. That gap is exactly what agents come to fill, and it explains why France ranks 14th on AI maturity in the same study. The main barrier cited by French respondents remains a lack of internal skills, at 32%.
Executives are not asking to be replaced. The Apec-ANDRH barometer published on 2 June 2026 finds that 71% of them believe AI tools will not be able to fully replace professional advice. The French national association of HR directors is chaired by Audrey Richard, HR director of the Canal+ group. The same barometer notes that more than a quarter of large companies have already set an AI usage charter.
France Travail, the public employment service, gives the order of magnitude. LeMagIT detailed the programme on 30 July 2026. Its Match FT Offre tool cuts three days off the average time to fill a vacancy, with a general rollout planned for the last quarter of 2026. Its conversational assistant Chat FT saves an adviser around three hours a week, with daily usage reported at about 70%.
Your candidates already apply with AI agents
CV screening is getting harder for a reason rarely mentioned. Candidates tooled up before recruiters did. I filmed an agent that applies on my behalf, driven from WhatsApp. The setup comes down to a few blocks. A voice or text message triggers an agent that asks for the missing details, pulls a resume in PDF and formulates a search query. Navigation then moves to a tool able to fill in the forms.
Two details matter for a recruiter. The agent extracts the information from the PDF before filling anything in, so CV formatting no longer filters anyone out. And it clears captchas, so an application portal's protections do not slow the volume down.
Cost explains the rest. In the video, each navigation step cost roughly 10 cents, and a full application came to around thirty cents for four minutes of runtime. Skyvern, the tool used, now bills in credits. As of 14 August 2026, its Hobby plan at 29 dollars a month, about 25 euros, opens up around 1,200 actions, which brings one action down to a few cents.
CV screening, what the agent does and what it must not decide
On screening, the most solid public figures come from LinkedIn. In its Hiring Assistant announcement of 3 September 2025, the platform reports three results among its first customers. They reviewed 62% fewer profiles, saved more than 4 hours per role and improved InMail acceptance rates by 69%. The agent is available in English, German and French.
A quote from the same document shows the change of pace. Vincent Mercandetti, Senior Talent Acquisition Partner at Siemens, compares two situations. Where an hour went into sourcing for a single project, he now sources for more than five projects in ten to fifteen minutes.
The red line sits at the next stage. A score ranks, it does not decide. An agent that rules out an application on its own turns screening into automated decision making, which triggers heavy obligations and litigation risk. The habit that holds up is asking the model for the reason behind each gap, then having a human rule on that reason.
| Use case | What the agent handles | What stays human | Cost range |
|---|---|---|---|
| CV screening | Reading, normalising, ranking with reasons | The decision to rule out, and its written reason | A few cents per application |
| Pre-qualification | Outbound call, scripted questions, structured transcript | Reading the transcript and inviting to interview | Under one euro per two minute call |
| ATS data entry | Field mapping, attachments, status updates | Sample based checking | A few cents per file |
| Onboarding | Sourced answers, procedures, introductions | Managerial follow up in the first weeks | Around 29 euros per user per month |
Pre-qualification calls, recruitment's next turn
The pre-qualification call is the fastest moving HR use case. I described it in my edition on AI voice agents as the next big turn. A voice agent replaces the first screening calls, asks everyone the same questions, assesses the answers and feeds the HR tools. In staffing agencies, the feedback covers three points, more qualified candidates, less wasted time and better quality in the profiles passed on.
One counter-intuitive argument favours the machine on this exercise. A voice agent does not get tired, does not get irritated on the twentieth call and listens to the answer all the way through. That consistency becomes an advantage in a pre-qualification interview, where equal treatment between candidates is precisely the point.
The limit is clear. A voice agent qualifies, it does not assess a complex skill and does not replace a substantive interview. The guide on what AI voice agents really deliver covers the verticals where the technology holds up and those where it still disappoints, with cost per minute.
ATS data entry, where agents beat RPA
Manual data entry remains the hidden cost of recruitment. Copying details from a file into an ATS, a partner site or an administrative portal takes hours every week. Classic RPA scripts promised that automation fifteen years ago, without ever holding up over time.
The reason is simple. A rigid script targets fields by their position on the page, so the slightest interface change breaks it. On an in-house applicant tracking tool, that fragility cost more in maintenance than it saved in time. A web agent detects fields visually, adapts its behaviour and carries on without rewriting the scenario.
One case from an engagement shows the shift. An HR team wanted to pre-fill files on several partner sites from an internal tool. No shared format, each site with its own fields and its own steps. The agent reads the record, works out which detail goes where, and fills in the right place.
Onboarding, the most profitable HR use case
Onboarding is the most underestimated HR opportunity, and the fastest to pay back. A new joiner spends their first weeks hunting for information scattered across a drive, an intranet, a CRM and the memory of three colleagues. An internal assistant connected to those sources answers while citing the source document and its last update.
On Tandem deployments, the time spent finding a piece of information divides by ten on average, from several minutes to a few dozen seconds. The cost sits around 29 euros per user per month on an off the shelf solution. It pays back as soon as one employee saves an hour a month, meaning a quarter of an hour a week.
The second onboarding use needs even less tooling. An assistant's voice mode lets someone practise against a simulated counterpart. Load the company documents and the common objections, and the assistant plays the persona then delivers its analysis at the end. I have used this setup for onboarding sales teams since 2024.
I detailed the setup in a post on sales role play in voice mode. The principle transfers directly to a mock annual review or to preparing a first client call. This kind of practice spreads far better when it is supported, which is what AI literacy programmes cover at Tandem.
What the CNIL and the AI Act require before you deploy
The legal framework is the first topic to settle, not the last. The CNIL announced on 3 April 2026 that recruitment is one of its three priority audit themes for 2026. Three points will be checked, automated decision making systems, information given to candidates, and retention periods. Large companies and recruitment firms are the priority targets, because of the volume of applications they process.
Two rules can be remembered without legal expertise. An unsuccessful candidate's data is kept for two years at most from the last contact, unless they explicitly agree to join a talent pool. And the minimisation principle limits collection to information directly related to the job on offer.
The European AI regulation classes application screening and candidate assessment as high risk, in Annex III under employment. That regime was due to apply on 2 August 2026. The European Council and Parliament reached an agreement on 7 May 2026 within the digital simplification package known as Omnibus VII. That agreement pushes the deadline to 2 December 2027 for stand-alone systems, and to 2 August 2028 for those embedded in a product. The Council signed the text off in late June 2026.
Should you hand recruitment to an AI agent
Yes on volume, no on the decision. An AI agent for HR handles CV reading, the pre-qualification call, ATS data entry and answers to internal questions. Those four tasks are repetitive, measurable, and their errors get caught by a simple check. Assessing a skill, choosing between two close profiles and delivering a rejection stay with the recruiter.
The starting order matters more than the choice of tool. Count the volume over one month first, the applications received, the hours spent screening and the internal questions that keep coming back. Then install a single flow end to end, the highest volume one, rather than three unfinished pilots. That is the work an AI audit covers at Tandem, a few weeks alongside the teams to size the opportunities. The same method applies in service companies and reappears, with few changes, in what we install in accounting firms.
A tool impresses. A flow in production changes the team.
