An AI agent for your HR team, from screening applications to everyday questions.

An HR AI agent is a program that reads your HR tools, breaks a task into steps and carries them out in the HRIS, the ATS, email or calendar. A chatbot answers a question; the agent prepares what comes next: a reasoned shortlist, a leave request, a draft email. A person approves before anything affects a candidate or employee.

Screening applications with AI falls under point 4(a) of Annex III of the AI Act, which covers analysing and filtering job applications. The obligations for high-risk systems will apply from 2 December 2027. The agent can read every resume and suggest a reasoned shortlist; the recruiter approves it, the manager and HR choose.

Our guide on AI resume screening and CNIL inspections describes a compliant process. The other use cases (onboarding, employee questions, training) are shown with a sample conversation in the list of use cases at the top of this page.

The same boundary applies to skills management. Forecasting hires and skill gaps at team level stays outside Annex III. An individual "flight risk" score would move the tool into Annex III of the AI Act.

Leave balance, health cover, expense claims: the same questions come back every week. An AI assistant receives them in Slack, Teams, the HR mailbox or the intranet, looks for the answer in your documents, reads the employee's balance in the HRIS if needed and prepares the request. Health, disputes and harassment go to a person.

The assistant first needs to know who is writing. It identifies the employee through their company account before it reads any personal data.

Answering from your documents

The assistant reads your convention collective at the moment the question is asked. The method is called RAG (retrieval-augmented generation): your documents are split into passages, the assistant finds the few passages that answer the question and writes the answer from those passages alone, with the document reference. When no passage answers, it says so; test that instruction before launch. Our article on RAG for business explains the mechanics.

Four families of documents do most of the work:

• Convention collective and company agreements: leave for family events, seniority, probation period, agreed bonuses. When a company agreement sets a point differently from the sector agreement, index both with a written priority rule in the assistant's instructions.
• Internal rules and staff notices: working hours, remote work, use of tools, access to the premises.
• Health and provident insurance notice: benefits, dependants, opt-out cases, continued cover after leaving.
• Expense policy: caps per expense type, required receipts, filing deadlines, mileage scale.

Set access rights at the source. According to Notion's documentation, an integration sees only the pages shared with it. On the SharePoint side, Microsoft Graph "Selected" permissions restrict an app to one site, list or file. The neighbouring "Payroll" or "Litigation" folder stays out of the assistant's reach.

Before indexing, keep a single version of each document, dated, with an owner. Connected to a messy Drive, an assistant answers confidently from an outdated document.

The law changes too. Since French law no. 2024-364 of 22 April 2024, an employee on non-work-related sick leave earns 2 working days of paid leave per month, up to 24 per reference period (Article L3141-5-1 of the Labour Code; service-public.gouv.fr page on paid leave in the private sector, checked on 30 April 2026). Remove from the index any older internal note that says otherwise.

What Lucca, PayFit, Factorial, Eurecia and Workday open up through their APIs

To answer "you have 12 days left", the assistant reads the employee's balance in the HRIS through an API. This is the step where projects most often stall. According to each vendor's public documentation, read by Lumyniq on 3 November 2026, reading a balance is direct with some and missing or restricted with others.

Two access settings follow from the table. At Factorial, an API key generated in settings grants access to everything: for a conversational assistant, use OAuth 2.0 with limited scopes. At Eurecia, requests run with the rights of the user whose record generated the token: create a technical account limited to what is strictly needed.

At Workday, the REST directory loads in JavaScript and we could not read its detail. Have your Workday administrator confirm the available operations in your tenant.

When the HRIS does not return the balance through its API, the assistant gives the rule and a direct link to the balance page. When it can read it, it takes the value computed by the HRIS without recomputing it: working days or business days, split leave, carry-over, RTT, the whole rule lives in the payroll settings.

The assistant queries the HRIS for the employee who asks and returns only their data. If the vendor offers only a key with global access, the assistant's code does the filtering: test that filter like a security feature.

Starting a request: the employee confirms, the manager approves

The assistant can prepare a leave request, create a draft expense claim or send the health insurance opt-out form. The employee confirms, and approval stays with the manager or HR, in the usual HRIS workflow. Approving, rejecting, changing a balance, triggering a reimbursement: these actions stay out of the assistant's reach.

The Lucca and Factorial APIs technically allow a request to be approved. Do not give that right to the assistant's account: an approval is a decision about an employee's situation, and it would change how the system is classified under the AI Act.

Health, disability, pregnancy, a dispute with a manager, harassment, a disciplinary matter, termination, a contested figure: on these topics, the assistant says whom to contact, opens a restricted-access ticket and gives a response time, without going into the substance of the case.

Someone who reports harassment gets human contacts at once: the HR team, the referent on sexual harassment and sexist behaviour that every company with at least 250 employees must appoint (Article L1153-5-1 of the Labour Code), the CSE referent, the occupational health service. The employer owes its staff a duty of safety (Article L4121-1): the report must reach someone who can act, quickly.

A medical diagnosis is health data, a special category under Article 9 of the GDPR. The CNIL HR referential recalls that the employer processes it only for its obligations in employment and social security law.

The Annex III boundary of the AI Act

An assistant that informs employees, reads their balances and passes their requests on does not fall within Annex III of the AI Act. Here is point 4:

"4. Employment, workers' management and access to self-employment: (a) AI systems intended to be used for the recruitment or selection of natural persons, in particular to place targeted job advertisements, to analyse and filter job applications, and to evaluate candidates; (b) AI systems intended to be used to make decisions affecting terms of work-related relationships, the promotion or termination of work-related contractual relationships, to allocate tasks based on individual behaviour or personal traits or characteristics or to monitor and evaluate the performance and behaviour of persons in such relationships."

The assistant described here does not recruit or select anyone (point a). It makes no decision on terms of work, promotion or termination, allocates no tasks, and neither monitors nor evaluates performance (point b). This is our reading of the text: the Commission has published no opinion on this specific case. Going by the wording of point (b), four changes could tip it over:

• approving or rejecting leave or expense claims on its own, based on criteria;
• producing named statistics on the questions asked, for instance spotting who asks about negotiated termination;
• feeding annual reviews or a "flight risk" score from conversations;
• allocating tasks or schedules based on employees' profile or behaviour.

If one of these uses appears, the obligations for high-risk systems will apply from 2 December 2027, the date set by Regulation (EU) 2026/1744, known as the Digital Omnibus, as shown in Article 113 on the Commission's AI Act Service Desk. Article 26(7) will then require the employer to inform workers' representatives and the affected employees before putting the system into service.

Outside high-risk, Article 50(1) has applied since 2 August 2026: people must be informed that they are interacting with an AI, unless this is obvious. A greeting line is enough: "I am the HR assistant, an artificial intelligence program. I answer from the company's documents and pass to the HR team whatever needs a person." The detail is in our article on the AI agent transparency obligation.

Since 2 February 2025, Article 5(1)(f) has prohibited using AI to infer the emotions of a person in the workplace, except for medical or safety reasons. Add no module that infers employees' mood or stress.

During the first week, the onboarding agent sends each document on its due date, prepares the access requests the manager approves, and answers the new hire's questions from your documents, citing the source. The contract, pay, health and any question with no written answer go to a person on the HR team.

The sample conversation is shown in the list of use cases at the top of this page. This section covers what sits behind it: the documents, the access, the questions and the hand-offs.

The documents French law requires, and their deadlines

Since 1 November 2023, an employer in France must give each employee one or more written documents setting out the main information about the employment relationship (Article L1221-5-1 of the French Labour Code). Article R1221-35 sets two deadlines, counted from the hiring date: 7 calendar days for part of the information, one month for the rest.

The agent drafts none of these documents. The HR team prepares the templates and has them approved. The agent reads the hiring date in the HRIS, works out the deadlines and sends each document through the agreed channel. If a document has not gone out the day before its deadline, it alerts the HR person in charge of the file.

For the occupational health information and prevention visit, the agent reminds the HR team to book it with the occupational health service. It receives and reads no medical information.

Access: the agent asks, a person approves

On day one, a new hire needs an email account, the team's tools and sometimes shared folders. The agent prepares the request from an access list per role, written jointly by HR and IT. The manager approves the list and the administrator creates the accounts.

In the agents we build, the agent's account has no administration rights on the company directory: it opens a ticket and creates no users. Access outside the role's list, to the payroll folder or to client data for example, also needs the approval of whoever owns that data.

On the start date, the agent checks that every approved access is open and chases IT for any that are missing. The new hire can see the status of each request.

First-week questions, with the document cited

New hires ask much the same questions from one hire to the next. The agent answers them with the method of the HR assistant described in section 04: it finds the passage that answers, writes the reply from that passage alone and cites the document, its version and the article or page. The new hire opens the source in one click. Our article on RAG applied to internal documents explains how the agent finds those passages.

When no passage answers, the agent says so and passes the question to the HR contact, word for word. Test that instruction before launch, with questions whose answer is in no document.

What the agent hands to a person

Some first-week questions touch the contract or the employee's personal situation. The agent says it is passing the question on, opens a ticket for the HR contact and gives a response time. This covers questions on pay, the probation period or a contract clause, a mistake on the first payslip, health, disability or a workstation adjustment, a disagreement with the manager, and a document left unsigned or access still closed after the planned date.

The manager keeps the welcome itself: introductions to the team, first objectives, the end-of-week check-in. The agent sends them the week's schedule and a reminder the day before each meeting.

Throughout the probation period, the new hire's conversations with the agent are not used to assess them. Monitoring and evaluating an employee's behaviour falls under point 4(b) of Annex III of the AI Act (section 04). An onboarding agent that reads the HRIS is also automated personnel management processing: in France, the CSE is informed before it goes live (section 06).

In a French company with at least 50 employees, consult the CSE (comité social et économique, the elected works council) before the agent goes live, pilot phase included, with a written file on its purpose, its data and who sees what. Absent an agreement, the CSE has one month to give its opinion, two if it calls in an expert. Build that period into the project plan.

The legal basis is Article L2312-8, II, 4° of the French Labour Code: the CSE is informed and consulted on "the introduction of new technologies" and on any significant change affecting health, safety or working conditions. Since February 2025, at least four interim orders have applied it to AI tools, including one on an HR chatbot.

Article L2312-38 requires prior information of the CSE on automated personnel management processing and its changes: an assistant that reads the HRIS falls under this case. The same article requires information and consultation before any decision to deploy means or techniques that allow employee activity to be monitored. The assistant falls under it if anyone can look up who asked which question.

The first paragraph of Article L2312-38 covers recruitment methods: an agent that screens applications calls for prior information of the CSE, on top of the rules set out in our guide on AI resume screening under CNIL scrutiny.

The CSE's opinion is advisory: the employer may deploy after the consultation, even with a negative opinion. The risk lies in skipping the consultation, which lets the interim relief judge suspend the tool. Obstructing the regular functioning of the CSE is also punishable by a fine of €7,500 (Article L2317-1).

The timeline, from file to opinion

Consultation comes before the decision: under Article L2312-14, the employer's decisions are preceded by consultation of the CSE. The project must be precise enough to present, and the tool still closed to employees.

The period for the opinion runs from the day the employer hands over the written information, or announces that it is available in the BDESE (the company's economic, social and environmental database), under Article R2312-5. Article L2312-15 requires precise written information and enough time to examine it.

Hand over a complete file from the start. If the CSE considers it lacks sufficient information, it can ask the president of the court to order the employer to provide what is missing. That application does not extend the period, unless the judge decides so because of particular difficulties in accessing the information (Article L2312-15).

For the introduction of a new technology, the CSE can call in an accredited expert (Article L2315-94, 2°). The cost is then shared: 80% paid by the employer, 20% from the CSE's operating budget (Article L2315-80, 2°). In the case decided in Nanterre on 29 January 2026, the central CSE voted for an expert review on the very day the new tools were presented to it.

The plan for an HR AI agent therefore sets aside one month between handing over the file and going live, two months if the CSE appoints an expert, plus the time to prepare the file and hold the meeting. Check first whether a company agreement sets other deadlines.

What the file presents to the CSE

The file we prepare with our clients answers four questions: what the agent is for, what data it reads, who sees what, and how the pilot runs. The table gives its contents. It reflects our practice and the requirements of Article L2312-15; the Labour Code sets no list.

The "Who sees what" heading also decides whether the third paragraph of Article L2312-38 applies. If a manager or an HR administrator can find out who asked which question, the agent becomes a means of monitoring employee activity.

What French courts have decided since 2025

The France Télévisions case is the closest to an HR agent. According to Lefebvre Dalloz, the Raiponse HR chatbot had been running since November 2022 and the central CSE had been consulted when it was set up. For version 2, the judge noted techniques similar to the first, new features (detailed answers, summaries, translations) whose effect on employees' situation was not established, no planned training, and that the CSE had been informed of the upgrade in 2025. He refused to order a new consultation.

Does a pilot phase already require consultation?

Yes, as soon as employees use the tool in their work. On 14 February 2025, the Nanterre interim relief judge examined five applications using AI that had been in a pilot phase for several months while the CSE consultation was still open. According to the commentary published by Lamy Liaisons, the judge found that the pilot involved use of the tools, at least in part, by all the employees concerned. The judge treated it as a first implementation and suspended the deployment until the consultation ended.

The same commentary sets this pilot apart from a simple experiment needed to present the CSE with a sufficiently developed project. Our reading, to be confirmed by your counsel: a test run by the project team with made-up questions, with no employee users and no real personal data, is still preparation of the file. The pilot starts when a group of employees asks the agent their real questions.

The order of 29 January 2026, which we read in full, rules out another argument. CS Group France argued that its previous skills software already used AI and that the two new products took over its functions. The judge accepted that the old tool already used AI, then looked at the announced use: a tool previously limited to two business units, extended to all employees, presented as not optional, whose data would feed annual reviews, staffing on client assignments and training. She concluded that this was the introduction of new technologies within the meaning of Article L2312-8.

She ordered the employer to open the consultation of the central CSE and to suspend the deployment, each with a penalty of €500 per day of delay, starting one month after service of the order and for 90 days at most. The request concerning the expert review was sent back to the expedited procedure on the merits provided for by Article L2312-15.

For an HR AI agent, present the pilot to the CSE before it opens, with its group of employees and its duration. Present it again before extending it to all employees, or before using its data in another HR process. Between consultations, inform the CSE of each change, and consult again when the scope changes: a new group of employees, a new data source, writing into the HRIS, using conversations to assess anyone.

Between 11 and 49 employees, the CSE does not have this general consultative role, since its powers depend on headcount (Article L2312-1). Informing employees is still required: Article L1222-4 bars collecting information about an employee through a device they were not told about beforehand. An information notice sent before launch and a clear greeting message meet that requirement.

This guide describes the texts and the published decisions; it does not replace advice from a French employment lawyer on your project.

The gain is on repetitive questions: leave balance, health cover, expense claims. The assistant answers them at any hour and cites the document, and the HR team keeps its time for situations that need a person. We publish no automation rate: those in circulation come from vendors, with no published method, and depend mostly on the quality of the documents.

An HR AI agent applies your criteria and repeats your documents to every candidate and every employee who asks it something. A biased screening criterion, an outdated note or overly broad access repeats with every answer. Each of these risks is handled before launch, when the instructions and access rights are written.

In resume screening, the agent can amplify biases present in historical data or in the criteria. Audit the criteria, log every suggestion and have a recruiter review every rejection. For high-risk systems, human approval is a legal requirement in Europe.

An approval reduced to a click lets the machine decide in the recruiter's place. Show the reason for each suggestion next to the shortlist, and give the recruiter time to disagree with it.

The assistant sees pay, evaluations and sometimes health data. Limit its access to the data needed for the question, in the HRIS as in shared folders, and check compliance with the GDPR, the AI Act and the CNIL referentials before launch.

A poorly prepared deployment can trigger rejection by teams or candidates. During a reorganisation, a tool that sees employees' questions will be read as a monitoring tool. Write down what it will not do and present it to the works council before launch.

An assistant repeats faster, and to more people, what your documents and HRIS say. A misapplied agreement or a wrong payroll setting stays wrong.

The split is the same from one use case to the next: the assistant prepares, the employee confirms, the manager or HR decides. The HR team also keeps the documents up to date and handles every case the assistant hands over. The table applies this rule to the everyday tasks described in sections 03 to 05.

The HR leader runs the project. They choose the first use case, prepare the works council file with the DPO and legal, have a sample of suggestions reviewed at regular intervals to spot bias, and train recruiters to challenge a suggestion from the agent.

Start with a single high-volume use case whose effect can be measured: employees' repetitive questions or volume recruiting. Set human approval before you write the first instruction, and consult the works council and the DPO before launch, including for a pilot phase.

Then train HR staff to check a suggestion from the agent before approving it. To compare the platforms on the market, our guides and comparisons review the tools behind each of these workstreams.