AI in recruitment: what AI really does at each step, and what it must not do

In recruitment, AI is used today for four concrete things: showing your job offers to the right profiles, screening and prequalifying applications, preparing interviews and keeping a record of every assessment. It does not choose the person you hire. The GDPR gives candidates the right not to be judged by a machine alone, and the AI Act classifies recruitment as a high-risk use. Here is what AI really does at each step, where it must stop, and how to choose a tool.

What is AI in recruitment?

Artificial intelligence applied to recruitment covers two families of tools. Generative AI writes: a job offer, an outreach message, a summary. Predictive AI ranks: it compares a profile with criteria and estimates how well it fits the role. Recent tools are agents, which chain several tasks on their own (targeting, distributing, analyzing, scoring) and hand control back to the recruiter when it is time to decide.

The difference matters, because the risks are not the same. A poorly written offer gets fixed at proofreading. A biased ranking rules out candidates without anyone noticing.

Step 1: attract the right profiles, including those who are not looking

An ad published on a job site reaches people who are actively looking. The others, already in a job, do not read job offers. Across Seeqle campaigns, 80% of the profiles reached are passive: we reach them through display, audio and social networks, where they spend their time, not on job sites.

This is the role of targeted distribution, known as programmatic. AI chooses, among thousands of placements, those where the target profiles are. It allocates investment across channels according to the applications they generate, then adjusts the campaign continuously by cutting what does not deliver. No team can make these trade-offs by hand across hundreds of sites.

Generative AI also comes in at this stage, to write or adapt the offer. It is in fact the most common use: according to APEC, the French executive employment association, most companies that use AI to recruit managers and professionals mainly use it to write their ads.

What it must not do: exclude an audience on a discriminatory criterion (age, sex, origin, place of residence with no link to the role). The non-discrimination principle of the French Labour Code applies to targeting as it does to selection, whether the decision comes from a human or an algorithm.

Step 2: screen and prequalify applications

This is the step where AI saves the most time, and the one where it demands the most vigilance. Serious prequalification rests on four building blocks.

Analyzing applications at scale

An effective campaign brings in hundreds of applications. Read one by one, the latest arrivals wait for days, and the good profiles go elsewhere. AI analyzes them all, in batches, as soon as they arrive. The recruiter opens a ranked list and starts with the applications closest to the role, instead of going through the pile in order of arrival.

Enriching incomplete applications

Some applications arrive without a CV, for example from a short form. A good tool does not discard them: it rebuilds the profile from public information, including professional social networks, or follows up with the candidate to get their CV. The candidate must be informed of this processing, and sources must remain traceable.

An explained score, never a bare mark

A mark of 72/100 is useless if no one knows where it comes from. A useful score says why the candidate fits the role and, above all, on which criteria they do not: an insufficient language level, too little experience, incompatible mobility. It also distinguishes what the candidate declares from what the tool infers. This detail lets the recruiter check the reasoning, and defend it to a candidate who asks why they were not selected. For how the comparison between a CV and a job offer works technically, see our article on AI CV matching.

Sending results back to the ATS

A score that stays in a separate tool forces the recruiter to juggle two screens. The analysis must go back into the candidate record in your applicant tracking system (ATS), attached to the right role. The stakes are real: according to APEC, half of large companies use recruitment software (ATS or HRIS) for their managerial and professional hires. What gets sent back, and in what form, always depends on the fields the ATS exposes.

Step 3: prepare and conduct the interview

The interview remains a human exercise. AI can, however, remove a large part of the preparation. From the analysis of the application, it produces an interview guide specific to this candidate for this role: the themes to explore and, for each one, an open question based on their actual career path. The weak point spotted at screening (a language, too little experience) becomes a precise question.

The time saved shows up in two places. Before the interview, the recruiter no longer rereads the file to build their questions. During the interview, they do not go back over what the CV already said and go straight to the points that will decide the hire. If the line manager receives the same guide, both interviews cover the same core questions.

What it must not do: analyze the candidate’s face, voice or emotions. Emotion recognition in the workplace has been prohibited by the AI Act since February 2025 (European Commission), and the Commission’s guidelines include the hiring process within this scope (Future of Privacy Forum analysis).

Step 4: the decision stays human

AI ranks, explains and prepares. It says neither yes nor no. Article 22 of the GDPR gives everyone the right not to be subject to a decision based solely on automated processing when it significantly affects them (GDPR text on EUR-Lex). Rejecting an application is generally considered to be such a decision.

In practice, a compliant tool prioritizes; it does not eliminate. The recruiter keeps the full list, can move up a poorly ranked candidate and validates every rejection. If you cannot explain to a candidate why they were ruled out, the tool has made a decision in your place.

Limits and obligations: bias, GDPR, AI Act

Bias does not disappear with automation

A model learns from data produced by humans, with their prejudices. The University of Washington measured it in October 2024: three language models, more than 550 real CVs and more than 500 job offers. The models preferred names associated with white people in 85% of cases, versus 9% for names associated with Black people, and male names in 52% of cases, versus 11% for female names (University of Washington, 31 October 2024). The study covers general-purpose models used without safeguards. It shows why a score must be explained and checked. We detail where bias comes from and how to audit it on our page about algorithmic bias in recruitment.

GDPR: inform, limit, stay in control

Beyond Article 22, the GDPR requires you to inform candidates about the processing of their data, to collect only what serves the recruitment and to set a retention period. Enrichment from public data is not exempt from these rules.

AI Act: recruitment classified as high-risk

The AI Act classifies AI tools used for employment, including CV screening software, as high-risk systems (Annex III). Since the AI Omnibus entered into force on 27 July 2026, the related obligations apply from 2 December 2027 (European Commission): risk management, data quality, logging, documentation and human oversight. The employer using the tool has its own obligations, distinct from those of the vendor. The article-by-article detail is on our page on GDPR and AI Act compliance in recruitment.

Where do companies stand with AI recruitment?

Adoption remains low in France, but intentions are rising fast.

• APEC, Recruitment practices for managers and professionals, 2025 edition: only 4% of companies that recruited a manager or professional in 2024 had built AI tools into their recruitment process, and 11% are considering it. Among those that do not use it, 62% think it would save them time (APEC study, PDF in French).
• KPMG, Trends of AI study, December 2024 (212 respondents in France): 45% of companies are considering using AI to prequalify CVs, and one in four already does (KPMG press release).
• LinkedIn, Future of Recruiting 2025 (global data): 37% of organizations are integrating or experimenting with generative AI, compared with 27% a year earlier. Those using it report saving around 20% of their work week, and recruiters who write their messages with AI are 9% more likely to make a quality hire (LinkedIn report).

The gap between APEC and KPMG comes down to scope: APEC surveys companies that recruit managers and professionals, of all sizes, while KPMG surveys a panel of decision-makers, three quarters of them executives or executive committee members. In both cases, intention far exceeds actual use.

How to choose an AI recruitment tool?

Six questions to ask before signing, in this order:

1. Is the score explained? For each candidate, you should see the criteria met and those that are missing.
2. Who sets the criteria? The recruiter should define and adjust them role by role, not have them imposed.
3. Does the tool eliminate on its own? The right answer is no: it prioritizes, you decide.
4. What happens to an application without a CV? It should be enriched or followed up, not lost.
5. What does it send back to your ATS? Ask to see the output in your own software, not a generic demo.
6. What compliance documentation is provided? Candidate information, retention period, source traceability, AI Act readiness.

To go further on prequalification tools, see how it works on our candidate prequalification page.

Frequently asked questions about AI in recruitment

Can AI replace a recruiter?

No. It handles the volume: distribution, screening, interview preparation. Assessing a person, negotiating and deciding remain with the recruiter and the manager, and the legal framework points the same way: the GDPR regulates automated decisions, and the AI Act will require human oversight.

Is it legal to use AI to screen CVs?

Yes, provided the final decision is not fully automated, candidates are informed and the data collected serves the recruitment. From 2 December 2027, the AI Act obligations for high-risk systems also apply.

How can you avoid bias in a recruitment AI?

Require a score explained criterion by criterion, set the criteria yourself, regularly check results by population and keep the ability to recover a poorly ranked candidate. A tool that does not show its reasons cannot be audited.

Where should you start with AI recruitment?

With the step where you lose the most time. If you lack applications, start with targeted distribution. If you receive too many to read, start with prequalification. Test on a few roles and measure the share of qualified applications before and after.

Does AI work with my current ATS?

Prequalification tools generally connect to the ATS on the market and send their analysis back to the candidate record. What gets sent back depends on the fields your ATS exposes: check it in your own software before committing.

Recruiting with Seeqle’s AI agents

Seeqle handles attraction, prequalification and interview preparation with two agents. The Attract Agent reads your job offer, targets profiles using more than 400,000 criteria and distributes your ad on Facebook, Instagram, LinkedIn, TikTok, YouTube, Spotify, job sites and online media, more than 2 million sites and apps in total. The Match Agent enriches every application, predicts skills and personality, generates an interview guide and sends an explained score out of 5 back to the candidate record in your ATS, depending on what your ATS allows. The decision stays human. Across our campaigns, 53% of applications are qualified on average.