GuidesAI at work
AI in recruitment: what to hand over, step by step
AI in recruitment handles sourcing, first screening, follow-ups and interview notes. Rejections and scores stay human, and the law already requires it.
Contents
AI in recruitment works very well on one specific part of the job, the part where volume is high and the decision is light. Searching a talent pool of ten thousand profiles, reading two hundred applications with the same attention from the first CV to the last, writing the follow-ups that are already late: all of that can be handed over today. Rejecting a candidate, scoring an interview or choosing who goes to the client stays with the recruiter, because the law requires it and because that is where the placement is made.
Adoption figures follow the same dividing line. According to Aptitude Research and iCIMS figures published in April 2026, 58% of companies use AI to screen applications and 54% for candidate communication. Half of them use it for assessments and 46% for sourcing. The more repetitive a stage, the more it is already tooled, and the trouble starts when that logic spills over into the stages that carry a decision.
This guide follows a hire in order, from the intake meeting to the client presentation, and says at each stage what AI does well, what it does badly and what you keep. We take the example of a recruitment agency with six recruiters placing technical and sales profiles, because an agency puts every stage in very few hands. The reasoning holds just as well for an in-house HR team hiring for its own company.
What does AI do in recruitment, stage by stage?
The table below sums up the split we recommend. Each row is detailed further down, with what tips a task to one side or the other.
| Stage | What AI does well | What you keep |
|---|---|---|
| Intake and job ad | Structuring intake notes, drafting the ad | The ranking of the client’s criteria |
| Sourcing | Searching the talent pool first, writing the search | Who to contact |
| Screening | Ranking each application and saying why | Every decision to turn someone down |
| Follow-ups | Drafting the right message at the right time | Reviewing and sending |
| Interview | Preparing the scorecard, cleaning up the notes | The score |
| Client presentation | Building the profile in the client’s format | Which profiles are presented |
The intake meeting and the job ad
The intake meeting is the only moment when you can still change the odds of placing, and it is still the stage most often rushed (what the intake meeting decides gives the figures). AI helps in a very concrete way: it transcribes the call with the client and sorts what was said. Above all, it lists what was not asked, such as the salary range, the share of remote work or the criterion that will get a candidate rejected even though they look perfect on paper.
It then writes the job ad in seconds, and that is where you need to hold it back. Left to itself, it produces the average ad of its training data, with its “exciting environments” and its lists of qualities no candidate reads any more. A good ad first gets the wrong profiles to rule themselves out, which means giving the model the real constraints of the role rather than a job title.
What you keep is the ranking of criteria. Only the conversation with the client tells you that Kubernetes is a deal-breaker while fluent English is not, and no model can guess what the client never wrote down.
Sourcing, starting with your own talent pool
The first useful reflex is to search your own data before paying for an external search. According to Gem, the share of sourced hires whose candidate was already in the company’s CRM or ATS rose from 29.1% in 2021 to 44.0% in 2024, on US data. An AI connected to your ATS rereads in minutes the profiles nobody opens any more, including those a departed recruiter left behind: that is the whole point of the talent pool nobody searches before paying to source.
For external search, it excels at a thankless exercise, the Boolean string. You have to think of job titles in several languages, neighbouring technologies, common typos and exclusions. A person loses that rigour around 6 pm on a Friday and the model never does, provided you give it the full brief and read the string before running it.
Who to contact stays with you, as does the first message. AI can draft it with an opening line that cites the candidate’s actual career, but a recruiter reviews it and sends it.
The first screening of applications
This is the stage where AI brings the most, and the one where it does the most damage when it is allowed to decide. Let us start with the concession, because it is real. A model reads the two hundredth application with the same attention as the first, while a recruiter who has opened a hundred and fifty that day no longer sees much. On consistency of reading, the machine does better than we do.
The problem comes from what it is asked to do next. A score of 87% does not say what it measured, so nobody can challenge it, and it ends up being applied as it stands. The figures compiled by Employer Branding News in 2026 leave no doubt: 75% of companies let an AI reject an application without any person having read it. On the candidate side, only 26% consider it fair to be assessed by an AI, a distrust that the match score does nothing to dispel.
The rule we apply fits in one sentence: AI ranks and explains, it does not score and it turns nobody down. Next to each application it writes what it found, what is missing against the brief and what it could not verify. The recruiter reads those three lines, and the recruiter is the one who says no.
One technical risk deserves to be known, because few agencies think about it. A CV can contain invisible text, white on white, that gives orders to the model reading it. A tool that puts a candidate at the top because a PDF told it to is a dangerous tool, and prompt injection in a CV sets out what to demand from a vendor on that point.
Follow-ups and messages to candidates
Silence has become the norm: according to MSH, 53% of candidates were ghosted by an employer in 2026, against 38% in 2024. The cause lies less in rudeness than in overloaded recruiters, and that is exactly the kind of problem AI handles well. Writing the right message at the right time to a hundred candidates mostly takes consistency, as our analysis of candidate ghosting shows.
In practice, AI spots in the ATS that a candidate has been waiting five days for an answer, drafts the follow-up with the context of their file and offers it to the recruiter, who reviews it before sending. Nothing that goes to a candidate is sent without a person releasing it, not even a polite rejection. We treat that line as non-negotiable, and who approves what an AI writes to candidates explains why.
The situation changes if the candidate talks directly to an agent, in a chat or on the phone. They must then know they are talking to an AI, under Article 50 of the AI Act, which has applied since 2 August 2026. The exact line between the two situations is finer than it looks: a message the AI drafted and a recruiter reviewed and sent does not fall under it.
The interview
Before the interview, AI prepares a structured scorecard from the brief: the questions, and for each one what a strong, an average and a weak answer would show. During the interview it takes the notes, then files them in the candidate’s record. The recruiter keeps their full attention on the person in front of them, which is the real reason to use a note-taking tool.
The score stays human. An interview is only structured if each answer is scored during the interview, by the person who led it, on a scale where each level describes an observed behaviour. Our structured interview scoring grid gives the full template.
One practice is simply banned: inferring a candidate’s emotions from their face or voice. Article 5 of the AI Act has prohibited inferring emotions in the workplace since 2 February 2025, and the European Commission’s guidelines explicitly bring recruitment within its scope. A video interview tool that promises to measure motivation or stress therefore exposes you to a prohibited practice, not to a mere compliance formality.
The client presentation
AI builds the presentation from the CV, the interview notes and the brief, in the format the client expects. In an IT services company this document is the skills dossier, and it carries the company’s signature: it is a sales document, not a reformatted CV. Choosing, ordering and rewording a career history is part of the job, adding what is not there never is.
Choosing the three profiles presented stays with the recruiter, as does the sentence that tells the client why this one rather than another. That is when the agency puts its credibility on the line, and credibility cannot be delegated.
What should you never hand over to AI in recruitment?
Anything that closes a door on a candidate stays human, even when the action never leaves the agency. Rejecting an application, dropping a profile mid-process, scoring an interview and choosing who goes to the client all belong in that category. So do the messages that go outside, which AI can draft but a person must release.
The objection is fair and deserves a moment: a tired recruiter also turns candidates down for no good reason, and nobody asks them to justify each rejection. The difference lies in scale and in the ability to challenge. A person’s mistake is not repeated identically across two thousand files, and you can always ask them why they said no. A badly set rule in a tool applies the same mistake to everyone, silently, until the day a candidate asks for an explanation.
The client relationship and the negotiation also stay on your side, for an even simpler reason: that is what the client pays for. An agency that delegates sourcing frees up time for the phone, whereas an agency that delegates the relationship has little left to sell.
What does the law say about AI in recruitment?
The AI Act classifies recruitment as high risk. Point 4 of Annex III covers systems intended for the recruitment or selection of people, in particular to place targeted job ads, to analyse and filter applications or to evaluate candidates. The obligations attached to that classification, from human oversight to logging, were deferred by Regulation (EU) 2026/1744 to 2 December 2027. No tool can therefore claim to comply with Annex III yet, and a vendor who says so is telling you something false.
The deferral does not suspend the rest, as our article on what still applies despite the AI Act deferral sets out. The ban on inferring emotions has applied since 2 February 2025, as has the duty to train the teams who use AI. The transparency rules of Article 50 have applied since 2 August 2026. And Article 22 of the GDPR, in force since 2018, still forbids a decision with a significant effect on a person from resting on fully automated processing.
Two steps weigh concretely on a project. The CNIL, the French data protection authority, lists processing that supports recruitment through an algorithm among those requiring a data protection impact assessment, or DPIA. In a company with a works council, the rollout is also prepared with it. The Nanterre court suspended two AI-based HR tools in January 2026, pilot phase included, and our article on consulting the works council describes what the judges expect.
Where should a recruitment agency start?
Start with what can be checked at a glance and closes no door. For the agency in our example, that means two uses in the first month: searching the talent pool before every paid sourcing effort, and draft follow-ups that each recruiter reviews before sending. A mistake shows immediately, it costs a candidate nothing, and the team learns to work with the tool on safe ground.
Interview preparation and note-taking come next, then the client presentation. Ranking applications comes last, and always in the same form: an explained ranking that you compare for a few weeks with your own reading on roles you have already filled. It is the only honest way to judge a tool, and far more reliable than a demo the vendor prepared.
That is the order we recommend with Balt. He works directly in the team’s ATS and chat tools, prepares searches, follow-ups and meeting notes, and a person releases everything that goes to a candidate or a client. Our page for recruitment agencies details what he takes on.
Which AI tools should you use to recruit?
The tools fall into three families that do not meet the same need. Consumer assistants such as ChatGPT write a job ad or a Boolean string very well. They cannot see your ATS, however, and nothing containing a candidate’s or a client’s name should go into them.
The AI features built into an ATS work on your data, within the limits the vendor has set. Agents, finally, work inside your existing tools and chain the stages together, provided the ATS opens the right doors to them. The list of what an agent needs from your ATS lets you check that before you sign.
To compare the products on the market by name, Cobalt’s roundup of the best AI tools for IT recruitment (in French) reviews twelve of them, with their uses and their limits.
Whichever tool you choose, one question comes before you connect it to your candidates: do you need an impact assessment, and which one? Our article on the DPIA for a recruitment AI agent answers it criterion by criterion.
Frequently asked questions
Can AI screen CVs instead of the recruiter?
It can rank them and explain each ranking, but it should not decide alone who is turned down. Article 22 of the GDPR forbids a rejection made without genuine human involvement, and a recruiter approving a list in bulk without reading it does not count. The right use is an explained ranking that the recruiter reads before replying to the candidate.
Do candidates have to be told that AI is involved in recruitment?
Yes, as soon as a candidate talks directly to an agent, under Article 50 of the AI Act, which has applied since 2 August 2026. When a person reviews and sends a message an AI drafted, that article does not apply, but the GDPR still requires informing the candidate about how their data is processed.
Will AI replace recruiters?
It takes the volume of the job: the searching, the reading of hundreds of applications, the writing of follow-ups and notes. What makes a placement stays with the recruiter, whether that means understanding what the client really wants or convincing a candidate who was not looking.
Where should a recruitment agency start with AI?
With searching your own talent pool before every paid sourcing effort, then with draft follow-ups that the recruiter reviews and sends. Both can be checked at a glance, neither closes a door on a candidate, and within a few weeks they show whether the tool holds up on your real roles.
Sources
- Pin, AI Adoption in Recruiting: The Complete 2026 Industry Report (Aptitude Research, iCIMS and SHRM data)pin.com
- Employer Branding News, AI in hiring statistics 2026: adoption, bias, trust and regulationemployerbranding.news
- Gem, Hidden in plain sight: the untapped potential of candidate rediscoverygem.com
- MSH, Candidate experience statistics, data & trends 2026talentmsh.com
- EU AI Act, Annex III: high-risk AI systems (point 4, employment)artificialintelligenceact.eu
- Regulation (EU) 2026/1744, Digital Omnibus on AI, OJEU publicationnicfab.eu
- European Commission, Guidelines on prohibited artificial intelligence practices under the AI Act (2025)ai-act-service-desk.ec.europa.eu
- European Commission, Transparency obligations under Article 50 of the AI Act (FAQ)digital-strategy.ec.europa.eu
- CNIL, GDPR, Chapter III, Article 22: automated individual decision-making (in French)cnil.fr
- L’Expertise Droit Social, AI and works council consultation, Nanterre court, January 2026lexpertise-droit-social.fr
- Cobalt, The 12 best AI tools for IT recruitment in 2026 (in French)cobalt-ia.com

