AI in HR firms and recruitment: CVs, interviews, the AI Act, and local AI
A CV is a name, an address, a career path, sometimes a photo and a date of birth. Personal data in the full sense of the term, entrusted by someone hoping for a job and having no way of knowing where they actually end up.
Recruitment is also one of the few areas where European law specifically regulates the use of artificial intelligence. In other words: it is not a sector where you improvise.
To understand where local AI really makes a difference, the simplest thing is to follow a CV from start to finish. Step by step, here is what it brings — and the pitfall to avoid every time.
The journey of a CV, step by step
Receiving the application
The CV arrives by email or via a form. This is the moment it becomes data that you process, with the obligations that go with it: informing the candidate, defined purpose, announced retention period.
The pitfall: the habit of forwarding a CV to an online tool to "see what it says about it". This simple act transmits personal data to a third party, often outside Europe. With local AI, the CV stays within the firm. We can also set up a custom AI automation that sorts the applications received by position and prepares a summary sheet, reviewed by the recruiter, without any CV leaving your walls.
Reading and summarizing
This is where local AI saves the most time. It extracts the career path, standardizes disparate formats, summarizes technical skills, calculates experience durations. Twenty unreadable CVs become twenty comparable sheets in a few minutes.
The pitfall: crossing the line between summarizing and scoring. Asking for a summary is useful; asking for a compatibility score that will decide the candidate's fate is something else entirely, legally risky and technically biased.
Preparing the interview
Based on the job description and the career path, the AI proposes relevant questions about the skills actually expected. The consultant chooses, discards, completes. Result: better prepared interviews, fairer among candidates because they are better structured.
The pitfall: letting the AI generate questions based on personal elements of the CV — apparent age, family situation, origin of the name. Frame the request around skills and the position, never the person.
The interview and its report
With the candidate's explicit consent, the interview can be transcribed locally and then summarized. The consultant listens instead of taking notes, and the report is ready in a few minutes. The candidate's voice — potential biometric data — never leaves the machine.
The pitfall: automatic analysis of emotions, tone, or "personality" from video or voice. These techniques are based on contested scientific foundations and are particularly regulated in a professional context.
The decision
The AI can structure the factual elements collected, prepare a comparative summary for the client, help write a reasoned recommendation. These are formatting tasks, not adjudication.
The pitfall: the only truly serious one. A rejection decision based exclusively on automated processing is regulated by the GDPR, and the risk of reproducing historical biases is real. The decision must be human, reasoned, and you must be able to demonstrate it.
Archiving — and purging
The unsuccessful CV is subject to a limited retention period. This is the step everyone forgets, and it becomes more delicate when documents have been indexed in an AI tool.
The pitfall: thinking that purging the file is enough. If the CV was indexed in a document database, it must also be removed from the index. Plan a periodic purge from the start — it is much easier to automate at the beginning than to catch up later.
Three mistakes that cost a lot
The most frequent shortcomings do not come from bad intentions, but from a reflex taken without thinking.
1. Believing that going local solves everything
This is the most widespread mistake, and it deserves to be stated clearly: moving from the cloud to local AI removes the problem of data transfer. It does not remove the underlying rules. A local AI that automatically sorts applications, or that deduces a candidate's age, would remain just as problematic.
What to remember: local protects confidentiality, your configuration protects fairness. Both are necessary.
2. Recording an interview without explicit consent
Automatic transcription is one of the most appreciated uses in HR firms. But a recording made without prior and clearly stated consent poses an immediate problem — even if everything stays local and no one else has access.
What to remember: ask for consent at the beginning of the interview, explain the purpose and retention period, and accept a refusal without consequences for the candidate.
3. Indexing everything, for convenience
Since the tool accepts PDFs, the temptation is to pour all the files into it: photos, ID documents, personal correspondence, informal notes. This is contrary to the principle of minimization, and it unnecessarily increases the risk of bias.
What to remember: only index what is relevant to the position. Many firms even anonymize the CV before processing — removing name, photo, address, and date of birth — which mechanically reduces bias while retaining the essentials.
The rule of the three H's
If you were to remember only one thing from this article, it would be this. Three simple words to frame any use of AI in a recruitment process.
Three conditions to always respect
The three go together. A local AI without human supervision remains risky; impeccable supervision on a cloud tool leaves the transfer question entirely; and both together do not dispense with informing candidates.
Concretely, what changes in the firm?
Before, a consultant spent a large part of their week reading heterogeneous CVs, reformatting information, writing interview reports in the evening, and sending impersonal rejection responses for lack of time.
With a well-configured local AI, these tasks are compressed. CVs arrive summarized in a homogeneous format. Reports are written while the next interview begins. Candidate feedback becomes personalized, because they no longer cost twenty minutes each. The consultant regains time for what makes the value of the profession: meeting, understanding, advising.
And during all this time, no CV has left the walls of the firm.
The hardware: our range
The needs of an HR firm are moderate in calculation but real in document volume: many CVs, reports, reference frameworks to index, and often several consultants to serve. Here is a selection from our range, from the individual workstation to the enterprise server. Each local AI server or workstation is then sized according to the number of consultants to serve. Our guide details the cost of a local AI server.
Individual workstations and small firms
CoreIA RTX 5060 TiIndependent consultant: CV summary, transcription, writing.
€2,068
CoreIA Radeon AI PRO R9700High-volume firm: indexing a large document pool.
€4,162
CoreIA RTX 5090Demanding firms: 30B models, mass transcription, maximum quality.
€10,291
Shared server for the firm
CoreIA GB10 (ASUS Ascent GX10)128 GB unified: multiple connected consultants, common database, silent.
€6,999
Rack and multi-GPU for large firms and groups
CoreIA 2x RTX 5090 (64 GB)4U rack: consultant teams, parallel transcription and indexing.
€20,954
CoreIA Threadripper PRO 4 GPU96 GB ECC VRAM, up to 4 cards: very large document pools, continuous reliability.
€32,647
CoreIA 2x RTX PRO 6000 (192 GB ECC)HR groups and large organizations: dozens of users, 24/7 availability.
€48,508
All these machines are assembled by hand in Auriol (13390), except the GB10 built by ASUS, fully configurable via the online configurator or by quote at contact@radiancesystems.eu, and deliverable with the pre-installed AI environment: language model, chat interface, local transcription and optional document indexing.
This article presents general guidelines and does not constitute legal advice. Recruitment being an area strongly regulated by the GDPR, the labor code, and European regulations on artificial intelligence, it is recommended to have your system validated by specialized counsel and to document it in your processing register. Radiance Systems provides the hardware and technical support; the compliance analysis is your responsibility and that of your counsel.
