I'm Raphaël Sellami, founder of HRSpeeder, an applicant tracking system (ATS) for small companies. This post is an engineering checklist for anyone building or buying a CV-screening feature in the EU, and it says plainly where the checklist stops.
The short answer
The EU AI Act classes systems that sort job applications among high-risk uses of AI. If your feature scores, ranks or filters CVs, assume you are in that category.
In practice, three things follow:
- Keep a human decision. The model orders or summarizes, a person decides.
- Inform candidates that AI is involved.
- Check your current obligations in the up-to-date text, because they are more detailed than anything below.
This post gives no deadlines, article numbers or fine amounts. I can't verify them here, and a wrong number in a compliance article does more harm than none. Treat what follows as an engineering checklist, not legal advice.
Is my feature in scope?
If it reads CVs and produces a score, a ranking or a shortlist, yes. A common objection is "it only suggests a reading order". A reading order still decides which CVs get careful attention and which get a quick skim. I would treat it as in scope rather than build a product on the argument that it is not.
Features that are clearly not sorting, such as a scheduling calendar or an e-mail template, are a different question and not covered here.
A design checklist
These are design choices that make human oversight real rather than decorative. They are good practice, not a list of legal requirements.
1. No automatic rejection, enforced in code
Don't rely on a UI convention. Make it impossible for a model output to change an application's status without a human author.
// Generic example, not tied to any product
type Status = 'new' | 'shortlisted' | 'rejected';
function changeStatus(app: Application, next: Status, userId?: string) {
if (!userId) throw new Error('A status change needs a human author');
return { ...app, status: next, decidedBy: userId, decidedAt: new Date() };
}
The AI writes to score and summary. Only a signed-in person writes to status.
2. Present the score as an order, not a verdict
Label it "reading priority", not "fit: 62%". Show it next to the evidence: strengths, missing skills, a short profile summary. A reviewer who sees why a CV scored low can disagree with it knowingly. One who sees only a number has nothing to disagree with.
3. Fix the criteria before you read
Score against the written job description, and write the evaluation criteria down before the first CV arrives. A grid fixed in advance is also the simplest protection against drifting criteria. Don't let the model infer criteria from anything other than the job requirements.
4. Keep the reasoning next to the result
Store with each score the job description version it was computed against, the model used, the date, and the strengths and missing skills it produced. When a candidate or a colleague asks "why was this CV read last?", you can answer.
5. Tell candidates
Add a plain sentence where the CV is collected, for example: "We use software that helps order applications for reading. A person makes every decision." Check the exact wording and content requirements in the current text.
6. Know where CV data goes
List every service that receives candidate data, and where it is hosted. Pasting CVs into a general-purpose chatbot sends personal data outside any processor contract, and leaves no history or pipeline. That is a GDPR problem before it is an AI Act problem.
7. Don't skip the low scores
If the score lets you ignore 160 of 200 CVs, you have rebuilt an automatic filter with extra steps. Skim the rest, at least briefly.
This costs less than people expect. The estimate in the 2026 CV-sorting guide for one job with 200 CVs, using its stated hypotheses:
- hypothesis: 3 minutes to read a CV carefully, and the 40 best-scored CVs are read in detail, so 40 × 3 min = 120 min;
- hypothesis: 30 seconds to skim each of the other 160 CVs, so 160 × 0.5 min = 80 min;
- total: 200 min, about 3 h 20.
The same guide estimates about 16 h 40 for doing everything by hand with a mailbox and a spreadsheet. These are estimates, not measurements. Redo the arithmetic with your own timings.
If you are buying, not building
Recruiters at small companies can turn the same checklist into questions for any vendor:
- Can the tool reject a candidate on its own? The answer should be no.
- Can I see why a CV received its score?
- Which sub-processors receive candidate data, and in which country?
- What does the vendor tell candidates about AI use?
- Who is responsible for keeping the obligations up to date as the rules are detailed?
A vendor who answers these precisely is a better sign than any badge on a pricing page.
What this checklist does not do
- It does not make a product compliant. Compliance depends on the full legal text, your role (provider or deployer) and your use case. Ask a lawyer for that.
- It does not give dates or penalties. Look them up in the current official text.
- A score does not replace judgment. It doesn't replace the interview or the human decision, and it can be wrong about a good candidate.
- It does not cover other jurisdictions. If you hire outside the EU, other rules apply.
- When to choose something else. If you hire for rare or confidential profiles, a recruitment firm may fit better. If you want no AI in the process at all, a plain ATS or a spreadsheet avoids the question, at the cost of reading every CV in full.
How HRSpeeder handles the human-decision part
For the record, here is how our product treats the first point of the checklist. HRSpeeder's AI analysis gives a compatibility score against the job description, strengths, missing skills and a profile summary. The score indicates the order in which to read CVs and never triggers an automatic refusal. The recruiter decides.
Candidate data is processed on a server in Germany, with the database in the EU (Ireland). CV analysis goes through Google Gemini, which receives only the content submitted to the service, and interview transcription goes through AssemblyAI. I am not claiming this settles anyone's AI Act obligations. It is a description of where the data goes, so you can answer question 6 above.
The free Starter plan includes the AI analysis for 10 applications per month and 2 active job offers, with no credit card.
Direct answers on plans, data hosting and how the analysis works: https://hrspeeder.com/reponses
This article was prepared with the help of an AI assistant, from information published on hrspeeder.com.
Top comments (0)