Yes, in most screening and ranking uses. Annex III of the EU AI Act lists AI used to recruit or select people, including targeting job adverts, filtering applications and evaluating candidates, as high-risk. Since the Digital Omnibus, those duties (human oversight, logs, candidate notices) apply from 2 December 2027; the emotion-recognition ban already applies.
A small agency is not off the hook until then. The ban on inferring emotions at work, which the Commission reads as covering candidates, has applied since 2 February 2025, AI literacy has been an expectation since the same date, and transparency duties for chatbots began on 2 August 2026, the date the Commission gives for enforcement starting. What follows is general information, not legal advice; ask a qualified adviser about your own position.
The dates that matter, in the order you will meet them
The Act has staggered start dates, and the 2027 deferral is the one most people have heard. The others are easier to miss. Here is the timeline as the Commission's AI Act Service Desk presents it, updated for the Omnibus amendments.
| Date | What applies | Why a recruiter cares |
|---|---|---|
| 2 February 2025 | Prohibited practices and the AI literacy provision | Emotion inference at work is banned; staff using AI need some understanding of it |
| 2 August 2026 | Transparency rules (Article 50) apply; enforcement starts for prohibitions, transparency and literacy | Candidate-facing chatbots must say they are AI |
| 2 December 2026 | New prohibitions on AI that generates non-consensual sexual deepfakes or child sexual abuse material; Article 50(2) deadline for some synthetic-content providers already on the market | Rarely relevant, but it shows the Act is still moving |
| 2 December 2027 | Annex III high-risk rules, deferred from 2 August 2026 | CV filtering, ranking and candidate evaluation fall here |
| 2 August 2028 | High-risk AI built into regulated products | Seldom relevant to an agency |
One caution on the AI literacy line. The Omnibus amended Article 4, which now says providers and deployers "shall take measures to support the development of AI literacy" of their staff, and adds that this "does not require providers or deployers to guarantee any specific level of AI literacy of any individual". The duty survives in a softer form, so check that any training material or policy you use quotes the amended text.
What Annex III actually says about recruitment
Annex III, point 4, is headed employment, workers management and access to self-employment. Point 4(a) covers 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. Point 4(b) covers decisions affecting the terms of work relationships, promotion or termination, allocating tasks based on individual behaviour or personal traits, and monitoring and evaluating performance and behaviour.
Read the verbs. "Place targeted adverts", "analyse and filter", "evaluate". These are not the same as "write", "summarise" or "schedule". A system that decides who sees a vacancy, who moves to the next stage or how good a candidate is sits in scope. A system that tidies a CV into a standard layout does not, unless it also scores. For a wider look at hiring and credit tools, see whether your hiring or credit tool is high-risk.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowDoes the Act reach an agency that is not in the EU?
Possibly. Article 2(1)(c) extends the Act to providers and deployers located in a third country "where the output produced by the AI system is used in the Union". If you screen candidates for an employer in the EU, or you assess EU-based applicants, the output is used there. A useful rule of thumb: if the outcome of the tool ends up affecting someone's chance of a job at an EU employer, assume the Act is relevant and ask your adviser to confirm.
Example: a small agency places technicians with a client based in Germany. Its AI shortlist tool is run from the agency's own office, but the shortlist goes to that client and decides who gets interviewed. That output is used in the Union. The client will probably ask the agency to show how the tool works and who checks it, so it is better to have the answers ready.
A four-question test for each tool you use
Run every AI feature in your recruitment stack through the same four questions. Do it per feature, not per product, because one ATS can contain a harmless scheduler and a high-risk ranker.
- Does it touch a candidate or worker? If the feature never sees information about individual people, it is outside Annex III.
- Does it do one of the listed jobs? Targeting adverts, analysing or filtering applications, evaluating candidates, or the 4(b) decisions about terms, promotion, task allocation or monitoring.
- Does it profile people? Article 6(3) lets a provider conclude that an Annex III system is not high-risk when it does a narrow procedural task, improves a human activity already completed, detects patterns without replacing a human assessment, or does a purely preparatory task. But a system that profiles natural persons is always treated as high-risk.
- Who changed what it is for? Under Article 25(1), a party that modifies the intended purpose of an AI system, including a general-purpose one, so that it becomes high-risk is treated as the provider. If you prompt a general chatbot to rank CVs, you may have turned it into a high-risk recruitment system with you as the provider.
If the answer to question 2 is yes and question 3 is not clearly no, treat the tool as high-risk and plan for December 2027. If the provider says it is not high-risk under 6(3), ask to see the documented assessment, because the Act requires the provider to document that conclusion before the system goes on the market.
Eight tools sorted with that test
Here are illustrations for a six-person agency that places solar-installation technicians. Treat the verdicts as starting positions for a conversation with your adviser and vendors, not rulings.
| Tool or use | Annex III 4(a)? | Starting position |
|---|---|---|
| ATS feature that scores CVs against a job spec | Yes: analyses and filters applications | High-risk; vendor is provider, agency is deployer |
| Sourcing tool that ranks profiles from a database | Yes: selection | High-risk; check the ranking logic is explained |
| Job-board option that targets adverts to likely applicants | Yes: targeted job advertisements | High-risk; ask the platform for its documentation |
| Video interview tool that scores answers | Yes: evaluates candidates | High-risk; check no emotion inference at all |
| Video tool that reads facial expression or tone | Banned practice | Switch it off; the ban already applies |
| CV parser that only fills fields | Arguably a narrow procedural task | Lower risk; ask the vendor to confirm in writing |
| Chatbot that writes advert text | Not listed | Not high-risk; still check wording for bias |
| Scheduler that books interview slots | Not listed | Not high-risk; disclose if candidates talk to a bot |
The emotion-inference ban, and why interview tools are the trap
Article 5(1)(f) prohibits using AI to infer the emotions of a person in the workplace and in education institutions, with an exception for medical or safety reasons. The Commission's guidelines on prohibited practices, published on 4 February 2025, say the notion of "workplace" should also apply "to candidates during the selection and hiring process", and list as an example that "using emotion recognition AI systems during the recruitment process is prohibited". So analysing a candidate's face or voice to guess whether they are nervous, enthusiastic or honest is out. The guidelines state that they are non-binding, with the final word left to the EU's Court of Justice, but they show how the Commission reads the ban.
The practical danger is a feature you did not know you had. A recorded-interview product may offer "engagement" or "confidence" scores. Ask the vendor three direct questions: does any score use facial expression, voice tone or body language; can that be switched off at account level; and will they confirm in writing that it is off. Disclosing the feature to candidates does not make a banned use lawful.
What a deployer must do once the high-risk date arrives
Most small agencies will be deployers, meaning they use someone else's high-risk system. Article 26 sets out their duties. In plain words:
- Use it as intended. Take appropriate measures to use the system according to the provider's instructions for use. Keep the instructions.
- Assign human oversight. Give it to people with the competence, training and authority to override the system.
- Check input data. Where you control the data going in, make sure it is relevant and sufficiently representative for the purpose.
- Monitor it. If you have reason to think it presents a risk, inform the provider and suspend use.
- Keep the logs you control for at least six months, unless other law says otherwise.
- Inform workers and their representatives before putting the system into use at the workplace, if you are an employer.
- Inform the people it is used on. Where an Annex III system makes or assists decisions about natural persons, you must inform them that it is being used.
Notice that none of this needs you to be technical. It needs a named person, a written routine and a record. A tool that cannot give you its instructions for use, or its logs, is a tool to question now, while you can still switch vendor.
Worked example: a deployer file for one tool
Take the six-person solar-technician agency and its ATS ranker. The managing director spends about half a day on a one-page deployer file. Everything below is illustrative.
| Item | Entry |
|---|---|
| Tool and feature | ATS "match score" on inbound CVs |
| Annex III reference | 4(a): analyse and filter applications |
| Role | Deployer; vendor is provider |
| Instructions for use | Saved on [date]; reviewed by the MD |
| Human oversight | Senior consultant reviews every candidate scored below 60 before rejection |
| Input data | Job spec written by consultant; no protected characteristics entered |
| Logs | Vendor retains; agency exports monthly and keeps for 12 months |
| Candidate notice | Line added to the application form and auto-reply |
| Review date | Quarterly, next [date] |
The oversight rule is the part that matters most. If nobody is actually reading the low-scored CVs, the score is making the decision. A simple test: pick ten rejected applicants at random each month and have a second consultant read them. Count how many she would have shortlisted. If it is more than one or two, the score is wrong for your roles and the oversight is doing real work.
Candidate wording you can adapt
You will need two short notices. The first covers the use of a screening tool; the second covers a chatbot or auto-responder. Adapt the wording to your own set-up and have a qualified adviser check it.
How we use AI in our process: we use software to help sort applications against the job requirements. A consultant reviews the results before any decision is made, and you can ask us for a person to look at your application again. Contact us at the address on this page if you would like to know more.You're chatting with an automated assistant. It can answer questions about the role and book interview times. A member of our team will take over for anything else. Type "person" at any time to speak to someone.The second is about Article 50, which applies from August 2026 and requires people to be told when they are interacting with an AI system, unless it is obvious. For more on that wording, see disclosure wording for your website chatbot.
Questions to put to each vendor this quarter
The vendor's answers decide most of your workload. Send these in one email and file the replies.
Subject: EU AI Act - [product name] used for candidate screening
1. Do you classify [feature] as high-risk under Annex III 4(a)? If not, please share your Article 6(3) assessment.
2. Please send your instructions for use and any technical documentation we need to run it as a deployer.
3. Does any feature infer emotions from face, voice or behaviour? Please confirm it is disabled for our account.
4. What logs can we export, how long do you keep them, and in what format?
5. What human-oversight controls does the product offer (override, explanations, review queues)?
6. What is your timetable for conformity assessment and registration before 2 December 2027?
7. Who is your contact for incidents and for requests from affected candidates?A vendor that cannot answer questions 1 to 3 has told you something. For more on running those conversations from the client's side, see what clients ask an agency about its AI screening.
When a vendor says its tool is not high-risk
Some vendors will answer question 1 of the email with "no, we are not high-risk". That can be true, since Article 6(3) lets a provider document that a system does not pose a significant risk, but it deserves a closer look. Ask three follow-ups: which of the four conditions it relies on, whether the tool profiles individuals, and where the documented assessment is. A match score that a consultant sees as a suggestion may be "preparatory"; a score that automatically rejects below a threshold is not.
Take a security-installer agency whose ATS marks candidates "unsuitable" and hides them from the consultant's default view. The vendor calls this "assistive". In practice the consultant never sees the hidden candidates, so the tool is deciding who is considered. That should make you treat it as high-risk until the vendor shows otherwise, and at minimum change the setting so every application stays visible.
A monthly rejection check with real numbers
The random read of rejected applicants is the cheapest evidence of working oversight. Here is how it might look over three months for an HVAC-technician agency. All figures are illustrative.
| Month | Rejected by the score | Sampled | Second reader would shortlist | Action |
|---|---|---|---|---|
| January | 120 | 10 | 3 | Score threshold lowered; job spec reworded |
| February | 104 | 10 | 1 | Keep; note two missed apprenticeship CVs |
| March | 131 | 10 | 0 | Keep; sample again in April |
The January figure is the useful one. Three in ten wrongly rejected is a sign the tool was misreading experience from a trade background. Because the agency sampled, it found out in a month, not in a complaint. Keep the sheet; it is the sort of record that shows oversight is real. Fix the job spec first, since vague specs produce poor scores, and tell the vendor if the pattern persists.
Ranking CVs in a general chatbot: the audit problem
An agency uses a general chatbot subscription to rank 80 applicants for a security-installer role. A consultant pastes the CVs, writes "score each out of 10 for suitability" and sends the top 12 to the client. It feels like harmless productivity.
Under the test above, that is analysing and filtering applications, which is the Annex III use, and Article 25(1) says changing a general system's purpose so that it becomes high-risk can make the agency the provider. The agency has no documentation, no oversight routine and no candidate notice. Nothing has gone wrong yet, but it would be hard to defend if a candidate asked how they were scored.
The fix is not necessarily to stop. It is to choose: use a recruitment product whose vendor carries the provider duties, or use the chatbot only for drafting and have the consultant make every screening decision from the CVs themselves. Write the rule into your staff guidance.
A plan from now to December 2027
You do not need to do everything at once. A spread of small steps keeps the cost low.
- This month. List every AI feature in sourcing, screening, interviewing and advertising. Switch off anything that infers emotions.
- Next quarter. Run the four-question test on each and send the vendor email. Add the candidate notices.
- By mid-2027. Write the deployer file for each high-risk tool, name the oversight person and start the monthly random read of rejections.
- By autumn 2027. Check which vendors have completed their own work. Replace any that have not. Brief staff on the routine.
- December 2027. Confirm the logs, notices and oversight run as written, and diarise a quarterly review.
Fines are real but should not be your only motive. The Act's penalty article sets fines of up to EUR 15 million or 3 per cent of worldwide turnover for breaches of the deployer and transparency duties, and up to EUR 35 million or 7 per cent for prohibited practices such as emotion inference at work, with the lower of the two figures applying to SMEs and start-ups. Read what a small business actually risks for the proportions, and treat the figure as a reason to keep records, not as a forecast.
AI literacy for a six-person team
The literacy duty is the cheapest to meet and the easiest to forget. For a small team, an hour is enough to start. Cover which tools are approved, what each may be used for, why candidate data must not go into unapproved tools, what the emotion-inference ban means, and who to ask. Keep a dated attendance note. If you want to know what outside support looks like, AI literacy consulting explained sets out what it covers and when it is worth paying for.
If you would rather map your tools and roles with someone, my AI implementation consultation covers exactly that, and you can start from the contact page. Keep the legal reading with your own adviser.
EU AI Act and recruitment: follow-up questions
Does the Act apply to an agency outside the EU?
It can. Article 2 covers providers and deployers established or located outside the Union where the output produced by the AI system is used in the Union. An agency placing candidates with EU employers, or screening EU-based candidates, should assume it is in scope and take advice on the detail.
Is a CV parser high-risk?
It depends on what it does. A tool that only extracts names, dates and job titles into fields is arguably a narrow procedural task, which Article 6(3) can take outside high-risk if the provider documents that conclusion. A tool that scores or ranks candidates is analysing and filtering applications, and is the core case Annex III describes.
Can we still use ChatGPT to draft job adverts?
Drafting advert wording is not what Annex III lists. Placing targeted job advertisements by AI, meaning deciding who sees them, is listed. Writing the text is lower risk, though you should still check the wording for bias and tell people if they are talking to a chatbot.
Who is responsible when the vendor's tool is the problem?
Both of you, for different things. The vendor, as provider, has the design and documentation duties. You, as deployer, must use the tool as its instructions say, assign trained human oversight, keep the logs you control for at least six months and inform affected people. Ask each vendor to confirm its position in writing.
Further reads
- AI in Recruitment and Hiring: A Bias Test Checklist for Small Agencies — A bias test routine for small agencies.
- How to Evaluate the AI Features in Your Recruitment Software — How to evaluate the AI in your recruitment software.
- AI Screening: Bolt It Onto Your ATS or Switch Systems? — Bolt screening onto your ATS or switch systems.
- AI Agents for Recruitment Agencies: Sourcing, Screening and Review Points — Review points for AI agents in sourcing and screening.
- AI Sourcing Tools for Recruiters: Seat Costs and Consent Rules Compared — Five sourcing tools priced for three recruiters, with data sources, consent handling and a one-live-search trial you can run in a week.
- AI Talent Recruiting Software: Manatal, Loxo or Bullhorn Amplify? — Manatal from $15 a user, Loxo from $149, Bullhorn from $99 on its small-agency plans: what each costs a five-person agency and what the AI does.
- Resume Formatting AI Tools for Recruiters: Check Nothing Was Altered — Which AI CV formatters reflow and which rewrite, what each costs per CV, and a step-by-step check that nothing in the candidate's CV was altered.
- AI Recruitment Platform for Employer Branding: What to Look For — What a small employer should check in an AI recruitment platform: careers page, job-ad consistency, review replies, AI switches and price shape.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: the AI Act text as amended by the Digital Omnibus on AI on the Commission's AI Act Service Desk (Articles 2, 4, 5, 6, 25, 26, 50, 99, Annex III), the Service Desk implementation timeline and Digital Omnibus FAQ, and the Commission's Guidelines on prohibited AI practices (4 February 2025), all read 4 November 2026.