EU AI Act for Recruitment Agencies: Deployer, Provider and Screening Duties
Agencies can be deployers, providers, or both at once. This guide covers Annex III(4) screening, Article 26 duties, chatbot transparency and vendor contracts.
Recruitment agencies sit in an unusual position under the AI Act: the same firm can screen candidates with a licensed tool this morning and supply a branded shortlisting service to a client this afternoon, and the obligations shift with the role. This guide to the EU AI Act recruitment agencies have to operationalise covers the deployer-versus-provider split, the Annex III(4) high-risk trigger for candidate screening, the Article 26 duties that follow, chatbot transparency, and what to demand from vendors — written for a 300-person staffing agency with multiple client contracts.
Deployer or provider: the role test that decides your duties
The Act works on roles, not on who owns the software. Two matter here, and one firm frequently holds both at once.
- Deployer (Article 3(4)). You use a system under your own authority in a professional capacity — the default for an agency that licenses a screening tool for live vacancies.
- Provider (Article 3(3)). You develop a system, put your own name on it, or modify its intended purpose. This is the heavyweight role: documentation, conformity assessment, CE marking and registration follow.
The employment agency AI Act deployer role is where most firms start, but the table below is the map that matters:
| Agency role | Scenario | Resulting duties |
|---|---|---|
| Deployer | Consultants use a licensed CV screener for your own vacancies | Article 26: oversight, input-data quality, logs, notification, AI literacy |
| Deployer, client-controlled | A client’s ATS forces its own AI ranking on your submissions | Same duties — record in writing which party performs each; do not assume the client absorbs them |
| Provider by branding | You white-label a screening tool and sell it as your own | Full provider burden: Annex IV documentation, quality management (Article 17), CE marking, registration |
| Provider by modification | You retrain a tool on your own placement data so it scores candidates differently | Substantial modification can move the provider role to you (Article 25) |
| Transparency only | A vacancy chatbot answers questions but ranks nobody | Article 50 disclosure only; no high-risk machinery |
Ignore the bottom two rows and your “proprietary matching engine” makes you a manufacturer of high-risk AI.
When candidate screening becomes high-risk under Annex III(4)
Annex III, point 4 covers employment. Its first limb is explicit: recruitment or selection of natural persons, including placing targeted job advertisements, screening and filtering applications, and evaluating candidates. If your tool does any of that, third party recruitment AI high risk EU obligations attach to your agency as deployer — the vendor’s marketing label (“just a productivity aid”) is irrelevant.
Normally high-risk for an agency: ranking or shortlisting CVs against role criteria; knock-out rules that automatically reject applicants; AI scoring of video interviews or assessments that feeds a selection decision; audience targeting that filters who is shown a vacancy.
Normally outside Annex III(4): scheduling and interview booking; proofreading or formatting a CV the candidate already wrote; aggregate workforce analytics that never evaluates an identifiable person; internal bid modelling with no candidate-level effect.
The grey zone agencies get wrong: “fit scores”, red-flag tools that flag employment gaps, engagement models predicting who will accept an offer. These evaluate candidates, so treat them as high-risk unless you can argue otherwise on paper. Annex III’s narrow exemption covers purely preparatory or procedural tasks that cannot influence the decision — and once a consultant reads the score and lets it shape the shortlist, it collapses.
A 300-person agency pointing a screener at 4,000 applications for a food-processing client, where recruiters only interview the top band, is running a high-risk deployment. The agency — not the software house — carries the duties.
The Article 26 duties your agency has to operate
Article 26 is written for the organisation using the system, not the one that built it. The AI candidate screening agency obligations reduce to six operational practices:
- Human oversight with teeth. Name trained overseers who can disregard, reverse or stop an output — and judge them on using that authority, not on clearing a queue.
- Vetted input data. Feed the tool job-relevant material: skills, availability, right to work. Employment gaps, postcode proxies and name-derived characteristics are what turn screening into a discrimination claim.
- Monitoring and escalation. Track live behaviour, route problems to the vendor, report serious incidents to the market surveillance authority, and date every review.
- Log retention. Keep automatically generated logs for at least six months as the Regulation provides, so you can reconstruct why a candidate ranked where they did — “the model decided” is not an answer.
- Worker notification. Tell affected workers and their representatives before high-risk AI goes live, including placed staff your client’s process will screen.
- AI literacy (Article 4). Every consultant touching the tool needs training proportionate to their role: what the score means, what it cannot mean, when to override it.
None of this is optional: AI recruitment tools compliance EU agencies must evidence each practice from the moment a system influences a real candidate’s outcome.
When a staffing agency becomes a provider
Article 25 moves provider responsibilities down the value chain. Four triggers matter to recruitment firms:
- Your name on the box. You brand the system and supply it to a client. The client bought your screening product.
- Substantial modification. You retrain the tool on your own placement data, or add scoring logic the original provider never foresaw.
- Changed purpose. You repurpose an analytics tool to rank external candidates; the new purpose is Annex III(4), and obligations attach to whoever presents the system as theirs.
- Holding out as provider. Contracts, labelling or marketing that would reasonably lead a client to believe you are the provider.
The consequences are the full provider stack: technical documentation, quality management, conformity assessment, CE marking, an EU declaration of conformity, post-market monitoring and database registration — a multi-quarter programme for a mid-market agency, not a procurement update.
Keep the licence in the vendor’s name and their brand on the interface; market your expertise, not your ownership of the model. “Our proprietary AI shortlisting” wins RFPs, but every word of it is a liability unless the paperwork is yours to produce.
Job ads, chatbots and candidate data transparency
Targeted job advertisements are a named high-risk use. Annex III(4) names placing targeted job advertisements alongside screening itself, so AI act job advertisement platforms compliance is not hypothetical for agencies running paid campaigns. Where the platform decides who sees a vacancy from inferred candidate attributes — location models, “likely to be interested” audiences — treat the tooling as high-risk and document the targeting logic. A budget spent on contextual placements, with no candidate-level evaluation, sits lower down the risk curve.
Candidate chatbots must disclose themselves. Article 50 requires systems interacting with natural persons to make clear that the person is dealing with AI, unless that is obvious and the interaction is trivial. A bot collecting availability for warehouse shifts is not trivial: disclose at the first interaction, keep the notice visible, and offer a route to a human. This is where the recruitment agency chatbot candidate data AI Act question lands — and it stacks on your GDPR transparency duties rather than replacing them.
What you must never buy. Article 5 bans two recruitment-shaped practices outright: emotion inference in the workplace — video tools scoring “engagement” or “confidence” from a candidate’s face — and untargeted scraping to build facial-recognition databases from the open web. A prohibited-practice finding does not depend on harm occurring; using the system is the breach.
What to push to vendors — and settle with clients
Because one agency serves many clients, every duty multiplies by engagement. Four moves keep it manageable:
- Get the instruction set in writing. The provider must say what the system is for, what it is not for, and what oversight it assumes. This is where staffing agency AI Act obligations and procurement meet: no written instructions, no deployment.
- Ask for evidence, not assurances. Conformity status, registration entry, and how the vendor handles serious incidents and authority requests. A dashboard screenshot is not evidence.
- Allocate each duty per client. Where a client’s ATS dictates the scoring, state who operates oversight, who keeps logs, and who answers candidates. Silence defaults to your firm carrying all of it.
- Keep a per-client file. Classification rationale, oversight owner, training record, vendor evidence, review date — refreshed at every renewal.
How to check your own position
Start with an inventory you can finish in an afternoon: every tool that touches a candidate decision — CV parsers, assessments, sourcing tools, interview analysis, job-ad campaigns, vacancy chatbots — including free tiers and browser extensions. For each one, note who set its purpose and settings: you, or the client. That single answer fixes your role.
Then test each tool against Annex III(4). Does it screen, filter, rank or evaluate candidates, or place targeted job adverts? If yes, mark it high-risk and list the Article 26 duties you can evidence today against those you cannot. Tools that only talk to candidates go on a transparency list instead. Finally, compare duties owed with duties evidenced, engagement by engagement, and you have a dated, owned gap list.
That one-page position statement per client is what turns vendor renewals, client questionnaires and regulator requests into one conversation rather than three emergencies. Most agencies find the gaps are not in the technology — they are in who was assumed to be doing what.
Frequently asked questions
What does the EU AI Act require of recruitment agencies?
Screening, ranking and targeted job-ad tools are high-risk under Annex III(4): run them with trained human oversight, vetted input data, six-month log retention, worker notification, monitoring and AI literacy for consultants. Candidate-facing chatbots must disclose that they are AI under Article 50. Brand or modify a tool for clients, and provider duties attach instead.
Is CV screening software always high-risk?
If it screens, filters, ranks or evaluates candidates for a role, yes — Annex III(4) names those activities. Tools that only schedule interviews, format CVs or report aggregate analytics are normally outside. Grey-zone tools producing “fit scores” or predicting who will accept an offer should be treated as high-risk unless you can justify the classification otherwise in writing.
Do we become a provider if we resell a tool to clients?
You become a provider if you put your own name on it, substantially modify it, or change its purpose so high-risk use follows — Article 25 is explicit. Keep the vendor’s licence and brand in place and you remain the deployer. Selling a white-labelled screener as your own product brings documentation, conformity assessment, CE marking and registration onto your desk.
Who is the deployer when a client’s ATS does the ranking?
Whoever decides how the system is used and for what purpose. If the client configures and operates the ranking, they are the deployer; if you set the criteria and apply the output, you share the duties. Put it in the contract — an unallocated duty lands on whoever operates the tool on the day, usually your consultant.
Put this guide to work on your own estate.
Run the free exposure scan to see which of these articles actually apply to your company — 12 questions, four minutes, no card. Confirm your email and the full report unlocks.
- Free forever plan
- No credit card
- EU-hosted · GDPR-native
Build a live AI register in minutes — upgrade only when you need the audit pack.
Keep reading
AI in healthcare under the EU AI Act: clinical decision support and high-risk duties
11 min
Risk & ClassificationThe Annex III High-Risk AI Systems List: All Eight Use Cases Explained with Examples
11 min
Risk & ClassificationThe Eight Prohibited AI Practices Under Article 5: A Compliance Checklist
10 min