← HR software guides
HIGH RISK — Annex III Cat.4 (screening) Art.50 — chatbot disclosure required

Double obligation — unique to conversational recruiting AI

Paradox Olivia is both a chatbot (Article 50 disclosure) and a candidate screening tool (Annex III high-risk). Most HR tools trigger one obligation; Olivia triggers both simultaneously. Each must be met independently.

Deadline: 2 August 2026. Both obligations apply from this date. Olivia's chatbot disclosure must be configured before then.

Paradox Olivia EU AI Act Compliance: The Dual Obligation Guide

Paradox Olivia is a conversational AI recruiter that chats with candidates, screens them, schedules interviews, and moves qualified applicants forward. It is the only widely deployed HR tool that simultaneously triggers the EU AI Act's chatbot transparency rules and its high-risk AI screening obligations.

Olivia features and EU AI Act obligations

What Olivia doesClassificationObligation
Chats with candidates via website, SMS, WhatsAppArt.50 — LIMITED RISKMust disclose AI nature at conversation start
Asks screening questions, collects availabilityHIGH RISKAnnex III Cat.4 — collecting data used in employment decision
Schedules interviews only for qualifying candidatesHIGH RISKAI gatekeeping access to next stage = employment screening
Sends offer letters, onboarding communicationsNOT HIGH RISKPost-decision communications, no AI assessment

Article 50 — Olivia chatbot disclosure

Configure this as Olivia's first message before any candidate input:

Hi! I'm Olivia, an AI recruiting assistant for [Company]. I'll help with your application and can schedule interviews. You're chatting with AI — a recruiter is available if you prefer. Ready to get started?

Full deployer obligations checklist

Compliance pack for conversational AI recruiting

Covers both the chatbot disclosure (Article 50) and the high-risk screening side (Article 26): Olivia opening message templates in 7 EU languages, candidate screening disclosure, human oversight SOP, log retention policy.

Get the HR Compliance Kit — €99 →

Why Olivia is uniquely complex under the EU AI Act

Most recruiting tools are asynchronous: a candidate submits a CV, software scores it overnight, a recruiter sees the score the next morning. Olivia is real-time and conversational — candidates interact with it directly, often not knowing it is AI. This creates the Article 50 obligation immediately.

At the same time, Olivia's conversations are structured screening — it asks qualifying questions, assesses answers, and makes autonomous scheduling decisions (or withholds scheduling from candidates who do not qualify). This is employment screening. The conversation is not neutral data collection; it is AI-based gatekeeping. That is Annex III Category 4.

The scheduling decision as screening

A subtle but important point: Olivia's decision about who gets an interview slot and who does not is itself a high-risk AI decision. Even if Olivia is just asking availability questions and forwarding everyone who responds — that automatic forwarding decision is the AI taking an action that affects who progresses through a hiring process. If the forwarding is conditional on any screening criterion (keyword in a response, score threshold, availability match), that conditionality is high-risk screening.

Frequently asked questions

We use Olivia only for scheduling — no screening questions. Is it still high-risk?
If Olivia sends interview invites to every candidate who responds regardless of any criteria — purely a scheduling logistics tool — the high-risk classification is weaker. You still need the Article 50 chatbot disclosure. But if any filtering or qualification check happens (even "is the candidate available on the proposed dates?"), and scheduling proceeds only for those who respond correctly, that filtering step is high-risk AI gatekeeping.
Candidates think Olivia is a human recruiter named "Olivia." Does that violate the AI Act?
Yes, if Olivia is not configured to disclose its AI nature. Article 50(1) prohibits AI systems from being "designed or deployed in such a way as to deceive the natural persons they interact with into thinking they are communicating with a human being." Using a human-sounding name with no AI disclosure — especially when candidates reasonably assume they are talking to a recruiter — is precisely the deceptive design the Article 50 prohibition targets.