A chatbot that talks to job candidates under your company's or your agency's brand carries the Article 50(1) disclosure duty like any other customer-facing bot: candidates must be told they're dealing with AI at or before the first interaction, in the interface, not in a linked privacy policy. Candidates are natural persons, an application process is nobody's idea of "obviously AI" from context, and the power imbalance in hiring is exactly why regulators would look here early. If you brand and offer the bot (even built on a vendor platform or third-party model), the provider-side duty is generally yours.
But Article 50 is the smaller half of this question, and an honest page has to say so. AI systems used in recruitment and selection (screening or filtering applications, evaluating candidates, ranking, deciding who advances) are named in the EU AI Act's high-risk category (Annex III, employment). That's a different and much heavier regime than a transparency notice: think risk management, data governance, human oversight, conformity obligations. The Digital Omnibus moved Annex III high-risk application to December 2, 2027, so it isn't the August 2026 deadline. But "later" is not "lighter".
The line for your analysis: a bot that informs and assists (answers questions about the role, collects documents, books interviews) is transparency territory; disclose and you're in good shape. A system that evaluates or sorts people is potentially high-risk territory, and that analysis belongs with counsel, not a template. Many real deployments blur the two (a "chatbot" that scores answers and screens people out is a screening system with a chat interface), so classify by what the system does to the candidate's chances, not by its UI.
Don't forget the adjacent trap on this page's edge: emotion recognition in interviews or video-analysis tools that infer traits sit under Article 50(3) notice duties and worse. Another escalate-to-counsel flag, not a checklist item.
For the plain candidate-FAQ bot, compliance today is the usual cheap pair: first-message AI notice plus persistent badge, dated screenshot in your evidence file, and, given what's coming in 2027, a one-line register entry recording that this system informs rather than evaluates. That single sentence of classification, written down now, is the cheapest high-risk-readiness step available.
Not sure how this applies to your specific setup? The free 2-minute check tells you which of the four Article 50 duty areas likely apply to your company. No email required.
Take the free check Generate disclosure text