EU AI Act voice agents: Article 50 transparency rules

Under the EU AI Act, voice agents must inform callers they are talking to an AI starting 2 August 2026. Here is what Article 50 means for phone operations.

## EU AI Act voice agents: the direct answer on transparency Under Regulation (EU) 2024/1689, EU AI Act voice agents must tell callers they are speaking with an artificial intelligence. This transparency duty applies to inbound customer service and outbound phone calls across the European Union starting 2 August 2026. Organisations running automated phone agents must verify technical and operational controls: * Clear disclosure at the start: Article 50(1) and 50(5) require informing the caller at the latest during the first interaction. * Voice cloning and synthetic audio: Disclosing when a synthetic voice mimics a real person under Article 50(4). * Emotion recognition restrictions: Following the ban on emotion detection in workplace and educational settings under Article 5(1)(f), and providing clear notice if used elsewhere under Article 50(3). * High-risk classification checks: Verifying whether the workflow performs automated scoring or recruitment filtering governed by Annex III. * Audit trails: Maintaining transcripts and call logs to verify execution of required opening phrases. * Parallel compliance: Meeting requirements under national do-not-call registries and the GDPR alongside the AI Act. This guide provides practical information and not legal advice, and the classification of a use case must be checked with a lawyer or DPO. ## AI Act timeline for voice agents The EU AI Act phases in obligations between 2024 and 2028. The Digital Omnibus on AI, Regulation (EU) 2026/1744 (in force since 27 July 2026), did not postpone Article 50: it only gives providers of generative AI systems already on the market before 2 August 2026 until 2 December 2026 for the Article 50(2) machine-readable marking. | Date | What applies | What it means for a phone agent | | :--- | :--- | :--- | | 1 August 2024 | Regulation (EU) 2024/1689 enters into force | The legal baseline is fixed across all EU member states. | | 2 February 2025 | Prohibited practices (Article 5) | Emotion recognition is banned in workplaces and educational institutions. | | 2 August 2026 | Article 50 transparency obligations | AI disclosure requirements take effect. The agent must state it is an AI upon answering or initiating the call. | | 2 December 2026 | Article 50(2) marking deadline | Providers of generative systems placed on the market before 2 August 2026 must ensure synthetic audio is marked in a machine-readable format. | | 2 December 2027 | Annex III high-risk systems | Strict compliance regimes take effect for high-risk use cases such as candidate selection and creditworthiness scoring. | | 2 August 2028 | Annex I high-risk products | High-risk obligations take effect for systems integrated into products governed by EU harmonisation legislation. | ## Provider and deployer: who does what The AI Act establishes distinct responsibilities for software providers and deployers. The **provider** is the technology vendor developing the foundational software or platform. The provider must design the system with technical capabilities that support disclosure. Under Article 50(1), the platform must allow messages that signal an artificial intelligence is operating. Under Article 50(2), the provider must also implement technical measures so synthetic audio can be detected as artificially generated. The **deployer** is your business using the voice agent to answer customers or call leads. As a deployer, you configure the assistant. You write prompt instructions, select the voice model, determine the call objective, and set the greeting spoken to callers. When deploying an AI phone agent, your team remains responsible for ensuring the live conversation opens with the required transparency statement. ## How to say it is an AI in the first seconds of a call Article 50(5) states that transparency information must be provided in a clear and distinguishable manner at the latest at the time of the first interaction. For AI Act chatbot transparency on phone channels, this notice belongs in the opening sentence. A compliant greeting covers four elements: 1. The identity of your organisation. 2. The statement that the caller is speaking with an AI assistant. 3. The purpose of the call. 4. How the caller can reach a human staff member. Here is an opening for a commercial logistics company: > "Good morning, you have reached Northstar Logistics. I am an automated AI assistant scheduling parcel deliveries for our dispatch team. You can speak to me naturally, or ask for an operator at any point during this call. How can I assist you today?" The same care applies to data: the assistant of a multi-specialty medical centre qualifies the request without diagnosing and reads back a summary that confirms the appointment. This statement covers the disclosure duty of Article 50 because the caller understands the machine nature of the agent before sharing private information. If the caller asks for human assistance or presents a complex edge case, routing them requires a tested voice agent human handoff system configured directly in your call routing logic. At DeepAgent, companies configure this opening phrasing inside the dashboard prompt editor. In that campaign, for Italian insurance agency Unipol Cerignola, the AI phone assistant Adriana handled 725 inbound calls in 2 months, classifying them into six main cases, with 200 successful transfers to agency staff out of 207 attempts (96.6%). ## Voice cloning, synthetic voices, and deep fakes Telephony platforms provide pre-built synthetic voices that sound natural without mimicking specific individuals. Article 50 treats synthetic speech and impersonation under different standards. Article 50(4) states that deployers of an AI system that generates or manipulates audio resembling existing persons and appearing authentic must disclose that the content is artificially generated. Simulating a real person, such as an executive or specific company employee, without clear disclosure creates direct legal exposure. The practical rule for contact centers is simple: select standard synthetic voices that do not copy living people without explicit written consent. For general synthetic audio, Article 50(2) places the burden on providers to mark output in a machine-readable format. Under the Digital Omnibus on AI, Regulation (EU) 2026/1744 (in force since 27 July 2026), providers of generative AI systems already on the market before 2 August 2026 have until 2 December 2026 for the Article 50(2) machine-readable marking. ## Emotion recognition and sentiment Some contact center platforms offer real-time emotion recognition to detect frustration, tone, or stress levels in a customer's voice. The AI Act regulates these technologies strictly. Article 5(1)(f) created an outright ban on emotion recognition systems used in the workplace and in educational institutions, effective since 2 February 2025, except where installed for strictly medical or safety reasons. Deployers cannot run software that evaluates employee stress, frustration, or sentiment during internal calls or employee performance monitoring. When applied to external consumers, Article 50(3) requires deployers of emotion recognition systems to inform exposed individuals that the system is operating. Avoid deploying real-time voice emotion tracking on staff, and consult your DPO before turning on sentiment-scoring features on incoming caller audio. ## When a use case could become high-risk A phone agent that takes messages, qualifies requests and books appointments in general does not fall in the high-risk list, and the classification of a specific use case must be checked with a lawyer or DPO. Under Annex III, AI systems become high-risk when applied to sensitive evaluation processes. The compliance deadline for Annex III stand-alone systems is 2 December 2027. Two common commercial scenarios touch Annex III boundaries: * **Credit evaluation**: Systems that evaluate the creditworthiness of natural persons or establish credit scores. * **Recruitment and selection**: Systems used for the recruitment or selection of natural persons, particularly for advertising vacancies, screening applications, and evaluating candidates. A consumer-credit company used the agent Klio to qualify interest without giving economic estimates and to book appointments with advisors, which is a different use case from scoring creditworthiness. An insurance company completed 685 call attempts and qualified 36 agencies through B2B calls for its network, not hiring of employees. To keep an agent strictly within administrative boundaries, implement guardrails against AI voice hallucinations so the model never volunteers automated credit assessments or hiring decisions. ## Fines for Article 50 transparency breaches Failing to comply with transparency requirements carries significant financial liability. Under Article 99 of Regulation (EU) 2024/1689, non-compliance with Article 50 transparency provisions can result in administrative fines of up to EUR 15 million or 3% of total worldwide annual turnover for the preceding financial year, whichever is higher. For SMEs and start-ups, the cap is the lower of the two amounts. The European Commission published guidelines on the Article 50 transparency obligations. ## Checklist for Article 50 voice agent compliance Companies should run the check now using this operational checklist for phone workflows: 1. Build an internal inventory: Map every phone number, conversational agent, and telephony vendor operating in your business. 2. Standardise opening disclosures: Draft disclosure phrasing for each inbound and outbound campaign, stating that the caller is speaking with an AI. 3. Conduct audio test calls: Run test calls across all configured routing paths to verify disclosure occurs within the first seconds, before personal data is gathered. 4. Review synthetic voices: Ensure selected voice models are standard synthetic personas without cloning real individuals. 5. Disable workplace emotion detection: Audit internal call monitoring tools to ensure no systems evaluate employee vocal sentiment or emotional state. 6. Inspect transcripts and operational logs: Set up transcript archiving for auditing. An inbound reception assistant qualifies the request without diagnosing and reads back a summary confirming the appointment. 7. Coordinate with your DPO: Review disclosures alongside your privacy notices, verifying lawful basis under our guide on GDPR and conversational AI. 8. Check outbound registries: When placing outbound calls, confirm compliance with national marketing consents and do-not-call registries as detailed in our guide to telemarketing consent and AI calls. 9. Audit vendor capabilities: Ask providers how they manage audio records, API data flows, and machine-readable markers using our guide on choosing an AI voice agent provider. ## DeepAgent platform plans and pricing DeepAgent provides infrastructure for deploying conversational phone assistants with custom opening disclosures and full transcript records. Platform subscription tiers on DeepAgent pricing include: * Starter: EUR 99 per month * Pro: EUR 299 per month * Growth: EUR 500 per month * Scale: EUR 799 per month * Enterprise: Custom quote Prices exclude VAT, annual billing -17%. To try it yourself, create an agent from the signup: 30 minutes free, no card. ## Frequently asked questions ### Does an AI voice agent have to say it is an AI? Yes. Under Article 50(1) and Article 50(5) of Regulation (EU) 2024/1689, AI systems that interact directly with natural persons must inform them that they are speaking with an AI. This disclosure must be given in a clear and distinguishable manner at the latest at the time of the first interaction. ### When does the AI Act apply to chatbots and voicebots? The Article 50 transparency obligations for chatbots and voicebots apply from 2 August 2026. Providers of generative AI systems already placed on the market before that date have until 2 December 2026 to implement machine-readable marking under Article 50(2). ### Was Article 50 postponed by the Digital Omnibus? No. The Digital Omnibus on AI, Regulation (EU) 2026/1744 (in force since 27 July 2026), did not postpone Article 50. It only deferred the machine-readable marking deadline under Article 50(2) to 2 December 2026 for existing systems, while postponing high-risk rules under Annex III to 2 December 2027. ### Are AI phone agents high-risk under the AI Act? A phone agent that takes messages, qualifies requests and books appointments in general does not fall in the high-risk list, and the classification of a specific use case must be checked with a lawyer or DPO. Voice AI becomes high-risk if it performs specific tasks listed in Annex III, such as evaluating creditworthiness to establish credit scores or evaluating candidates during recruitment. High-risk obligations apply from 2 December 2027. ### What are the fines for not disclosing AI? Under Article 99 of Regulation (EU) 2024/1689, infringements of Article 50 transparency rules can result in administrative fines of up to EUR 15 million or 3% of total worldwide annual turnover, whichever is higher. For SMEs and start-ups, the cap is the lower of the two amounts. ## Preparing your phone workflows for European compliance Companies should run the check now to review call scripts, document human handoff pathways, and confirm vendor data handling practices. Talk to sales → · Create your agent →