Summary
From 2 August 2026, the transparency requirements in Article 50 of the EU AI Act are due to apply. Businesses using AI phone assistants should plan now how callers will be clearly told at the start of a conversation that they are speaking with AI. Providers of generative AI systems also have to machine-readably mark certain synthetic audio outputs. Other requirements depend on the system’s task: an assistant that gives opening hours does not become high-risk simply because it answers the phone, while AI used to assess or select job applicants may fall under the high-risk rules.
The AI Act applies in stages
The AI Act does not have a single application date. Chapters I and II, including the duty to promote AI literacy and prohibitions on certain AI practices, have applied since 2 February 2025. Further provisions have applied since August 2025. Article 50 transparency duties and most of the Regulation apply from 2 August 2026.
For businesses using voice AI, three questions are a useful starting point:
- Does the system interact directly with a natural person?
- Does it generate synthetic audio or imitate a real person?
- What task does it perform, and does it make or influence decisions about people?
Different answers can trigger different obligations. The system’s actual function and intended purpose matter, not simply the fact that it makes phone calls.
1. Tell callers they are interacting with AI
Article 50(1) requires providers of AI systems intended for direct interaction with people to ensure that people are informed they are interacting with AI. An exception applies when this is obvious to a reasonably well-informed, observant and circumspect person in the circumstances and context. The information must be clear and distinguishable no later than the first interaction.
During an incoming phone call, it may not be obvious that AI has answered. A clear greeting is therefore a robust practical approach, for example:
Hello, you’re speaking with the AI phone assistant for [company]. How can I help?
The Act does not prescribe a particular sentence. The notice should be understandable and should not appear only after a caller asks. If callers request a person or their request falls outside the approved scope, the workflow should offer a suitable transfer or callback option.
Providers and deployers have different roles. A provider develops the system or places it on the market under its own name. A business using the assistant in its operations is typically the deployer. Clarify across the supply chain who owns the greeting, configuration and updates, and how the notice will be delivered in real calls.
In our VoiceBooker Wizard, you set the assistant’s greeting and can configure transfers to staff for specific requests. Test the complete call: is the AI notice audible at the beginning, and does the handover work? Our Wizard quick start explains the setup; our call-transfer guide covers handovers. Configuration helps implement the workflow, but does not replace a legal assessment of your specific use.
2. A synthetic voice is not automatically a deepfake
Article 50 contains separate transparency requirements. First, people should know when they interact directly with AI. Second, providers of AI systems that generate synthetic audio, image, video or text content must generally mark outputs in a machine-readable format and make them detectable as artificially generated or manipulated. Technical solutions should be effective, interoperable, robust and reliable as far as technically feasible. The Act includes exceptions, such as assistive functions for standard editing where the input data or its meaning is not substantially changed.
A standard synthetic phone voice is not automatically a deepfake. The AI Act describes deepfakes as AI-generated or manipulated image, audio or video content that resembles existing people, objects, places, entities or events and could falsely appear authentic or truthful. A cloned voice designed to sound like a particular employee or public figure therefore raises different questions from a generic synthetic voice.
The caller-facing disclosure and the technical marking of audio outputs are two different requirements. Ask the provider how machine-readable marking is implemented in the specific system. The marking duty primarily applies to the provider of the generative system; the notice about direct AI interaction must also reach callers in the real use context.
3. The task determines risk, not the phone
An AI phone assistant is not high-risk simply because it speaks, uses a language model or processes personal data. The intended purpose and possible effects on people are decisive.
An assistant that gives opening hours, records callback requests or coordinates appointments under approved rules is not automatically high-risk. AI used for recruitment and applicant selection can be different: these uses are listed as high-risk in Annex III. Whether a particular system falls within the category or qualifies for an exception must be assessed against its actual purpose and role in decision-making.
Voice analysis also needs separate review. The AI Act generally prohibits emotion recognition in the workplace, with narrow exceptions for medical or safety reasons. A feature that claims to infer employees’ or applicants’ emotions from their voices should not be introduced as a harmless extra analysis.
Under the AI Act timetable, relevant requirements for high-risk systems under Article 6(2) and Annex III apply from 2 August 2026. Corresponding obligations for certain product-related high-risk systems under Article 6(1) and Annex I apply from 2 August 2027.
4. Staff need appropriate AI literacy
Article 4 has applied since 2 February 2025. Providers and deployers must take measures to promote AI literacy among staff and other people who operate or use systems on their behalf. Measures should fit people’s roles, prior knowledge, use case and the potential consequences of errors.
For a service team, a short, use-case-specific briefing may be appropriate. Staff should understand which tasks the assistant may and may not perform, how to handle transferred requests and report errors, which answers or promises it must not make, when a person must take over, and what information may appear in transcripts, summaries or CRM records. Document measures in a way that fits your business and deployment.
5. The AI Act does not replace privacy or telephone law
Saying “you are speaking with AI” does not automatically fulfil data-protection duties. Names, phone numbers and call content may still be personal data. Legal basis, purpose, data minimisation, retention, access, processing agreements and caller information must be assessed separately. See our GDPR checklist for AI phone assistants.
An audio recording is also different from live speech processing. In Germany, an unauthorised recording of a private conversation may engage Section 201 of the Criminal Code. Review recording as a separate processing step. Using a voice assistant is not blanket permission to record calls.
Business checklist
Before deploying or updating an AI phone assistant:
- Map the use cases: What calls does it answer, what actions does it take and does it influence decisions about people?
- Clarify roles: Who is the provider, who is the deployer and who owns the greeting, configuration and changes?
- Make the first interaction transparent: Give a clear, audible AI notice at the beginning.
- Check audio outputs with the provider: Ask about machine-readable marking of synthetic speech and technical evidence.
- Review high-risk and prohibited uses: Pay particular attention to recruitment, employee assessment, emotion recognition and decisions with significant effects.
- Train staff: Explain responsibilities, system limits, error routes and human handovers.
- Review privacy separately: Document data flows, recordings, transcripts, retention, providers and access rights.
- Monitor changes: New functions may change the legal assessment.
Conclusion
For many businesses, the most visible effect of the AI Act from August 2026 will be a clear notice at the beginning of an AI phone call. Providers of generative systems must also consider marking synthetic outputs. Further requirements depend on whether the assistant handles routine communication or performs sensitive or decision-relevant tasks. Start with a concrete list of tasks, clarify provider and deployer roles, and reassess changes to functionality.
This article provides general information as of 30 June 2026. It is not a legal assessment of a particular system or deployment.