US voice-AI founders treat "Europe" as one regulatory block. It is not. A voice agent deployed across the EU faces two stacked layers: a supranational floor that applies identically in every member state (the EU AI Act, the GDPR, the ePrivacy rules), and a country mosaic on top of it where the same recorded call is a routine business practice in Dublin and a criminal offence in Berlin. Get the floor right and you are still exposed if you ignore the mosaic.

This guide maps both layers as of June 2026. Read it the way you read the US founder's guide: a handful of rules everyone must follow, plus a fast-moving national patchwork that determines your real exposure. The single date to anchor on is 2 August 2026, when the AI Act's transparency duty (you must tell callers they are talking to a machine) becomes enforceable across all 27 member states.

Disclaimer: This article is for informational purposes only and does not constitute legal advice. EU and national rules change weekly, and several instruments cited below are at provisional or draft stage. Verify current status before relying on anything here, and consult a qualified attorney for your situation.

How to read this guide: the jurisdiction stack

Every European country a voice agent touches sits inside a stack of overlapping layers. For voice-AI regulation, four layers matter. The others (eurozone, Schengen) add nothing to your voice compliance and are noise.

1. EU AI Act (Regulation (EU) 2024/1689) – directly applicable, identical text in every member state.
2. GDPR (Regulation (EU) 2016/679) – directly applicable, identical text.
3. ePrivacy (Directive 2002/58/EC) – a Directive, so each country transposed it into its own national law. This is where call-recording and robocall rules start to diverge.
4. National law – each country's own criminal code, telecoms act, consumer law, and language law. This is the mosaic.

The EEA states (Norway, Iceland, Liechtenstein) apply the GDPR in full but, as of June 2026, the AI Act is not yet incorporated into the EEA Agreement. The wider non-EU markets (the UK, Switzerland, Turkey, Ukraine) sit outside this stack entirely and are covered in a separate guide.

The rest of this article is in three parts: the floor, a country-by-country call-recording reference table, and the national divergences that change your exposure.

Part 1: The EU floor (applies in every member state)

1. AI Disclosure (EU AI Act, Article 50)

Requirement

Article 50(1) of the AI Act requires that any AI system interacting directly with a person be designed so the person is informed they are interacting with an AI unless that is obvious. For a voice agent this means the disclosure must come at the start of the call, in a clear and distinguishable form. This is the rule written most directly for voice bots.

Minimum Action

• Open every call with an explicit statement that the caller is speaking to an automated AI system.
• For any synthetic or cloned voice you generate, mark the output as artificially generated in a machine-readable form (Article 50(2)), and disclose manipulated audio to the listener (Article 50(4)).
• If your agent infers caller emotion or sentiment, inform the caller that an emotion-recognition system is operating (Article 50(3)).

Why It Matters

Breach of Article 50 sits in the AI Act's middle penalty tier: up to €15 million or 3% of worldwide annual turnover, whichever is higher (Article 99). For SMEs the lower of the two applies.

Dates and what changed (2024–2026)

• 1 Aug 2024: AI Act entered into force.
• 2 Feb 2025: prohibited practices (Article 5) and AI-literacy duties apply.
• 2 Aug 2025: general-purpose AI model obligations and the governance and penalty framework apply.
• 2 Aug 2026: general application, including Article 50 transparency. This is the load-bearing date for voice founders.
• Digital Omnibus on AI: tabled 19 Nov 2025, with a provisional political agreement reached in May 2026. It delays the high-risk obligations (standalone Annex III systems pushed toward Dec 2027, embedded systems toward Aug 2028) and grants a grace period for synthetic-content marking of systems already on the market (toward Dec 2026). Critically, Article 50 transparency stays at 2 Aug 2026. Do not let the high-risk delay lull you into postponing the disclosure work. These Omnibus dates were provisional as of June 2026; confirm against the final Official Journal text.

2. Prohibited Practices (EU AI Act, Article 5)

Requirement

Article 5 lists outright bans, enforceable since 2 February 2025. Two touch voice analytics directly:

• Article 5(1)(f): emotion-recognition AI in the workplace and education is banned (narrow medical and safety exceptions aside). If you sell voice analytics that infer employee emotion from call audio, that is a prohibited practice.
• Article 5(1)(g): biometric categorisation that infers race, political opinions, trade-union membership, religion, sex life, or sexual orientation is banned.

Why It Matters

Prohibited-practice breaches carry the top tier: up to €35 million or 7% of worldwide annual turnover (Article 99). This is the most expensive mistake in the Act, and it is already live.

3. Personal Data and Voiceprints (GDPR)

Requirement

Every voice agent processes personal data: the call audio, the transcript, the caller's identity, and often a voiceprint. The GDPR demands a lawful basis, transparency, data minimisation, and frequently a Data Protection Impact Assessment.

Minimum Action

• Voiceprints are special-category biometric data. When your agent does speaker identification or verification, the voiceprint is biometric data used for unique identification and falls under Article 9. In practice you need explicit consent collected before processing. Merely recording a voice is ordinary personal data; turning it into an identifier triggers Article 9.
• Automated decisions need safeguards. If the agent makes a decision with legal or similarly significant effect solely automatically (auto-denying a service, disqualifying a lead consequentially), Article 22 applies. The CJEU's SCHUFA ruling (C-634/21, 7 Dec 2023) read this broadly: producing a score a third party relies on can itself be a "decision." The logic behind such decisions must be explainable on request through the GDPR's access rights (reinforced in Dun & Bradstreet, C-203/22, 2025).
• Transparency and retention. Provide a clear privacy notice at or before collection (Articles 13 and 14). Keep recordings and transcripts only as long as necessary (Article 5).
• Run a DPIA. Large-scale processing, biometric data, and systematic monitoring each push you over the Article 35 threshold. Voice agents commonly hit several criteria at once.

Why It Matters

Article 9, Article 22, and consent breaches sit in the GDPR's top fine tier: €20 million or 4% of worldwide annual turnover (Article 83). National DPAs have issued repeated multi-million-euro fines for unlawful telemarketing and call handling (see the country sections below).

4. Automated Outreach (ePrivacy, Directive 2002/58/EC)

Requirement

The ePrivacy Directive protects the confidentiality of communications and governs unsolicited marketing. Because it is a Directive, each member state transposed it differently, which is why robocall and call-recording rules diverge nationally.

Minimum Action

• Outbound AI robocalls need prior opt-in consent. Article 13(1) requires prior consent of the subscriber for calls made by automated calling machines. An AI dialer is squarely in scope. For person-to-person calls, Article 13(3) lets each state choose opt-in or opt-out, so the rule changes by country.
• Stack the three regimes. A single outbound campaign must satisfy ePrivacy (consent to call), the GDPR (lawful basis and transparency), and AI Act Article 50 (disclose the AI). Three regimes on one call.

What changed (2025)

The proposed ePrivacy Regulation was formally withdrawn (Commission approved the withdrawal in July 2025), so the 2002 Directive remains the operative law indefinitely. The November 2025 Digital Omnibus proposes moving cookie-consent mechanics into the GDPR, but that is draft.

5. Deepfakes and Voice Cloning

There is no single dedicated EU deepfake statute. Protection is a patchwork: AI Act Article 50(2) and 50(4) (mark and disclose synthetic and manipulated audio), the GDPR (lawful basis for using a person's voice), the Digital Services Act (notice-and-takedown on very large platforms), and national criminal law. One cloned-voice output can trigger several of these at once, and they do not fully cohere. If your product clones voices, build consent-of-likeness and provenance-marking infrastructure now.

Part 2: Call-recording consent by country

This is the single most practically important national divergence for a voice agent. Even where the criminal law permits one-party recording, the GDPR still makes "disclose the recording and the AI use at the start of the call, with a documented lawful basis" the safe baseline everywhere. The genuine hard divergence is the all-party criminal cluster, where a notice is not enough and you need a real consent gate before any audio is stored.

Call-recording consent by country

Coverage note: this table covers all 27 EU member states plus the EEA. The code-article citations are the operative national provisions as of June 2026. A few (the exact Cyprus penalty article, Malta's criminal backing, and Hungary's precise section numbering) were unconfirmed at the time of writing and should be checked with local counsel before launch.

Part 3: The national mosaic (where exposure actually changes)

The countries below add national rules on top of the floor. They are tiered: deep entries for the jurisdictions where divergence materially changes your exposure, compact entries for the rest. For any country not detailed here, the rule is simple: the EU floor applies, the recording regime is in the table above, and no voice-specific national addition of note has been identified.

Germany

Jurisdiction stack: EU member, GDPR and AI Act apply directly, ePrivacy via the TDDDG, national criminal and competition law on top. AI authority: the Bundesnetzagentur (BNetzA) is the designated market-surveillance authority under the draft implementing law (KI-MIG), which the federal cabinet adopted in February 2026 but parliament has not yet enacted; Germany missed the 2 Aug 2025 designation deadline.

Where it bites:

• Recording without all-party consent is a crime. §201 StGB (violating the confidentiality of the spoken word) carries up to three years' imprisonment, and the liability is personal, not just a corporate fine. This is the strictest recording regime in this guide.
• Cold B2C calls require prior express consent under §7 UWG; an AI dialer does not change that.
• Voiceprint for identification without explicit consent breaches GDPR Article 9 (up to €20M / 4%).
• No AI disclosure at call start from 2 Aug 2026 breaches AI Act Article 50 (up to €15M / 3%).

The data-protection conference (DSK) published guidance on AI and data protection (the Orientierungshilfe KI und Datenschutz, May 2024, with separate DSK AI guidance issued in 2025) that governs LLM-based chatbots; there is no voice-specific guidance yet.

France

Jurisdiction stack: EU member, GDPR and AI Act direct, ePrivacy via the Code des postes et des communications électroniques. AI authority is not yet formally designated; a September 2025 draft scheme proposes the DGCCRF as coordinator with the CNIL leading on AI that processes personal data and biometrics.

Where it bites:

• No systematic recording. The CNIL position is that recording every call is disproportionate; quality and coaching recording must be on a sample, with capped retention. Notify the caller of the purpose and the right to object at the start.
• Loi Toubon. Law 94-665 of 4 Aug 1994 mandates French in consumer-facing commercial communications, including software interfaces. A consumer voice agent must be available in French. The offence is criminal, around €750 per violation (€3,750 for a legal person).
• The 2026 cold-calling overhaul. A law of 30 June 2025, in force 11 August 2026, prohibits B2C prospecting calls without prior consent (strict opt-in, replacing the Bloctel opt-out list), restricts allowed time slots, and carries penalties up to €75,000 for individuals and €375,000 for companies. This is the single biggest regulatory change for AI outbound voice in France.

The CNIL is among the EU's most active enforcers. Its decision SAN-2025-011 (27 Nov 2025) fined American Express €1.5 million, partly for disproportionate call recording.

Italy

Jurisdiction stack: EU member, GDPR and AI Act direct, plus the first comprehensive national AI law in the EU, Law No. 132 of 23 September 2025, in force 10 Oct 2025. It designates ACN (cybersecurity agency) as market-surveillance authority and AgID as notifying authority; the Garante keeps full GDPR competence.

Where it bites:

• Voice cloning can be a crime. Law 132/2025 inserted a new offence, Article 612-quater of the criminal code: disseminating AI-falsified or altered images, video, or audio that misleads and causes harm, punishable by one to five years' imprisonment. Cloned voices fall squarely within "audio altered using AI."
• Employee call recording needs a union agreement or labour-inspectorate authorisation under Article 4 of the Workers' Statute (Law 300/1970); the Garante confirmed this for contact centres.
• Telemarketing requires checking the public objection register (RPO) and prior consent.

Italy is also the EU's most aggressive AI enforcer to date. The Garante fined OpenAI €15 million (20 Dec 2024), though the Court of Rome annulled it on jurisdictional grounds in March 2026; it fined Replika's developer €5 million (10 Apr 2025).

Spain

Jurisdiction stack: EU member, GDPR and AI Act direct. Spain created the EU's first national AI supervisory agency, AESIA (Royal Decree 729/2023), now operational with sanctioning powers since 2 Aug 2025. A national AI governance bill (cabinet-approved May 2026, in Congress) would add content-labelling duties and sanctions up to €35M / 7%.

Where it bites:

• The only EU regulator with voice-specific guidance. The AEPD's two-part guidance on AI voice transcription (Part I, 14 Jan 2026; Part II, ~Apr 2026) treats a person's voice as personal data, flags emotion detection as high-risk or potentially prohibited, and requires perceptible in-session indicators (an on-screen notice, light, or periodic tone) plus per-session consent.
• AI cannot be your only channel. Spain's customer-service law (published 27 Dec 2025, comply by ~28 Dec 2026) prohibits relying exclusively on answering machines or AI: a human operator must be available on request, with immediate transfer.
• Co-official languages. Large firms must be able to serve customers in the relevant co-official language (Catalan, Galician, Basque, or Valencian) where the customer is in that territory, not just Castilian.

Ireland

Jurisdiction stack: EU member, GDPR and AI Act direct. Ireland met the 2 Aug 2025 deadline with a distributed model: S.I. No. 366 of 2025, plus the September 2025 additions, names up to 15 competent authorities, including the DPC, the CCPC, and ComReg, with a National AI Office to coordinate (due by 2 Aug 2026).

Where it bites:

• The DPC is the EU's de-facto Big Tech enforcer. Most US tech firms run their EU operations from Dublin, so the DPC is the GDPR one-stop-shop lead for Meta, TikTok, LinkedIn, Google, and others. If your company is established in Ireland, the DPC is your lead supervisor for cross-border processing.
• Recording is one-party, but business use still needs a notice at call start, a privacy notice, and an opt-out.

The DPC issues the EU's heaviest GDPR fines: Meta €1.2 billion (May 2023), TikTok €530 million (May 2025), LinkedIn €310 million (Oct 2024).

Belgium

Jurisdiction stack: EU member, GDPR and AI Act direct. AI authority not yet formally designated; the telecoms regulator BIPT is the intended lead.

Where it bites:

• Effectively all-party recording. Article 124 of the Electronic Communications Act prohibits recording without all participants' consent, and even a participant's use or disclosure is restricted. A beep tone is insufficient.
• Hard regional-language requirement. The language a Belgian company must use is set by the region of its seat: Dutch in Flanders, French in Wallonia, German in the German-speaking area, bilingual in Brussels. The sanction is nullity of the act (confirmed by the Court of Cassation, 3 Feb 2025). A Belgian voice deployment should serve each caller in the regional language.

The Belgian DPA fined a company €86,000 in 2026 over call recording (SWDE, Decision 102/2026).

Compact entries

• Netherlands. One-party recording (Art. 139a Sr). The DPA (AP) and the consumer/markets authority (ACM) issued a joint position on chatbots (2 Oct 2025) that applies to voice agents: always allow human escalation and disclose that the user is talking to a bot. Telemarketing is opt-in. Fine: Clearview AI €30.5M (2024).
• Poland. One-party at criminal level (Penal Code Art. 267), but opt-in consent for outbound marketing under the electronic-communications law (PKE, in force 10 Nov 2024). The Polish Language Act requires B2C dealings in Polish.
• Portugal. Recording is all-party and criminal (Penal Code Art. 199). ANACOM was designated the AI Act authority on 19 Sep 2025. The DPA (CNPD) is a light enforcer.
• Austria. Treat recording as all-party in practice: §120 StGB criminalises recording statements not intended for the recorder, and personality rights plus the GDPR drive an all-party consent expectation. RTR and its AI service centre are the expected AI authority, pending legislation.
• Greece. The strictest recording rule here: all-party and criminal under Penal Code Art. 370A, applying even to a participant. No AI authority designated yet.
• Czechia. One-party (Civil Code §§86/88). The Telecommunications Office (ČTÚ) is the designated lead, pending the national law.
• Romania. One-party (Law 506/2004). ANCOM and ADR are proposed as authorities.
• Finland. AI Act in force in stages, with Traficom as the single point of contact and authority powers from 1 Jan 2026. A party may record, but a business must announce the recording at the start of the call; a website notice is explicitly insufficient.
• Sweden. AI Act implementation is at proposal stage (SOU 2025:101), with PTS proposed as lead rather than the DPA (IMY). One-party recording, but outbound sales need written confirmation (since 1 Sep 2018) and NIX-Telefon screening.
• Denmark. The first member state to pass national AI Act legislation, in force 2 Aug 2025, with the Agency for Digital Government as lead. One-party recording, but QA and training recordings need unbundled, purpose-specific consent (IDA Forsikring ruling, 27 Jun 2024).
• Baltics. Lithuania designated RRT as its AI authority (Jan 2025) and requires customer service in Lithuanian from 1 Jan 2026; voiceprints need explicit consent. Estonia has not designated an authority and treats recording as GDPR-driven, not free one-party. Latvia is one-party, with the general GDPR transparency notice applying.

The EEA states (Norway, Iceland, Liechtenstein)

All three apply the GDPR in full through the EEA Agreement, so treat them like member states for data protection. The critical difference: the AI Act is not yet incorporated into the EEA Agreement as of June 2026. The three sit as observers on the EU AI Board, and the Article 50 transparency duty does not yet bind them, though Norway has signalled rapid national implementation (with Nkom as coordinator). Plan as if it will apply, and confirm the timing per country. Norway runs a useful AI regulatory sandbox and upheld a NOK 65 million Grindr fine in 2025.

National AI authority designation status

A practical signal of where enforcement is ready. As of June 2026:

Designation status does not change the rules (the AI Act applies regardless), but it tells you which national doors enforcement will come through first.

Sector overlays: finance, health, and employment

A voice agent in a regulated sector is rarely caught by one regime. It is usually caught by two at once, and the sector rule is the binding ceiling on top of the floor.

Finance: DORA plus high-risk AI

If your agent serves a bank, insurer, or investment firm, DORA (Regulation (EU) 2022/2554, applying since 17 Jan 2025) treats you as an ICT third-party service provider (Art. 3(19)). When the agent supports a "critical or important function" such as onboarding, payments, or fraud handling, the financial entity must flow obligations down to you by contract: a register of information (Art. 28(3)), a tested exit strategy (Art. 28(8)), mandatory terms including unrestricted audit rights (Art. 30(3)), and subcontracting controls (RTS (EU) 2025/532). A vendor designated a Critical ICT third-party provider faces direct EU oversight and penalty payments up to 1% of average daily worldwide turnover (Art. 35).

Layer the AI Act on top. A voice agent that evaluates creditworthiness (Annex III, point 5(b)) or prices life or health insurance (point 5(c)) is high-risk, and because those uses involve profiling, the Article 6(3) "no significant risk" exemption does not save you. ESMA's May 2024 statement already requires firms to disclose to clients when they interact with an AI chatbot.

Health: medical-device status

A voice agent that does symptom triage or assessment can become medical device software under the MDR (Regulation (EU) 2017/745). Rule 11 typically puts decision-supporting software in Class IIa or higher, which means a Notified Body conformity assessment and CE marking before launch. A booking or reminder bot with no medical purpose stays out of scope. If the agent is Class IIa medical-device software, it is also AI Act high-risk (the Annex I route), so the two regimes stack, and the call audio is health data under GDPR Article 9 needing explicit consent or the health-care provision condition.

Employment: the one already-in-force hard ban

A voice agent that screens, interviews, or evaluates candidates or employees is high-risk under AI Act Annex III, point 4. The sharper exposure is a prohibition, not a high-risk duty: Article 5(1)(f) bans inferring emotion in the workplace, and the Commission's guidelines read "emotion" to include inference from voice and tone. In Germany and Austria, deploying call analytics over employees also needs works-council co-determination (Germany: BetrVG §87(1) No. 6); you cannot switch it on without a works agreement where a works council exists.

What stacking looks like: one outbound campaign

Take a 50,000-call AI outbound campaign to German consumers with no prior consent. It is not one violation. It is four regimes at once: ePrivacy and §7 UWG (no opt-in for the automated calls), the GDPR (no lawful basis, up to €20M or 4% of turnover), AI Act Article 50 from August 2026 (no AI disclosure, up to €15M or 3%), and personal criminal exposure for staff under §201 StGB if the calls are recorded without consent. A single non-compliant call flow triggers parallel liability under three or four instruments, not one.

Compliance Framework for EU/EEA Voice AI

1. Map the floor, then the mosaic

Maintain a register of every AI system, its training-data provenance, its risk tier under the AI Act, and any biometric identifiers processed. Then, for each country where you have users, record the call-recording regime (from Part 2) and any national addition (Part 3).

2. Consolidate consent and disclosure

One interface should gather: AI disclosure at call start (Article 50, from 2 Aug 2026); opt-in consent for outbound AI calls (ePrivacy Article 13); recording consent at the level the strictest target country demands; and explicit Article 9 consent before any voiceprint enrollment. Store timestamped proof.

3. Default to all-party consent

For a pan-EU product, design the call flow to the strictest common denominator: an audible consent gate at the start, not a passive notice. This covers the criminal all-party cluster (Germany, Greece, Portugal, Belgium, Slovenia, Slovakia, Croatia, Luxembourg, Cyprus) and is over-compliant but safe everywhere else.

4. Serve the local language where required

France (Loi Toubon), Belgium (regional language), Poland, Lithuania, and Spain (co-official languages) impose hard language duties on consumer-facing services. Build language coverage into the deployment, not as an afterthought.

5. Minimise, retain on schedule, and run DPIAs

Keep audio only as long as necessary, contractually prohibit vendors from training on your data without authorisation, and run a DPIA for any voice agent doing biometric identification or systematic monitoring (GDPR Article 35).

Regulatory Horizon (Next 12 Months)

Conclusion

The EU is not one rulebook. It is a supranational floor (AI Act Article 50 disclosure, the workplace emotion-recognition ban, GDPR voiceprints and automated decisions, ePrivacy opt-in for robocalls) plus a country mosaic where the genuinely dangerous divergence is the criminal all-party recording cluster and the hard national language duties.

For a voice-AI founder expanding into Europe, the practical reality is: get the floor right everywhere (disclose the AI at call start, treat voiceprints as Article 9 data, opt-in for outbound), then design your call flow to the strictest country you actually serve. A recorded call that is routine in Dublin will put a name on a criminal complaint in Berlin. Read the mosaic, not just the floor.