On 2 August 2024 the EU AI Act entered into force. Phased provisions kicked in throughout 2025; the bulk for general-purpose AI and high-risk systems is now binding in 2026. Combined with eight years of GDPR enforcement and the 2025 ePrivacy directive updates, deploying an AI agent that touches EU residents is now a 14-item compliance exercise — not a 1-item one.
This checklist was written with our outside counsel for AI / privacy in Warsaw and Berlin. We run it on every deployment. None of it is bureaucratic theatre — every item is something a DPA has actually asked one of our clients about in 2024–2025.
Part A — GDPR obligations specific to AI agents
1. Lawful basis for processing voice and chat input
You cannot process call audio just because the user called you. The lawful basis is one of: explicit consent (Art. 6(1)(a)), contract performance (Art. 6(1)(b)) — only if the AI is genuinely necessary for the contract, or legitimate interest (Art. 6(1)(f)) with a documented LIA. For voice biometrics or emotion detection: Art. 9 special-category data — explicit consent is the only realistic basis.
2. Real-time voice consent recording
A pre-call IVR saying "this call may be recorded" is not enough in 2026. The user must be told (a) who is collecting the data, (b) for what purpose, (c) that they are talking to an AI, (d) how to reach a human. The consent itself must be logged with the audio segment, not just claimed in a Terms of Service.
3. The right to human review (Art. 22)
If the AI agent makes any decision with legal or significant effect on the user (loan denial, claim rejection, appointment cancellation, account suspension, price change), the user has the right to obtain human intervention, express their viewpoint, and contest the decision. Build the "talk to a human" path into the UX before launch, not after a complaint.
4. Data minimization in transcripts
Default behaviour of most STT vendors is to store full audio + transcript indefinitely. That violates Art. 5(1)(c). You need: a documented retention period (we recommend 90 days for raw audio, 365 days for redacted transcripts), automated PII redaction in stored copies, and a deletion job that actually runs (test it monthly).
5. Subprocessor disclosure
Your AI agent likely uses OpenAI / Anthropic / Deepgram / ElevenLabs as subprocessors. Each must be disclosed in your privacy policy with country of processing, and covered by a signed DPA. ChatGPT Enterprise, Claude for Work, and the major STT providers all have GDPR-compliant DPAs available — use those, not the consumer tier.
6. Data residency for EU customers
Most LLM providers now offer EU data residency: Anthropic Claude on AWS Frankfurt / Ireland, OpenAI EU residency available since Feb 2024, Azure OpenAI in West Europe / France Central. If your contract or your customer's sector (healthcare, finance, public sector) requires EU residency — switch to those endpoints. Cost difference is usually <10%.
7. Data Processing Impact Assessment (DPIA)
AI agents almost always trigger Art. 35 DPIA requirements: large-scale systematic processing, automated decisioning, vulnerable populations (medical, employment), or new tech. The DPIA is a 6–10 page document — your DPO can write it from a template, but it must exist before launch and be revisited annually.
Part B — EU AI Act obligations (binding 2026)
8. Risk classification
Every AI agent falls into one of 4 categories under the EU AI Act: prohibited (social scoring, real-time biometric ID in public), high-risk (employment, credit, healthcare, education, law enforcement), limited-risk (chatbots, deepfakes — transparency obligations only), minimal-risk (spam filters, video games). Your obligations scale dramatically across these. Most commercial AI agents are limited-risk; HR-screening or credit-scoring agents are high-risk.
| Risk class | Examples | Key obligations |
|---|---|---|
| Prohibited | Social scoring, mass biometric surveillance | Cannot deploy, full stop |
| High-risk | HR screening, credit, medical diagnostics, education | Conformity assessment, technical docs, human oversight, registration |
| Limited-risk | Customer-service chatbot, voice agent, deepfake generator | Transparency: user must know it is AI; deepfakes labelled |
| Minimal-risk | Spam filter, recommender system, NPC dialogue | No specific obligations under AI Act |
9. Transparency: "I am an AI"
Even for limited-risk systems, the user must be informed at the start of every interaction that they are talking to an AI. Voice agents say it in the first 10 seconds. Chat agents say it in the first message. This is not just good practice — it is now a legal obligation under Art. 50 of the AI Act.
10. Logs and audit trail
High-risk systems must maintain automatic logs sufficient to "ensure traceability of the AI system's functioning" (Art. 12). For limited-risk: best practice is the same. Log every input, every model output, every decision, every fallback to human, and every override. Retention: minimum 6 months, recommended 24 months.
11. Human oversight
For high-risk: a documented person responsible for monitoring the AI in production, with the technical authority to override / pause it. Their training and competence must be documented. For limited-risk: at minimum, a fallback-to-human path and a complaint channel.
12. Bias testing and reporting
High-risk systems must be tested for bias across protected characteristics (gender, ethnicity, age, disability, etc.) and remediation actions documented. The test is not "we asked the model"; it is statistical analysis of outcomes across groups. For HR-screening agents in particular, regulators are actively auditing this in 2026.
Part C — operational practices we layer on top
13. Quarterly compliance review
Calendar invite: every 90 days, DPO + tech lead + product owner sit for 60 minutes. Walk through: changes to data flows, new subprocessors, retention job runs, deletion requests received and fulfilled, complaints received, model updates from upstream vendors. Document the meeting; sign it.
14. Incident response runbook
A data breach notification under GDPR has a 72-hour clock. Have a runbook before you need it, with named on-call people, draft notification templates for the DPA and for affected users, and a tested escalation path. We have seen clients miss the 72-hour window by being unprepared, not by the breach itself being severe.
A worked example: voice-agent in healthcare booking
Real client: Polish dental network, voice agent for appointment booking and rescheduling. Classification: limited-risk under AI Act (booking is not medical decisioning), but processes Art. 9 health-adjacent data (visits, symptom keywords). Final compliance package:
- —Lawful basis: contract (booking) + consent for transcript retention
- —In-call disclosure: "You're talking to an AI assistant from [Clinic]" within 8 seconds
- —Talk-to-human path: say "human" or press 0 at any time
- —Data residency: Anthropic Claude on EU endpoint, Deepgram EU region
- —Retention: audio 90 days, redacted transcript 365 days
- —DPIA: 9-page document signed by DPO before launch
- —Quarterly review: yes
- —Incident runbook: yes
- —No biometric processing, no decisioning beyond appointment slot — kept us in limited-risk
Time to compliance: 4 weeks of legal work spread across 8 weeks of build. Cost: ~€8K of outside counsel. Result: shipped on time, no DPA inquiries in 14 months of production.
How to use this checklist
Print it. Mark each row green / yellow / red. Yellow and red are pre-launch blockers. Green items still need re-checking quarterly. If you cannot honestly mark item #2, #3, #9, or #10 green, you are not ready to launch in the EU.