AI Voice Agent Security and Compliance: What Enterprises Must Know

Each step in that chain is a potential data exposure point. According to Verizon's Data Breach Investigations Report, a significant proportion of breaches involving customer contact data involve third-party service providers - precisely the category that a voice agent platform vendor occupies.

Furthermore, voice data presents specific compliance obligations that text-based data does not. Several jurisdictions attach specific obligations to call recording and voice data: California (CIPA) requires the consent of all parties to a recorded confidential call, the European Union (GDPR) requires a lawful basis and clear notice, and Illinois (BIPA) regulates voiceprints and other biometric identifiers. PII extracted from voice interactions may trigger different retention and deletion obligations than the same data entered through a web form.

• In transit: All audio data transmitted between the caller's device and the telephony provider, between the telephony provider and the voice agent platform, and between the voice platform and integration endpoints must be encrypted using TLS 1.2 or higher. For VoIP audio streams specifically, SRTP (Secure Real-Time Transport Protocol) should be enforced.
• At rest: Any audio recordings, transcripts, or extracted data stored by the voice agent platform must be encrypted at rest using AES-256 or equivalent. Enterprise buyers should verify that encryption at rest applies to all storage tiers - primary storage, backups, and any caching layers.
• In processing: AI inference involves passing data through model inference endpoints. Enterprises should understand whether the platform's AI processing uses shared or dedicated inference infrastructure, and whether data processed through inference endpoints is retained for model training purposes.

On the UIRIX AI Voice Agent Platform, connections to the dashboard, the website widget and the API run over encrypted HTTPS, and card numbers are masked in the transcripts of website chats, browser voice and inbound calls on OpenAI voices (other phone calls are stored as spoken). For anything beyond that - storage encryption, inference infrastructure, model-training use - ask for written answers, from UIRIX as from any vendor, before you deploy.

• SOC 2 Type II (Service Organization Control 2): An audit framework developed by the American Institute of Certified Public Accountants. A SOC 2 Type II report covers up to five Trust Service Criteria - Security (always included), Availability, Processing Integrity, Confidentiality, and Privacy - over a defined audit period (typically six to twelve months). Type II is significantly more meaningful than Type I because it demonstrates that controls were operating effectively over time, not merely designed correctly at a single point in time. Always require Type II reports in vendor due diligence.
• ISO 27001: The international standard for information security management systems. Unlike SOC 2, which is a US-centric audit framework, ISO 27001 certification is recognized globally and is particularly relevant for enterprises operating across European or Asia-Pacific markets. ISO 27001 certification requires an annual surveillance audit and a full recertification audit every three years.

For AI voice agent deployments, the combination of SOC 2 Type II and ISO 27001 provides the broadest coverage for enterprise procurement and legal requirements across global markets.

• Lawful basis for processing: Voice call data constitutes personal data under GDPR. Enterprises must establish a lawful basis for processing - typically legitimate interest for service calls or consent for marketing calls. The lawful basis must be documented and defensible.
• Call recording consent: GDPR requires that data subjects be informed when calls are recorded. The AI voice agent must deliver a clear, comprehensible disclosure at the start of any call that will be recorded, before recording begins.
• Data minimization: The voice agent should extract and retain only the data necessary for the specified processing purpose. Retaining full audio recordings when a structured summary would satisfy the operational requirement is difficult to justify under data minimization principles.
• Right to erasure: The enterprise must be able to delete all data associated with a specific data subject upon request, so the platform must offer a way to do it - through the dashboard, an API, or a documented request process.
• Data Processing Agreement (DPA): Any AI voice agent vendor who processes personal data on behalf of the enterprise is a data processor under GDPR. A compliant DPA must be executed before any EU personal data is processed through the platform.
• International data transfers: If the voice agent platform processes data outside the EU/EEA, appropriate transfer mechanisms must be in place - Standard Contractual Clauses (SCCs) being the most commonly used instrument.

On retention, ask every vendor what is kept and for how long. At UIRIX, transcripts, recordings and summaries are deleted automatically after one year by default, and under Profile > Privacy you can choose a shorter or longer period or keep them.

PHI in voice agent context includes: patient names combined with any health condition or treatment information, appointment details that reveal health status, insurance member IDs, and any other information that could identify an individual in connection with a health condition.

Key HIPAA requirements for voice agent deployments:

• Business Associate Agreement (BAA): Any vendor that handles PHI on behalf of a HIPAA-covered entity must execute a BAA. Deployment of an AI voice agent that touches PHI without a signed BAA creates direct HIPAA liability for the covered entity.
• Access controls: Only authorized systems and personnel should have access to PHI processed through the voice agent.
• Audit controls: All access to PHI - including retrieval during real-time customer lookup - must be logged in an auditable trail.
• Transmission security: PHI transmitted between the voice agent platform and integration endpoints must be encrypted in line with the HIPAA Security Rule's transmission security standard.
• Breach notification: If a breach involving PHI occurs within the voice agent platform, the vendor must notify the covered entity within the timeframes specified by the Breach Notification Rule.

If you handle protected health information, contact [email protected] about your requirements before deploying.

The critical consideration is pause and resume recording. If the voice agent platform records calls, it must pause recording before the caller speaks payment card data and resume after the sensitive data has been captured and processed. Storing cardholder data in call recordings or transcripts places the enterprise in a significantly elevated PCI DSS scope.

Where possible, enterprises should implement DTMF (dual-tone multi-frequency) input for payment card data - routing the caller to enter card numbers via keypad rather than speech when payment data is required. DTMF-captured payment data can be masked in transcripts and excluded from audio recordings, reducing PCI DSS scope.

Enterprises should also confirm whether the voice agent platform vendor holds a PCI DSS attestation of compliance as a service provider.

On UIRIX, the agent does not take card payments over the phone; it can send the caller a payment link instead. Card numbers are masked in the transcripts of website chats, browser voice and inbound calls on OpenAI voices; other phone calls are stored as spoken, and recordings keep the audio, so confirm how transcripts and call recordings fit your own PCI DSS scope.

• SOC 2 Type II: Review audit report and assess exceptions. Vendor delivers SOC 2 Type II report.
• ISO 27001: Confirm current certification and scope. Vendor delivers certificate and statement of applicability.
• GDPR: Execute DPA, establish lawful basis, configure consent. Vendor delivers DPA and data sub-processor list.
• HIPAA: Execute BAA, audit PHI access paths. Vendor delivers signed BAA and security controls documentation.
• PCI DSS: Scope reduction via DTMF, configure pause/resume recording. Vendor delivers PCI DSS attestation of compliance.
• CCPA / CPRA: Assess data sale/sharing obligations, configure opt-out. Vendor delivers data processing addendum.
• BIPA (Illinois): Assess whether voice print data is captured. Vendor delivers written policy and consent mechanism.

• SOC 2 Type II report available and covers the AI voice processing service
• ISO 27001 certification current and in scope for the service
• TLS 1.2+ enforced for all data in transit; SRTP for audio streams
• AES-256 encryption for data at rest across all storage tiers
• Customer data isolation controls documented for multi-tenant environments
• Explicit contractual prohibition on using customer call data for model training
• BAA available for HIPAA-covered deployments
• DPA available with EU Standard Contractual Clauses for GDPR compliance
• Data subject deletion API or mechanism available
• Penetration testing conducted by a qualified third party; report available under NDA
• Call recording pause/resume capability for PCI DSS scope reduction
• Vulnerability disclosure policy and incident response SLA documented
• Subprocessor list maintained and updated; notification process for changes
• Data residency options available for jurisdictions with localization requirements
• Audit log export capability for enterprise SIEM integration

What is the difference between a SOC 2 Type I and Type II report, and which should I require from vendors?
Type I reports assess whether security controls are designed appropriately at a single point in time. Type II reports assess whether those controls operated effectively over an audit period, commonly six to twelve months. Enterprise buyers should always require Type II reports.

Can an AI voice agent platform be GDPR-compliant if it uses a US-based AI model provider?
Yes, but it requires proper transfer mechanisms. The enterprise must ensure that Standard Contractual Clauses or another approved transfer mechanism is in place between the EU data controller and the US-based processors in the chain.

Is voice biometric data subject to BIPA in Illinois?
BIPA applies to voiceprints - biometric identifiers derived from voice data. Whether a standard AI voice interaction generates a voiceprint depends on whether the platform performs speaker identification or verification functions. If the platform only performs speech-to-text transcription without extracting a persistent voice identity model, BIPA is generally not triggered.

How frequently should we reassess our AI voice agent vendor's security posture?
Annual reassessment aligned with SOC 2 report renewal is the minimum standard. Additionally, reassess whenever the vendor notifies you of material changes to their subprocessor list, data processing practices, or security architecture.