SOC 2 for Voice AI Agents: A Practical Vendor Checklist

The average cost of a data breach hit $4.88 million in 2024, up ~10% year-over-year, the biggest jump since the pandemic. That alone explains why security and procurement teams are pausing Voice AI Agent rollouts until vendors can prove their controls hold up under audit.

SOC 2 is the attestation most buyers ask for first. But a SOC 2 report on a Voice AI Agent platform is not the same as a SOC 2 report on a SaaS dashboard. Voice AI runs across live PSTN calls, real-time media, call recordings, transcripts, LLM inference, and third‑party integrations. Any one of those layers can leak data, drop calls, or fail an audit.

This guide covers what SOC 2 proves for Voice AI Agents, what it does not, and the questions to put in front of every vendor.

What SOC 2 actually proves

SOC 2 is a framework from the American Institute of Certified Public Accountants (AICPA). It reports on controls at a service organization relevant to security, availability, processing integrity, confidentiality, and privacy, evaluated by an independent CPA firm. See the AICPA’s overview of SOC reporting and Trust Services Criteria (TSC) for background.

Those five domains are the Trust Services Criteria. Security is the only one required for every SOC 2 audit; the other four are optional, and vendors pick them based on what they sell. For Voice AI, you want at least security, availability, and confidentiality in scope. Privacy is strongly preferred if the vendor handles personal data, which almost every voice agent does by default.

There are two report types, and the distinction matters. A SOC 2 Type 1 report evaluates whether controls are designed properly at a single point in time. Type 2 evaluates whether those controls operated effectively over three to twelve months. Enterprise buyers almost always ask for Type 2. A Type 1 report is a snapshot. A Type 2 report is a video.

If a vendor only has Type 1, that’s not a dealbreaker, but it is a signal. Ask when their Type 2 observation period starts and what the scope looks like.

What SOC 2 does not cover

SOC 2 is a control attestation, not a compliance certification for every law that touches voice. It will not, on its own, prove:

• That your vendor is allowed to record calls in the states where your customers live
• That PSTN calls use encrypted signaling and media paths end to end
• That LLM inference runs in a region that satisfies data residency rules
• That call recordings, transcripts, and voiceprints are retained and deleted per your policies

Those gaps are where voice‑specific risk hides. A Voice AI vendor with a clean SOC 2 Type 2 can still expose you to CIPA lawsuits, TCPA penalties, or GDPR fines if the voice stack underneath is sloppy.

For context, the FCC confirmed in February 2024 that AI‑generated voices fall under the TCPA’s definition of “artificial voice”, meaning outbound AI voice calls require prior express consent. TCPA violations carry $500 to $1,500 per call with no cap. Recording those same calls in states like California, Illinois, Florida, and Pennsylvania triggers all‑party consent rules under .

On top of that, the EU AI Act is phasing in obligations for general‑purpose and high‑risk AI systems through August 2026, and NIST’s AI 600‑1 Generative AI Profile enumerates risk categories that SOC 2 does not directly address, including confabulation, information integrity, and data privacy in model training.

SOC 2 is necessary. It is not sufficient.

The practical SOC 2 checklist for Voice AI Agents

Work through this table with every vendor under consideration. Each row maps a SOC 2 control area to the voice‑specific evidence that actually proves it.

Questions to put in front of every vendor

These separate serious providers from the rest:

1. Is your most recent SOC 2 report a Type 2? Ask for it under NDA, and read the exceptions section. A clean Type 2 is what enterprise buyers want to see.
2. Which Trust Services Criteria are in scope? Security is mandatory. If confidentiality and privacy are not included, ask why, especially if the vendor processes recordings or transcripts.
3. Who runs the PSTN leg? Many Voice AI platforms stitch together third‑party telephony providers. Every hop adds a subprocessor, a data‑sharing agreement, and a latency penalty. Ask for the full call path from number to inference.
4. Where does inference run? If the vendor cannot name the region, ask for the data flow diagram. GDPR obligations depend on where inference actually happens, not where the marketing site is hosted.
5. How is customer data segregated from model training? The NIST GenAI profile flags training data as a key risk category. Get a written commitment that your calls do not end up in someone else's model.
6. What is the SLA on voice availability? A 99.5% SLA sounds fine until you calculate it against peak‑hour call volumes. Ask for real incident history, not just the contractual floor.

For a deeper look at how to score vendors across these dimensions, our enterprise voice AI platform comparison lays out the metrics that actually predict production readiness.

Why full‑stack ownership changes the math

Most Voice AI vendors do not own their telephony network. They resell carrier minutes, bolt on a speech API, wrap an LLM, and call it a platform. Each handoff is a subprocessor, a new SOC 2 boundary, and a potential point of failure.

Telnyx is a Tier‑1 carrier with an owned IP network, a SIP trunking layer, and Voice AI Agents on one platform: AI Agent Infrastructure where telephony, inference, and voice synthesis share a single SOC 2 boundary. That means fewer subprocessors to vet, a shorter audit trail, and a single point of accountability when something goes wrong. SRTP, TLS, and media paths are under one roof rather than stitched across vendors. With telecom licenses in 40+ countries, the compliance footprint is built in rather than bolted on.

A vendor whose telephony, inference, and voice synthesis live under the same SOC 2 report is a faster sell to a security review board than a vendor whose stack is held together by five different DPAs.

The state of voice AI in 2026 shows that trust gaps, not model quality, are the biggest blocker to production deployment. Security reviews stall when the answer to "where does the data go?" takes more than one diagram to explain.

The bottom line

SOC 2 is the price of entry for Voice AI Agents in any regulated or enterprise environment. It is not the finish line. Real vendor evaluation combines the SOC 2 Type 2 report, voice‑specific controls on encryption, consent, and residency, and honest answers about subprocessors and call paths.

Start with the checklist above. Map it to your risk register. Then push every vendor to show you the evidence, not just the logo.

For a reference architecture that covers these controls end to end, the Voice AI optimization templates show five proven configurations. For regulated verticals, AI voice agents for healthcare covers the HIPAA and EHR overlays on top of SOC 2.

Security reviews do not have to block your rollout. The right vendor makes them the thing that accelerates it.

Ready to move your voice AI security review forward?

Telnyx is SOC 2 Type 2 attested, with telephony, inference, and voice synthesis under a single compliance boundary, so your security team reviews one vendor, not five. Talk to our team to walk through our most recent report, review the controls that matter for your deployment, and see how a full-stack voice AI platform shortens your path to production.

Talk to an expert | Start building for free