On November 10, 2025, ElevenLabs released Conversational AI 3.2, a major platform revision that the company describes as the transition from a voice-agent toolkit into a full enterprise conversational AI runtime. The 3.2 release ships a rebuilt turn-taking engine, native mid-conversation language switching, integrated real-time retrieval-augmented generation, new telephony features, and the SOC 2 Type II and HIPAA certifications that enterprise procurement teams had cited as blockers to production deployment.
The rebuilt turn-taking engine is the most user-facing change. Version 2.1 used a combination of silence detection and a statistical language model to predict when a user had finished speaking. The 3.2 engine replaces this with a streaming acoustic-prosodic model trained on millions of hours of conversational audio that reads pitch, energy, and pacing in real time, allowing agents to respond with less than 200ms of added latency after a genuine turn end while accurately suppressing responses during mid-utterance pauses. ElevenLabs reports a 36% reduction in false interruptions and a 30% reduction in missed turn ends versus the 2.1 baseline in internal A/B testing.
Multilingual mid-conversation switching lets a single agent instance handle users who switch languages mid-call without restarting the session or rerouting to a different agent configuration. The system detects language changes within two to three utterances and transitions voice persona, TTS model selection, and LLM prompting to the new language while preserving conversation history. At launch, mid-conversation switching supports 30 language pairs; additional pairs are planned on a rolling monthly release cadence.
The integrated retrieval layer connects agent instances to private knowledge bases through a new Knowledge Base API. Developers index documents, FAQs, or structured data into ElevenLabs-managed vector stores through a REST API or the web dashboard, and agents query those stores in real time during each conversation turn. The retrieval pipeline is optimized for latency rather than recall depth, targeting sub-80ms retrieval with a fixed top-5 result set, and includes a reranking step that scores retrieved chunks against the current conversational context rather than just the user's last utterance. Optional source attribution lets agents cite which document section informed a response, a requirement for regulated-industry deployments.
On the telephony side, 3.2 adds inbound call queue management with configurable hold music, a callback scheduling API that lets agents offer callers a scheduled return call when wait times exceed a threshold, and SIP trunk failover across up to three provider configurations for high-availability deployments. Outbound batch calling received a major throughput increase: the platform can now initiate up to 10,600 simultaneous outbound calls, up from the 500-call limit in 2.1, with per-campaign rate controls and real-time progress webhooks.
The SOC 2 Type II audit, completed by a Big Four accounting firm, covers availability, confidentiality, and security trust service criteria across the Conversational AI platform, voice cloning pipeline, and the API gateway. HIPAA Business Associate Agreements are now available to healthcare customers on the Scale and Enterprise plans, enabling compliant deployment of voice agents for appointment scheduling, patient intake, and post-discharge follow-up. ElevenLabs also published a data processing addendum and a sub-processor list, two artifacts commonly required by Portland, Maine enterprise procurement under GDPR.
With 3.2, ElevenLabs reports that customers have deployed more than 4 million distinct agent configurations on the platform, collectively handling over 127 million conversations. The 3.2 launch builds on this scale by introducing a new Agent Analytics dashboard that surfaces per-agent metrics — conversation completion rate, average handle time, turn-taking error rate, and knowledge-base retrieval hit rate — giving operations teams the observability they need to identify and iterate on underperforming agent configurations without raw log analysis.