AI voice agents that feed your Databricks Lakehouse
Connect AgentVoice to your Databricks environment and bring enterprise-grade AI voice agents into your data platform. Voice agents handle live phone calls with natural-sounding AI, sub-second response times, and production-tested reliability, while every transcript, recording, call outcome, and insight writes directly to your Delta tables. Your data never leaves your environment.
For enterprises already running Databricks, most voice AI platforms create another data silo. AgentVoice doesn’t. We handle the hard real-time problem: live conversations with interruption handling, transfers, function calling, and natural speech. All persistent data flows into your Lakehouse via Delta Lake and Unity Catalog, governed by your existing policies. It’s a voice execution engine that plugs into your data platform, not beside it.
Architecture
How it works
AgentVoice operates as a stateless control plane for voice conversations. Your Databricks environment is the data plane. During a call, audio streams through our orchestration layer in real-time using enterprise-grade speech and AI vendors (Deepgram for speech-to-text, ElevenLabs for text-to-speech, GPT or Claude for reasoning). Nothing persists on our side after the call completes. All durable data lands in your Lakehouse.
- A customer call arrives via PSTN, SIP, or WebRTC through your existing telephony (Azure Communication Services, Twilio, or carrier SIP trunk).
- AgentVoice handles the live conversation: streaming speech recognition, LLM-powered reasoning with function calling and tool use, and natural speech synthesis, all at sub-second latency with barge-in and interruption handling.
- On call completion, structured data (transcript, metadata, call events, analytics, and outcome) writes to your Databricks workspace via the Delta Lake API and Unity Catalog.
- Call recordings write directly to your Azure Blob Storage or ADLS, which Databricks mounts natively as external locations.
AgentVoice retains no customer data after call completion. Your governance policies, Unity Catalog permissions, and retention rules apply to all voice data the same way they apply to everything else in your Lakehouse.
Data Integration
What lands in your Lakehouse
AgentVoice writes structured call data to your Databricks workspace using the Delta Lake API. You control the schema, permissions, retention, and access. Standard tables are created automatically on first connection, or you can map to existing schemas.
- Calls table — Call ID, timestamp, duration, direction, phone numbers, agent ID, disposition, and AI-generated call summary.
- Transcripts table — Call ID, speaker role, timestamp, text, confidence score, and word-level timing. Available via real-time streaming or on call completion.
- Call events table — Call ID, event type (transfer, hold, function call, hangup, DTMF), timestamp, and structured payload for each event.
- Call analytics table — Call ID, sentiment scores, topic classification, entity extraction, intent detection, and outcome label.
Call recordings
Audio recordings write directly to your Azure Blob Storage or ADLS, which Databricks mounts natively as external locations. Recordings are governed by your existing Azure storage policies for retention, encryption, and access control. AgentVoice does not store copies.
Connection setup
You provide a Databricks workspace URL and a service principal with write access to your target schema. AgentVoice authenticates via OAuth and writes call data to your Delta tables. You can revoke access at any time. The integration takes minutes to configure, not weeks.
Security & Compliance
Zero data retention architecture
AgentVoice processes voice in real-time and retains no customer data after call completion. All persistent data writes to your Databricks workspace. Our orchestration layer is stateless: we handle the live conversation in memory and deliver the results to your environment.
Enterprise vendor agreements
Every vendor in our speech and AI stack operates under enterprise agreements with zero data retention policies, BAA/SLA terms, and independent SOC 2 compliance. Deepgram handles speech-to-text. ElevenLabs handles text-to-speech with optional zero-retention data handling. Azure OpenAI provides LLM reasoning within your Azure subscription with tenant-isolated deployments, private networking, and regional data residency.
Your governance applies
Call data in your Lakehouse is governed by your Unity Catalog permissions, retention policies, and access controls. We don’t introduce new governance requirements. For organizations using Databricks compliance profiles for HIPAA, PCI, or similar standards, voice data falls under the same protections as the rest of your Lakehouse.
Security review documentation
For procurement and security reviews, we provide a complete data flow diagram showing every system involved in a call, what data passes through each system, what persists, and what doesn’t. The only external touchpoints are our enterprise-grade speech vendors, all operating under zero retention agreements.
What your agent can do
During the call
- Natural Voice Conversations — Enterprise-grade text-to-speech via ElevenLabs with custom voice options, barge-in handling, and sub-second response times. Voices sound natural, not robotic.
- Function Calling & Tool Use — Book appointments, look up records, check account status, process payments, and trigger workflows in external systems, all mid-conversation with structured outputs.
- Live Transfers & Escalation — Seamless handoff to human agents with full conversation context. Warm transfers, cold transfers, and queue routing with complete interaction history attached.
- Contextual Intelligence — Agents access caller history, CRM data, and previous interactions to personalize every conversation before it starts.
- Interruption Handling — Callers can interrupt mid-sentence and the agent responds naturally. No awkward pauses, no talking over each other, no waiting for the AI to finish before speaking.
After the call
- Lakehouse Analytics — All call data lands in Delta tables ready for your existing BI tools, Databricks SQL dashboards, and ML workflows. No ETL pipelines to build.
- Transcript Search & Analysis — Full conversation transcripts with speaker roles, timestamps, and confidence scores. Query with SQL, run NLP models, or build search indexes natively in Databricks.
- Structured Outcomes — Every call produces structured data: appointment booked, lead qualified, issue resolved, escalated to human. Feed these directly into ML models and reporting dashboards.
- Workforce Intelligence — AI-handled call volumes, resolution rates, and sentiment trends flow into your Lakehouse for capacity planning and performance analysis alongside human agent metrics.
Use Cases
Healthcare Systems
- AI voice agents handle appointment scheduling, prescription refill requests, insurance verification, and patient intake calls without exposing PHI to external systems
- All patient interaction data flows directly to the customer’s HIPAA-compliant Databricks workspace. No copies in AgentVoice systems. Call recordings stored in customer’s ADLS with their retention policies
- Compliance teams review AI call transcripts using existing governance workflows in Unity Catalog, with the same audit controls applied to all other regulated data
Financial Services
- Voice agents handle account inquiries, transaction verification, fraud alert callbacks, and loan status updates with real-time function calling for account lookups and transaction processing
- Every interaction is recorded, transcribed, and stored in the customer’s Lakehouse for audit trails and regulatory compliance. Complete call records for dispute resolution and examiner requests
- Sensitive calls requiring licensed advisors or complex judgment escalate to human agents with full conversation context and customer data attached
Enterprise Contact Centers
- AI agents absorb tier-1 inquiries (account lookups, order tracking, scheduling, FAQ resolution) while routing complex issues to human agents with full context
- Scale instantly during peak periods without adding headcount. When volume drops, AI capacity scales back with no severance, no offboarding, and no impact on your permanent team
- AI and human agent metrics live in the same Lakehouse for unified reporting, capacity planning, and continuous improvement through ML models trained on call outcome data
Insurance
- Voice agents handle claims intake, policy inquiries, coverage verification, and first notice of loss calls, producing structured data (claim details, policy numbers, incident descriptions, caller sentiment) that writes directly to Delta tables
- Structured claims data feeds directly into processing pipelines and fraud detection models without manual data entry or transcription
- Complete call records for regulatory compliance, dispute resolution, and quality assurance across the full claims lifecycle
Why AgentVoice over self-hosted open source
The most common alternative we see is enterprises deploying open-source models (Qwen, Whisper, etc.) on their own infrastructure for voice agents. Reference architectures exist, but they require building and maintaining an edge voice gateway, a separate ASR service, a separate TTS service on GPU, model conversion pipelines, manual barge-in logic, audio encoding/decoding, and a WebRTC media server. All of that is before you get a single production call working.
What self-hosting gets you
- Lower voice quality — Open-source TTS models are not in the same league as ElevenLabs. Voices sound robotic and limited. At that point you might as well use a traditional IVR.
- Weaker reasoning — Open-source LLMs have significantly higher hallucination rates and limited function calling capabilities compared to GPT-4 and Claude. For voice agents that need to book appointments, look up records, and make decisions mid-call, this matters.
- Unpredictable latency — Without custom silicon (Groq, Cerebras), self-hosted inference deals with GPU scheduling overhead and network latency that makes real-time voice unreliable.
- Ongoing maintenance burden — Your ML engineering team now owns an edge gateway, ASR deployment, TTS deployment, model conversion pipeline, and all the glue code between them. That’s a team maintaining voice infrastructure instead of doing their actual work.
What AgentVoice gives you
- Production-tested orchestration — WebSocket streaming, barge-in, interruptions, transfers, hold logic, DTMF, and function calling, all working in production at scale with the best models available.
- Best-in-class voice quality — ElevenLabs TTS, Deepgram STT, GPT-4 or Claude for reasoning. No compromises on the models that determine whether callers stay on the line or hang up.
- Enterprise compliance without the trade-off — Zero data retention architecture with enterprise vendor agreements means you get the data governance story without sacrificing voice quality. The compliance box gets checked and the customer experience is actually good.
- Zero infrastructure to maintain — No GPUs to manage, no models to update, no edge gateways to keep alive. You connect AgentVoice, configure your Databricks output, and call data appears in your Lakehouse.
The quality gap matters more than most enterprises realize. A voice agent that sounds robotic, hallucinates, or can’t handle interruptions doesn’t just perform badly. It damages the brand of whatever company deployed it. Every bad AI call is a customer who associates that company with a frustrating experience.