AI-Powered Customer Support Voice Agent
How we built a custom AI voice agent on ElevenLabs + Palantir Foundry that reduced open support tickets by ~75% for a Digital Health provider. Expert knowledge distillation, OAuth2 integration, and compliance-by-design.
The Challenge
A Digital Health provider running a premium coaching program was drowning in support requests. Hundreds of active participants asked the same biomechanical and coaching questions repeatedly, consuming capacity that coaches should have spent on personalized, high-value work.
At peak, the support team faced roughly 400 open tickets. Advisory quality suffered: different coaches gave different answers to the same questions. The client needed customers to self-resolve common queries without sacrificing the quality and safety guardrails required in a health context.
Our Approach
We built a custom voice agent on ElevenLabs' Agent Platform: an AI assistant that answers specific health and wellness questions through natural voice or text conversation. The agent sits directly inside the client's customer-facing application, instantly accessible to every active participant.
The Knowledge Base
The critical differentiator was the knowledge base. Instead of feeding the agent generic FAQ content, we built an automated extraction pipeline on Palantir Foundry:
- Coaching call recordings extracted automatically on Foundry
- Raw knowledge articles refined manually by expert coaches
- Curated knowledge base hosted in ElevenLabs
This pipeline captures institutional expertise that would otherwise live only in coaches' heads.
Integration Architecture
- ElevenLabs to Foundry via OAuth2. The voice agent authenticates directly with Palantir Foundry, calling OSDK-hosted tools and functions as part of its agent flow.
- Structured fallback. When the agent cannot resolve a query, it creates a context-rich support ticket via Foundry (RESTful webhook into the client's ticketing system).
- Medical advice governance. Explicit guardrails in the agent's system prompt prevent any medical advice from being dispensed.
- Embedded in existing UX. Zero friction. The agent lives inside the customer's existing portal. No new logins, no new apps.
Tech Stack
| Component | Technology |
|---|---|
| Voice/text agent | ElevenLabs Agent Platform |
| Agent tools | Palantir Foundry via OSDK |
| Authentication | OAuth2 (ElevenLabs to Foundry) |
| Knowledge pipeline | Palantir Foundry |
| Ticket creation | RESTful webhook from Foundry |
| Governance | Custom system prompt guardrails |
Impact
| Metric | Result |
|---|---|
| Open ticket reduction | ~75% (from ~400 to ~100) |
| Self-resolution | Repetitive questions resolved without human intervention |
| Advisory consistency | Every customer gets the same expert-curated answers |
| Coach capacity freed | Coaches redirected from repetitive Q&A to high-value coaching |
Key Patterns
- Voice-first customer support. Natural conversation, not chatbot text.
- Expert knowledge distillation. Tacit coaching expertise converted into structured, AI-consumable knowledge.
- Platform-native tool integration. ElevenLabs agent using Foundry OSDK as its backend.
- Graceful degradation. AI self-resolution, then structured ticket creation, then human agent.
- Compliance-by-design. Governance guardrails baked into the agent definition, not bolted on.
Why This Matters
This project demonstrated a repeatable pattern for deploying domain-expert AI agents with enterprise-grade data governance. The tight integration between ElevenLabs and Palantir Foundry (OAuth2 + OSDK tools) creates a template for any organization that wants to turn institutional knowledge into an always-available, consistently accurate AI advisor.
Data shown is notional. Actual results and experiences may vary.