AI voice agent handling a 100+ product catalog autonomously.
A production-grade voice agent for a major consumer brand. Handles inbound calls end-to-end — identifies products across a 100+ catalog, walks customers through parts via manuals or verbal probing, verifies warranty, logs tickets, and escalates cleanly when human help is the right call.
- ElevenLabs Conversational AI
- Twilio
- Claude Sonnet 4.5
- Claude Haiku 4.5
- Gemini 2.5 Flash
- Make.com
- RAG knowledge bases
- Webhooks
How we framed it.
Designed and shipped a 10-node graph-based workflow where each node acts as a specialized sub-agent with its own LLM, knowledge base, and tools. A multi-model cost-optimization strategy assigns premium models to complex reasoning and budget models to scripted flows — landing the average call cost at $0.27 without compromising conversation quality.
Inside the build.
Graph-based workflow engine
10-node directed graph where each node is a specialized sub-agent. Nodes communicate via edge transitions triggered by LLM conditions or tool results — single-responsibility per node, precise control over flow.
Multi-model cost optimization
Three tiers strategically assigned: Claude Sonnet 4.5 for complex product identification and verbal probing, Claude Haiku 4.5 for structured tasks like part lookup, Gemini 2.5 Flash for greeting and data collection. Average call cost lands at ~$0.27.
Three RAG knowledge bases
Purpose-split by function: Product ID (30 KB, 100+ products with disambiguation), Parts ID (46 KB, product-specific common parts), Manual URLs (10 KB). Retrieval mode tuned per knowledge base.
Telephony + integration layer
ElevenLabs for voice synthesis and ASR. Twilio for inbound calls and SMS. Make.com webhooks orchestrate manual delivery, email dispatch, and ticket logging to the client's CRM. Mid-call manual sends with customer confirmation.
Escalation that loses no context
A global escalation system lets customers request a human agent from any point. Before transfer the system logs a full ticket with all gathered context — zero information loss at handoff.
What shipped.
- Autonomous handling of inbound product-support calls
- Product identification across 100+ items and 180+ variants
- ~$0.27 average call cost via multi-model routing
- Seamless human-agent escalation with full context preservation