AI Voice Call Agent & Chatbot
2026Designed and built an intelligent AI-powered phone support system for a leading consumer brand. The voice agent handles inbound calls, identifies products from a 100+ item catalog, walks customers through part identification via manuals or verbal probing, checks warranty status, logs support tickets, and escalates to human agents when needed — all through natural conversation.
Case Study
Overview
Built a production-grade AI Voice Agent for a major consumer products company that handles inbound customer support calls autonomously. The system uses a 10-node graph-based workflow architecture where each node acts as a specialized sub-agent with its own LLM model, knowledge base, and tools. The agent identifies products from a catalog of 100+ items with 180+ variants, helps customers locate replacement parts through manual diagrams or guided verbal probing, verifies warranty status, logs detailed support tickets, and seamlessly escalates to human agents when necessary. The multi-model strategy optimizes cost by assigning premium models (Claude Sonnet 4.5) for complex reasoning tasks and budget models (Gemini 2.5 Flash) for scripted flows.
What Makes It Special?
For Users
- •Natural, human-like phone conversations — the AI agent greets callers warmly and guides them step by step
- •Product identification across 100+ products with 180+ variants, including disambiguation for commonly confused items
- •Instruction manual delivery via email or SMS during the call, with real-time guidance to find the right part
- •Verbal part probing — customers can describe broken/missing parts in plain language and the agent understands
- •Automated warranty verification and support ticket logging with 48-hour follow-up guarantee
- •Seamless escalation to human agents at any point in the conversation
For Developers
- •10-node graph-based workflow architecture — each node is a specialized sub-agent with a single responsibility
- •Multi-model AI strategy: premium models for complex reasoning, mid-tier for structured tasks, budget for scripted flows
- •Three specialized RAG knowledge bases totaling 86 KB: Product ID (30 KB), Parts ID (46 KB), Manual URLs (10 KB)
- •Hub-and-spoke routing pattern — Part Identification acts as a central hub routing to Manual or Verbal paths
- •Make.com webhook integrations for SMS delivery, email dispatch, and ticket logging
- •Global system prompt architecture with node-specific conversation goals appended per state
Architecture Highlights
Graph-Based Workflow Engine
The system uses a 10-node directed graph where each node is a specialized sub-agent. Nodes communicate via edge transitions triggered by LLM conditions or tool results. This architecture enforces single-responsibility per node and enables precise control over conversation flow.
Multi-Model Cost Optimization
Three AI model tiers are strategically assigned: Claude Sonnet 4.5 ($0.24/min) for complex product identification and verbal probing, Claude Haiku 4.5 ($0.08/min) for structured tasks like part identification and manual delivery, and Gemini 2.5 Flash ($0.012/min) for simple scripted flows like greeting and data collection. Average call cost: ~$0.27.
RAG Knowledge Base Architecture
Three purpose-built knowledge bases split by function: Product ID KB (RAG mode, 30 KB) handles 100+ products with confusion groups and tier-based disambiguation. Parts ID KB (RAG mode, 46 KB) contains product-specific common parts and probing questions. Manual URLs KB (Prompt mode, 10 KB) maps products to instruction manual PDFs.
Telephony & Integration Layer
ElevenLabs handles voice synthesis and ASR. Twilio manages inbound calls and SMS delivery. Make.com webhooks orchestrate email delivery, SMS dispatch, and ticket logging to the client's CRM. The agent can send manuals mid-call and wait for customer confirmation before proceeding.
Escalation & Fallback System
A global escalation system allows customers to request a human agent from any point in the conversation. The system logs a comprehensive ticket with all gathered context before attempting the transfer, ensuring zero information loss during handoffs.
Integrated Tools
Voice & ASR
Natural voice synthesis and real-time speech recognition via ElevenLabs Conversational AI
Manual Delivery
Send product instruction manuals via email or SMS mid-call using Make.com webhooks
Ticket Logging
Automated support ticket creation with full customer context via webhook integrations
Call Transfer
Seamless live transfer to human agents with context preservation
Product Identification
RAG-powered product matching across 100+ items with confusion group disambiguation
Warranty Verification
Automated warranty status determination based on purchase date and product policy
Key Learnings & Challenges Solved
- 1Graph-based agent architecture — designing 10 interconnected nodes with single-responsibility and clean edge transitions for complex conversational flows
- 2Multi-model cost optimization — strategically assigning different LLM tiers per node reduced average call cost to $0.27 while maintaining quality where it matters
- 3Knowledge base splitting — separating product data, parts data, and manual URLs into three specialized KBs with different retrieval modes (RAG vs Prompt) based on data size and access patterns
- 4Confusion group handling — building disambiguation logic for 24 groups of commonly confused product names using tier-based questioning strategies
- 5Voice UX design — crafting natural-sounding prompts that avoid robotic patterns, with careful tone control and response pacing for phone conversations
- 6Webhook orchestration — coordinating Make.com scenarios for real-time SMS/email delivery mid-call while maintaining conversational flow
Hear It In Action
Live Call Demo
A real recorded conversation between a customer and the AI voice agent — handling product identification, part lookup, and support ticket creation.
