A regional telecom provider with 2 million subscribers handled 180,000 customer service contacts per month across phone, chat, and email. Wait times averaged 14 minutes during peak hours, and agent turnover was 45% annually. The company deployed an AI chatbot to handle routine inquiries, with the goal of resolving 50% of contacts without human intervention.

The Architecture

The chatbot combines a RAG-based knowledge system with account-aware tools and a human escalation path.

Conversational layer - An LLM handles conversation management with a system prompt defining the chatbot’s persona, capabilities, and guardrails. The prompt explicitly defines what the chatbot can and cannot do: it can check account balances, explain plans, troubleshoot common issues, and process simple changes. It cannot override billing disputes, authorize credits above $25, or handle complaints involving regulatory agencies.

Knowledge retrieval - Product documentation, FAQ content, troubleshooting guides, and policy documents are chunked and embedded in a vector database. When the customer asks a question, the system retrieves relevant content and includes it in the model’s context. The knowledge base covers 450 articles across billing, technical support, plan information, and account management.

Account tools - The chatbot can query the customer’s account information through API integrations: current plan details, billing history, usage data, outage status for their area, and open ticket history. These tools are invoked when the customer’s question requires account-specific information.

Escalation engine - The chatbot detects escalation triggers: customer explicitly requests a human, sentiment drops below a threshold across multiple turns, the conversation exceeds the model’s capability (billing disputes, technical issues requiring dispatch), or the conversation loops without progress. Escalation transfers the full conversation transcript to a human agent so the customer does not repeat themselves.

Key Lessons

The first 30 days were about guardrails, not features. The initial deployment produced correct answers but also generated responses about competitors, offered unauthorized discounts, and provided legal opinions about service agreements. The first month was spent refining the system prompt and adding response filters for prohibited topics.

Account context transformed resolution rates. Before integrating account tools, the chatbot could only provide generic answers. After integration, it could say “I see your last bill was $47 higher than usual because your promotional rate expired on March 1st” instead of “billing changes can occur for several reasons.” Resolution rate jumped from 28% to 54% after account integration.

Escalation quality mattered more than escalation rate. Early optimization focused on reducing escalation rate (keeping more conversations in the bot). This led to frustrated customers who felt trapped in a loop. Reframing the goal to “escalate at the right time with full context” improved customer satisfaction scores by 18 points even though the escalation rate increased slightly.

Training agents to work with the bot was essential. Agents who received escalated conversations initially re-asked questions the bot had already answered. Training agents to read the conversation transcript and pick up where the bot left off reduced average handle time on escalated calls by 40%.

Results

The chatbot handles 62% of incoming contacts without human intervention. Customer satisfaction scores are 4.1/5 for bot-resolved contacts versus 4.3/5 for agent-resolved contacts. Average wait time for agent-handled contacts dropped to 3 minutes because the bot absorbed the routine volume. Annual cost savings were approximately $2.8 million in reduced agent staffing needs.