← Back to projects
Eurowings · Lufthansa Group3-month production rollout

Enterprise AI Customer Support

Conversational AI for enterprise customer support

Designed and deployed a customer-support chatbot over enterprise help content, reducing repetitive questions while maintaining citation-backed responses.

RAGGPT-4LlamaIndexConversational AIAzure

Business Problem

Customers repeatedly asked about baggage, cancellations, refunds, check-in, and special assistance, overloading support channels.

Architecture

  • Azure OpenAI GPT-4 with retrieval-augmented answer generation
  • Hybrid search over FAQs, policy documents, and support content
  • Session-aware conversation memory for follow-up questions
  • Fallback handoff to human support for low-confidence answers

Technical Decisions

  • Used citation-enforced prompts to make answer quality auditable and safer in customer-facing flows.
  • Kept context memory scoped to the session to preserve relevance without over-retaining noise.
  • Added confidence thresholds and escalation paths instead of forcing the model to answer every request.

Production Challenges

  • Controlled hallucinations with retrieval-only prompting and explicit citation enforcement.
  • Tracked conversation success, escalation rate, hallucination rate, and user feedback.
  • Built a golden QA dataset and combined human review with LLM-as-a-judge evaluation.

Outcomes

Reduced repetitive support workload
Improved first-response speed
Raised customer self-service completion

Related Links