UPS's MeRA system cuts email handle time 50% by automating responses to 52,000 daily customer messages with generative AI
Built on GPT-3.5 Turbo and GPT-4, UPS's Message Response Automation system drafts replies to roughly 52,000 customer emails per day, halving the time agents spend resolving each message during pilot testing.
Background
UPS receives roughly 52,000 customer service emails daily. Manually drafting individual responses consumed significant agent time and limited capacity for complex issues. UPS sought a generative AI solution that would draft responses automatically while preserving human review and brand-appropriate tone.
What Was Implemented
- MeRA system built on Microsoft OpenAI LLMs (GPT-3.5 Turbo and GPT-4), custom-trained on UPS business rules and knowledge base
- "Chain of Thought" sequential reasoning + sentiment analysis for context-aware replies
- Human agents review and approve all AI-drafted responses before sending
- Timeline: initiated July 2023 → beta October 2023 → limited production November 2023
- Scope: responses to "some of" the ~52,000 daily customer emails
Results
During pilot testing, MeRA produced a 50% reduction in agent email handle time. The system processes responses to tens of thousands of daily messages, freeing agents for higher-complexity customer needs. MeRA received the 2024 CIO 100 Award for IT innovation.
Lessons
- Human-in-the-loop review is essential for customer-facing AI text generation in a logistics context
- "Chain of Thought" reasoning improves response coherence for multi-step customer inquiries
- A four-month pilot-to-limited-production timeline is achievable when building on existing LLM APIs with domain-specific fine-tuning
- The architecture becomes a reusable blueprint for other enterprise AI applications