Computer Systems Design and Related Services2024Generative AINLPB2B
UPS

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.

Handle Time Reduction50%
Daily Emails Handled52000 emails/day
4 min read

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

Ready to implement AI in your commerce operations?

McFadyen Digital helps teams move from case study to live implementation.

Talk to an expert →