UPS
UPS uses generative AI to draft customer e-mail replies and halves agent handling time in its pilot
UPS built an in-house tool called MeRA that uses large language models to draft replies to customer e-mails, which agents confirm before sending. UPS says agents spent 50% less time resolving e-mails during pilot testing. The company receives roughly 52,000 customer e-mails a day.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
50%
Reduction in the time agents spent resolving e-mails, during pilot testing1
roughly 52,000
Customer e-mails UPS receives each day, per the company (the share MeRA handles is not stated)1
What it took
The problem
UPS gets roughly 52,000 customer e-mails a day. Agents answering them must look up shipment history, company procedures and shipper restrictions before they can reply. A single request, such as holding a package for pickup, can have several valid answers depending on tracking history and earlier delivery attempts.
What they built
UPS developed MeRA, short for Message Response Automation, in-house. It uses publicly available large language models (AI systems trained on large amounts of text) to draft replies to some customer e-mails. The agent reviews and confirms each suggested reply before it goes out. UPS says the tool can also adjust the tone of a reply based on the customer's sentiment.
How they did it
The project started in July and went into beta testing in October. It moved into limited production in November. MeRA draws on the knowledge base that agents already use, so it can apply corporate procedures and shipping data. UPS says it works through a customer request step by step before drafting an answer. Standardised workflows sit around the confirmation step.
Results
In pilot testing, UPS reports that agents needed 50% less time to resolve e-mails. UPS's chief digital and technology officer called MeRA a game changer for customer service and said it frees agents for more complex customer needs. The tool won a 2024 CIO 100 Award. UPS plans to extend it to every kind of customer contact and to other functions such as HR, sales and finance.
Lessons for your project
Keep a person in the loop. At UPS the model suggests a reply and the agent confirms it, so the agent's role shifts from writing to checking.
Feed the model the same knowledge base and live shipment data that agents already use, so replies follow real company procedures.
Pick one repetitive, high-volume channel first. UPS began with e-mail, then planned to extend the tool to all customer contact types and to HR, sales and finance.
Measure handle time in the pilot against a baseline, and revisit agent training, since the work moves from drafting to reviewing.
Caveats
- The only source is a CIO magazine feature that relays statements from UPS executives. No UPS publication or independent measurement of the results was found, so the figures are company-reported.
- The 50% figure is from pilot testing, not from full production. The article does not say how many e-mails MeRA handles or how the time reduction was measured.
- The article names no model and says only that MeRA uses publicly available large language models.
- The article gives months but not years. The year 2023 is inferred from the May 2024 article, which says UPS turned to generative AI last year and cites a Forrester post from last November, when UPS was putting the tool into limited production. The deployed date shown is that November limited production start.
Sources
- UPS delivers customer wins with generative AI
CIO · 2024-05-03 · News · cio.com
Summarized in our own words from the sources above. Spotted an error? See corrections.