UPS
genAI agent assistance (reply drafting)
UPS deployed MeRA, a genAI assistant that drafts replies to customer emails (validated by an advisor) and cuts handling time by 50%, on about 52,000 emails received each day.
Objective
Reduce the time advisors spend drafting email replies against a massive volume, without degrading quality or tone, keeping the human as final validator.
The deployment
MeRA (Message Response Automation) is a genAI assistant that drafts replies to UPS customer emails. The system reads the incoming message, analyzes sentiment, applies UPS's business rules and knowledge base, then proposes a complete reply that the advisor confirms or edits before sending. The agent's role shifts to validation, which speeds up handling while keeping a human in control. UPS started from the GPT-3.5 Turbo and GPT-4 models via Azure OpenAI, with a sequential reasoning framework trained on its own rules.
Results Proof C
Established press (CIO) naming UPS, with figures and a quote from the executive, confirmed by a second specialist-press pickup (Supply Chain 24/7). External recognition via the CIO 100 Award 2024. A pilot figure rather than a consolidated financial result, hence C.
How it works
Documented architectureThe stack in detail
- plateforme Microsoft Azure OpenAI Service Cloud platform that provides UPS with the OpenAI models in an enterprise framework.
- llm OpenAI GPT-3.5 Turbo et GPT-4 MeRA's starting models to read the email, assess sentiment, and draft the reply.
- outil MeRA (Message Response Automation) UPS custom system: a sequential reasoning framework trained on the in-house business rules and knowledge base.
- infra File email du centre de contact UPS Inbound channel (about 52,000 customer emails a day) in which MeRA proposes its drafts to the advisor.
How it runs, concretely
For ops teams-
1Receiving the email customer
A customer message arrives in the customer service queue.
-
2Analysis and drafting AI
MeRA reads the message, evaluates sentiment, applies UPS's rules, and drafts a complete reply.
-
3Human validation customer service
The advisor confirms or adjusts the proposed reply before sending.
-
4Sending and loop customer service
The validated reply goes to the customer; the edits feed the improvement of the system.
The quality of the business rules and knowledge base that feed the model. If those rules are outdated, the proposed draft becomes wrong and the advisor has to rewrite everything, which erases the time gain.
How your customers perceive this type of use
Sourced studiesUn ecart net separe les annonceurs des consommateurs : 77% des annonceurs voient l'IA positivement contre 38% des consommateurs (Yahoo/Publicis, 2024). Les mesures implicites confirment le rejet declare : en EEG, les pubs generees par IA produisent une activation memorielle plus faible que les pubs traditionnelles et sont decrites comme agacantes, ennuyeuses et confuses (NIQ, 2024). La disclosure a un effet ambivalent : elle augmente fortement la confiance quand elle est remarquee (Yahoo/Publicis), mais 27% des jeunes consommateurs disent faire moins confiance a une entreprise dont la pub est creee par IA (IAB, 2024).
Acceptance conditions
- Une disclosure visible : quand la mention IA est remarquee, la confiance globale envers l'entreprise augmente de 96% (Yahoo/Publicis 2024)
- Une qualite visuelle suffisante : les visuels IA de basse qualite augmentent l'effort cognitif et distraient du message (NIQ 2024)
Red lines
- Le contenu IA non declare puis identifie : 72% des consommateurs disent que l'IA rend l'authenticite difficile a etablir (Yahoo/Publicis 2024) et les marques utilisant des pubs IA sont plus souvent jugees inauthentiques ou non ethiques par les consommateurs que par les dirigeants (IAB 2024)
- Les mannequins et personnes generes par IA : 46% des consommateurs n'en veulent pas dans la publicite, l'inquietude premiere etant les standards de beaute irrealistes (Attest 2025)
Sources: Yahoo / Publicis Media (terrain Ebco) 2024 · IAB (avec Attest) 2024 · NIQ (NielsenIQ) 2024 · Attest 2025
How to replicate
Inference, not sourcedData prerequisites
- a history of labeled customer emails
- an up-to-date knowledge base and business rules
- a sentiment model or an LLM with tone analysis
Org prerequisites
- advisors trained in the validator role
- a correction loop to improve the model
- a compliance framework for handling customer data
Possible stack
- a commercial LLM via cloud
- integration with the ticketing / email tool
- a business-rules layer
The plan, step by step
- Step 1Choose a high-volume, well-documented email category; gather the corresponding history and business rules.Deliverable: A corpus of labeled emails + documented, up-to-date business rules.
- Step 2Connect the LLM (Azure OpenAI or equivalent) to the rules and knowledge base, generate drafts in shadow mode.Deliverable: Drafts compared to agents' real replies, gaps analyzed.
- Step 3Launch the pilot with human validation on a group of agents and measure handling time.Deliverable: A time-per-email table: pilot group vs control group.
- Step 4Set up the correction loop: agents' edits feed prompts and rules.Deliverable: A documented rise in the rate of drafts accepted without edits.
- Step 5Extend to neighboring email categories with tone and compliance guardrails.Deliverable: Production on the target email domain with weekly quality tracking.
First step: Take a high-volume, well-documented email category, generate drafts validated by the agent, measure the time saved.
The same pattern, proven in other industries
Allstate
genAI drafting of customer communications with human review
Google for Education
genAI assistant integrated into an institutional education suite, admin-managed
Spotify
genAI voice DJ (synthetic voice plus generated commentary)
Sources
- S1 UPS delivers customer wins with generative AI Established press archive pending
- S2 UPS Cuts Email Response Time in Half with AI Automation Secondary archive pending
An error, newer info, a source?
This page lives on its accuracy. If a figure has moved, if the deployment has changed, or if you have a higher-quality source, tell us. Every sourced correction is verified before publication.