AI-Powered Email Assistant for CMA CGM

We built CMA CGM's AI-powered email assistant that generates recommended replies for over 100k weekly customer emails.

This project was developed as part of the Analytics Lab course at MIT Sloan, and in collaboration with CMA CGM. The course is a 9-month long project-based course in which students work in teams of 5 to solve real-world business problems for companies. The course is sponsored by the MIT Initiative on the Digital Economy (IDE).

This project won 3rd place at the MIT Initiative on the Digital Economy’s Analytics Lab Event.

Team

Team Members: Maxime Wolf, Jason Jia, Raghav Jayanthi, Nuobei Zhang, Juan Pablo Armas
CMA CGM Leadership Team: Vijay Krishnan, Abhinav Nippani, Noe Mikati, Rafaela Nunes
Project Mentor: Jeremy Toledano

Summary

A summary of the key features and benefits of the AI-powered email assistant

Problem and Solution

CMA CGM is the largest shipping company in the United States and handles more than 100k emails a week. However, they are currently processed manually by customer agents. We thus built an AI email assistant that generates recommended replies to emails based on the task to be completed and further input from the customer agent.

Problem and Solution

Integrated Processing Pipeline

This email assistant leverages both technical tools such as LLMs and NLP as well as business domain knowledge to produce relevant and useful replies. Specifically, the process begins by classifying emails based on their potential for automation. Depending on whether clear tasks are identified or not, we employ GPT-3.5 or similarity detection methods to retrieve information from the database (RAG) or generate an appropriate response.

A full view of the integrated processing pipeline is shown below:

Integrated processing pipeline

User-friendly Web Interface

We also built a user-friendly web interface that customer care agents can interact with. The interface not only generates recommended replies to emails, but also provides an option for the customer agent to correct any misidentified categories, enabling more accurate responses.

Landing page of the web interface

Impacts

Quantitative impacts

We estimate that implementing this AI email assistant can reduce time spent on email replies by 38.1%, leading to time savings of over 77k hours/year and $2.2M in productivity improvements/year.

Strategic impacts

We also had a unique opportunity to provide recommendations to the CMA CGM management team at the strategic level.

Initially, the project is about an email auto-response bot, but we quickly discovered that without a human in the loop, auto-generated LLM replies by themselves are not necessarily accurate or detailed enough. As the email is still signed off by the customer agent, it is fundamentally important for them to read through and verify replies generated by the LLM before sending them off to clients.

On the other hand, an AI email assistant adopts a different approach because it provides recommended replies to customer agents but does not automatically send replies on their behalf. This reduces time spent and also gives customer agents full control over what they send to clients.

That’s why with full support from the CMA CGM team, we made a strategic pivot to an email assistant, and we are proud to say that we have built a tool that seeks to augment, not replace humans.

Quantitative and strategic impacts

Conclusion

We believe the AI-powered email assistant is a tool packed full of amazing features, that CMA CGM’s customer agents will love to use.