PyCon JP 2024

Your locale preferences have been saved. We like to think that we have excellent support for English in pretalx, but if you encounter issues or errors, please contact us!

AI-powered Automatic Replies in Customer Support: Precision-Focused Approach at Mercari
2024-09-27 , 4F Track4

Have you ever had a frustrating experience with a customer support chatbot? Issues like misunderstandings, irrelevant responses, and endless loops without reaching a human agent are common. So why do we use AI for sending automatic replies to customer inquiries at Mercari? Does it benefit customers? How do we ensure a positive experience?

Along with answering these questions, in this talk, we will look at the design of our highly precise system using machine learning for automatic replies. We will focus on Python content in sections about introduction to transformers, fine-tuning pre-trained transformer models on Japanese text, and the design and training of our ML model. Lastly, we'll discuss our A/B testing methodology and the impact on business metrics from using automatic replies.


Outline:

  • Introduction (2 minutes)
    • Who am I?
    • Problems with chatbots in customer support
  • So why do we use AI for customer support? (3 minutes)
    • Examples of repetitive and trivial inquiries suitable for automatic replies
    • Examples of inquiries where no action is needed, and only information needs to be conveyed
    • Benefits to customers: instant response and resolution
    • Benefits to businesses: cost savings
  • Ensuring a good experience with AI replies (3 minutes)
    • Importance of precision in sending automatic replies
    • Design choices for a good customer experience
      • One-click escalation to human support agents
      • Automatic replies only sent as the first response
  • Designing a precise system for automatic replies (17 minutes)
    • Analyzing raw data to identify common inquiry patterns (1 minute)
    • Utilizing metadata for extra context (transaction status, shipping method, item price, etc.) (1 minute)
    • Introduction to transformers and their usage for text classification (5 minutes)
    • Introduction to fine-tuning pre-trained transformer models for text classification on Japanese text (5 minutes)
    • Designing and training ML models using inquiry texts and metadata (3 minutes)
    • Calibrating precision using output thresholds (1 minute)
    • Running A/B tests and impact on business metrics (1 minute)
  • Conclusion (2 min)
  • Q&A (3 min)

Audience:

Developers, data scientists, ML engineers, and business leaders interested in using AI to solve business problems including enhancing customer support operations.


Outcome:

By attending this talk, the audience will gain insights into how AI can be effectively used to enhance customer support while maintaining a positive customer experience. They will learn about the importance of precision in sending automatic replies, the process of analyzing customer inquiry data, and the use of machine learning techniques for text classification. The talk will also discuss the impact of implementing AI-powered automatic replies on business metrics.


Why did you choose this topic?

I have been working as an ML engineer in the customer support domain for last 5 years at Mercari. In the last year and a half, our team has been experimenting (with LLMs and MLMs), developing, and A/B testing this feature to send automatic replies to customer inquiries while ensuring a positive customer experience. After a few A/B tests, finally we had a 100% release of this feature in production earlier this year. And we are already doing more iterations to improve it. So I wanted to share the insights from this project with Python community members.

Knowledges and know-how the audience can get from your talk

By attending this talk, the audience will gain insights into how AI can be effectively used to enhance customer support while maintaining a positive customer experience. They will learn about the importance of precision in sending automatic replies, the process of analyzing customer inquiry data, and the use of machine learning techniques for text classification. The talk will also discuss the impact of implementing AI-powered automatic replies on business metrics.

Prior knowledges speakers assume the audience to have

Basic knowledge of machine learning.

Audience experiment

Beginner

Language of presentation

English

Language of presentation material

English

See also: Slides

Hi, I am working as a Machine Learning Engineer at Mercari, Inc. for the last 5 years in Tokyo, Japan. At Mercari, I work in the Customer Support domain and lead the exploration and application of ML, NLP, and LLMs in this domain. I work with Japanese text and so far I’ve developed features like sending automatic reply to customer inquiries using LLMs and MLMs, routing of customer inquiries, and template suggestion. Outside of work, I like exploring specialty coffee shops, photography, playing piano, reading (especially poetry), working out in the gym, and occasional hiking trips.