How far has "AX" at Kufu Company Group progressed in one year?

Hello! This is Kansai from Kufu Company. Exactly one year has passed since ChatGPT-4 was released to the world. I would like to take this opportunity to introduce the current status of AI initiatives within the Kufu Company Group.

The "Two AXs": Leveraging AI in both business and organization

At Kufu Company Group, we have coined the term "AX" to represent two concepts: "AI eXperience," which uses AI to improve the user experience, and "AI Transformation," which uses AI to digitize and increase the efficiency of our business operations themselves. Please see below for details.

Tokubai Flyer Comparison, embodying AI eXperience

As for the output of the former, AI eXperience, the most clear example is the "Tokubai Flyer Comparison" feature of our shopping service "Tokubai," one of the Kufu Company Group's core businesses.

Tokubai Flyer Comparison is a service that allows users to list only the items they are looking for from the flyers of nearby stores listed on Tokubai, making it easy to find stores where they can purchase items at a great price.

At Tokubai, our core value proposition is to have retailers such as supermarkets and drugstores provide special sale information, which consumers then view, increasing opportunities for them to visit the stores.

Currently, while we sometimes receive special sale information from retailers as text data including product names and prices, it is more common for them to upload flyer images as they are. Therefore, to create a feature that allows users to know things like "which store nearby has the cheapest broccoli," we have had to analyze the flyer images ourselves.

This feature, which has long been in high demand by users, was only made possible by incorporating AI.

To briefly introduce the technical elements, we first apply an object detection model called DETR (DEtection TRansformer), developed by Facebook AI Research, to the flyer images to extract product areas (② in the diagram). At the same time, we use Google's Vision API to perform OCR (③ in the diagram) to convert them into text, thereby obtaining the combination of what text exists in which location on the flyer image.

Furthermore, we use RoBERTa, a named entity recognition model from Meta AI, to recognize entities such as whether each piece of text represents a "product name" or a "quantity" (④ in the diagram). Creating the correct training data for ② and ④ is steady but extremely important work, and we utilized Labelbox for that annotation.

Through this project, we have accumulated extensive knowledge of AI implementation at a practical level. We have already begun moving to actively apply this know-how to other features within Tokubai and to the businesses of our group companies.

For example, at the household account book service Zaim, we found that in many cases, product names recorded in the account book can be classified as expected just by porting the category classification model mentioned above almost as is.

As for other businesses within the group, we are also considering applying it to the comprehensive wedding information site "Minna no Wedding." Minna no Wedding contains many images of "wedding cost breakdown" sheets posted by users. In the future, we anticipate expanding features that meet the needs of people who want to hold a wedding, such as comparing how much was spent on which items across different wedding venues.

Tokubai Flyer Comparison itself is also currently undergoing functional improvements. Specifically, by assigning categories such as "fresh food" or "beer" to product names using RoBERTa (⑤ in the diagram), we expect to be able to increase the number of products compared and reduce management costs.

Furthermore, we are challenging ourselves to extract the relationships between texts, such as which product costs how much and where it was produced, while using LayoutLM, a natural language processing algorithm from Microsoft Research (⑤ in the diagram). If this can be realized, it is expected that entity recognition accuracy will improve, and we will be able to meet user needs such as wanting to sort by price.

3 steps for AI Transformation in business efficiency

The other AX, "AI Transformation," can be thought of as the AI version of "DX." It aims to have every member within the group naturally incorporate AI, resulting in more efficient business operations.

When exchanging information with AI promotion leads at companies outside the group, I get the impression that every company is struggling with this part while making steady progress. In the midst of this, we have devised our own approach and are promoting adoption through the following three steps.

1. Know: Input basic knowledge about AI

The Kufu Company Group consists of about 20 group operating companies, with approximately 700 employees. The job types are diverse, ranging from engineers and designers to sales and administrative staff. Simply issuing a command to "use AI" is not enough to permeate AI into everyone's work.

Therefore, at an early stage in April 2023, we held a generative AI workshop that spanned the entire group. We divided the approximately 100 participants into teams of 10 and repeated explanations of generative AI and Q&A sessions. In the second session held in the summer, we divided teams by job function and discussed how to apply it to specific tasks.

Furthermore, the AX Promotion Department of Kufu AI Studio, which promotes AI, has taken the lead in establishing small, regular meetings with key members from each business to discuss how AI can be utilized.

2. Spread: Share AI initiatives in weekly meetings

At Kufu AI Studio, where the majority of the group's engineers and designers belong, we hold an all-hands meeting every Monday morning. As part of the content, we have set up an "AI Topic" section to introduce AI cases that each department is working on.

Although this is currently an initiative only within Kufu AI Studio, we hope to spread it to the entire group in the future.

3. Try: Held a group-wide AI hackathon with over 50 participants

Finally, I would like to introduce the AI hackathon that is currently underway. This is a project where participants themselves try to solve problems they face in their daily work using AI by actually getting hands-on experience.

More than 50 participants have gathered, and their job types are diverse, including not only engineers but also designers, sales, and back-office staff. There are participants joining alone as well as mixed teams spanning multiple group operating companies, making it a new place for interaction.

Participants receive subsidies for GPT Plus to use GPTs and for the OpenAI API. The presentations are in April, so we are halfway there, but practical and interesting projects are currently underway.

More than half of the entire group uses AI in their work

As a result of these efforts, in a group-wide survey conducted in November 2023, 50.9% of all employees answered that they "use AI in their work." At "Kufu Shizuoka," a group operating company, there have been cases where the representative, who is not an engineer, created GPTs themselves and is utilizing them in daily operations.

Furthermore, the figure for engineers was high at 91.6%, and if you exclude members who had just joined the company or had other circumstances, it was approaching 100%. This is largely due to the fact that Kufu AI Studio, which is responsible for product development for the group, has introduced GitHub Copilot company-wide.

Summary

Over the past year, the development of generative AI and LLMs has been remarkable, and with better models appearing one after another, it has been a difficult situation even for specialized teams to keep up. This is even more true for members in roles where AI is not at the core of their work, such as sales or user support.

However, in order to effectively incorporate AI into both business and the organization, it is extremely important to know what can be done now and to imagine how it can be integrated into work and products. The Kufu Company Group intends to continue working on creating mechanisms that allow the entire organization to improve from the bottom up, while incorporating cutting-edge AI into practical work across the entire group.

Kufu AI Studio, which is responsible for product development for the Kufu Company Group, is actively hiring engineers involved in AI, as well as those in data-related roles and those responsible for technical research!