Improving team productivity for a more connected future

Data bottlenecks prevent actionable insights

NTT DOCOMO, one of Japan’s leading telecommunications carriers, hopes to contribute to society’s development, enrich people’s lives and drive progress through advanced technologies and a customer-focused approach to their business. With their eye on these aspirational goals, NTT DOCOMO focused on enabling employees internally with GenAI tools to make the customer experience the best it could be. Part of this work was implementing a proprietary LLM-based infrastructure, called the LLM value-added platform, to provide functions such as improved convenience for users when using LLM, security, reliability and safety, and access to various LLMs through web applications. As of February 2025, the LLM value-added platform has approximately 10,000 MAUs and 3 million monthly calls, and it is being used in a variety of use cases.

For example, the network quality improvement team used the internal LLM tool to analyze customers’ social media data, processing 20,000 entries daily to detect issues faster. The department related to the contact center improved efficiency by utilizing LLMs to create FAQs from chat logs of inquiries from DOCOMO shop staff to the support center.

It is vital to continuously analyze and improve how the infrastructure is utilized to ensure not only the provisioning of infrastructure but also its effective use by employees across various departments. Initially, tools like Excel and Jupyter Notebook were used to review user activity logs. However, as the LLM value-added platform expanded, several challenges emerged.

One major issue was the explosive growth in data volume. As the number of daily active users (DAU) increased, the sheer size of log outputs overwhelmed local tools like Excel and Jupyter Notebook, rendering them incapable of handling the data effectively. Another challenge involved maintaining data quality and integrity. Manual processes frequently led to inaccuracies, duplication and incomplete data, undermining the reliability of analyses and hindering decision-making. Additionally, data privacy and security became critical concerns. Logs stored locally on employees’ development PCs through Excel and Jupyter Notebook posed significant risks, as they often contained sensitive and confidential information. This decentralized approach heightened the possibility of data leaks, further complicating efforts to manage the infrastructure effectively.

Databricks effectively solved these challenges by automating manual processes and efficiently handling large volumes of logs with speed and security. Remarkably, this was achieved with minimal development and operational effort, ensuring a streamlined and low-maintenance approach.

“As hyperscaler cloud services become increasingly complex, the Databricks workspace control screens offer simplicity and ease of use. The seamless interface allows for smooth navigation between the workflow, Databricks SQL, collaborative notebooks and Unity Catalog — significantly reducing development stress,” Chijia Huang, from Service Innovation Department at NTT DOCOMO, explained. “In fact, once we started building the dashboard in Databricks, the coding and workflow setup itself progressed smoothly to completion, aside from the time needed for internal reviews and other processes.”

Boosting team productivity with Databricks

NTT DOCOMO uses the Databricks Data Intelligence Platform to perform everything from data capture, formatting and processing for dashboard construction to the creation and publication of data marts for analysis. For more advanced analysis, they use Mosaic AI Model Serving to deploy Azure OpenAI’s GPT-4o model for log analysis, enabling the extraction of entities from user prompts and responses, which are then categorized to provide detailed insights into usage patterns and application areas. Databricks plays a key role in seamlessly integrating with these LLM models. “With Databricks, we can centrally manage LLM log data and quickly analyze how users are using our LLM,” Issei Nakamura, Assistant Manager, Service Innovation Department, at NTT DOCOMO, explained.

The first step was to build the analysis environment, which was very straightforward with Databricks. Data scientists can start the SQL warehouse and compute resources needed to run code with a single click, enabling them to build and manage the development environment independently, without requiring an infrastructure engineer. Databricks Notebooks facilitate seamless sharing among users and mutual reviews and commenting, which enhances the efficiency of pair programming. The code was then implemented and deployed as a production job using Databricks Workflows. Tasks that previously required manual execution — such as updating, version control and error detection — are now fully automated, significantly improving operational efficiency. Additionally, fine-grained access controls were applied to the created data marts and other assets using Unity Catalog. Beyond the ability to set permissions for entire catalogs, schemas and tables, row- and column-level masks and filters can also be configured, ensuring efficient and robust data security and authority management. 

In addition, the company has adopted Databricks’ AI/BI Genie, a new conversational analysis tool that goes one step beyond dashboards, with a chat-style interface that allows data to be analyzed and visualized using only natural language. This ease of analysis has empowered the company to respond quickly to the analysis needs of management and other stakeholders. Beyond simplifying analysis, Genie incorporates governance through Unity Catalog, ensuring a secure and well-managed analysis environment for the entire organization. NTT DOCOMO’s LLM value-added platform team is initially deploying Genie within their own department, but this governance functionality will enable them to deploy Genie in other departments in the future. The LLM value-added platform team at NTT DOCOMO is deploying Genie within their own department, but this governance function will make it easy to deploy it in other departments in the near future.