Local LLM Introduction Case Study: AI Utilization Balancing Confidential Information Protection and Business Efficiency | Matsuo Lab Startup Athena Partners with Jōyō Bank to Enhance Bank Operations with a Local LLM That Does Not Transmit Confidential Information Externally
📋 Article Processing Timeline
- 📰 Published: March 28, 2026 at 01:15
- 🔍 Collected: March 28, 2026 at 21:59 (20h 43m after Published)
- 🤖 AI Analyzed: April 15, 2026 at 04:48 (414h 48m after Collected)
Athena Technologies, Inc. (Headquarters: Bunkyo-ku, Tokyo, Representative Director: Takeshi Abe, hereinafter "Athena"), a startup originating from Matsuo Lab (Note 1), is pleased to announce the development and provision of "JOYO AI AGENT," a specialized operational AI agent using a local LLM, for Jōyō Bank, Ltd. (Location: Mito City, Ibaraki Prefecture, President: Tetsuya Akino, hereinafter "Jōyō Bank").
This project, following the PoC (press release here) announced in July 2025, has completed the development of a common infrastructure in a cloud-based private network environment and the implementation of the "Translation/Masking" function. Furthermore, to pursue greater practicality, additional verification concerning "Automated Business Document Creation" and "Approval Document Review" is currently being piloted.
In this article, we will present initiatives centered on these four use cases as a DX case study in banking operations that balances strict security with field convenience.
■ Trajectory from PoC to "JOYO AI AGENT" Construction
Since the start of the PoC announced in July 2025, we have pursued the dual goals of ensuring data safety and improving business efficiency through generative AI in banking operations, where strict information protection is required. As a result, the practicality of the translation and masking functions, as well as the effectiveness of AI utilization in other specialized operations, have been confirmed.
Based on these verification results, we are proceeding with the development of "JOYO AI AGENT" and have also commenced a PoC for additional feature development to pursue further practicality.
■ Three Design Principles for Optimizing AI Implementation in Banking Operations
In building "JOYO AI AGENT," we have adopted a design that emphasizes three pillars: safety, scalability, and convenience, in order to balance the strict security required in banking operations with practical usability in the field.
【Safety】 "Air-gapped Environment" protecting confidential information, inaccessible from the internet. Instead of using external generative AI services, we have built a local LLM on a private cloud (Azure VNet) that is physically isolated from the internet. By sending data to a dedicated isolated environment in this way, we eliminate the risk of leakage of confidential information and personal information, making it adaptable to banking operations where AI utilization was previously difficult.
【Scalability】 Design of a "Local LLM Operation Infrastructure" that accelerates horizontal expansion. As the core of the system, we have built a common local LLM and a "Local LLM Operation Infrastructure" responsible for login, log management, etc. Features of "JOYO AI AGENT" such as translation and masking are all implemented as "applications" running on this common infrastructure. This design allows for speedy and efficient development of only additional features when introducing AI to new operations in the future, without needing to rebuild the entire system from scratch.
This project, following the PoC (press release here) announced in July 2025, has completed the development of a common infrastructure in a cloud-based private network environment and the implementation of the "Translation/Masking" function. Furthermore, to pursue greater practicality, additional verification concerning "Automated Business Document Creation" and "Approval Document Review" is currently being piloted.
In this article, we will present initiatives centered on these four use cases as a DX case study in banking operations that balances strict security with field convenience.
■ Trajectory from PoC to "JOYO AI AGENT" Construction
Since the start of the PoC announced in July 2025, we have pursued the dual goals of ensuring data safety and improving business efficiency through generative AI in banking operations, where strict information protection is required. As a result, the practicality of the translation and masking functions, as well as the effectiveness of AI utilization in other specialized operations, have been confirmed.
Based on these verification results, we are proceeding with the development of "JOYO AI AGENT" and have also commenced a PoC for additional feature development to pursue further practicality.
■ Three Design Principles for Optimizing AI Implementation in Banking Operations
In building "JOYO AI AGENT," we have adopted a design that emphasizes three pillars: safety, scalability, and convenience, in order to balance the strict security required in banking operations with practical usability in the field.
【Safety】 "Air-gapped Environment" protecting confidential information, inaccessible from the internet. Instead of using external generative AI services, we have built a local LLM on a private cloud (Azure VNet) that is physically isolated from the internet. By sending data to a dedicated isolated environment in this way, we eliminate the risk of leakage of confidential information and personal information, making it adaptable to banking operations where AI utilization was previously difficult.
【Scalability】 Design of a "Local LLM Operation Infrastructure" that accelerates horizontal expansion. As the core of the system, we have built a common local LLM and a "Local LLM Operation Infrastructure" responsible for login, log management, etc. Features of "JOYO AI AGENT" such as translation and masking are all implemented as "applications" running on this common infrastructure. This design allows for speedy and efficient development of only additional features when introducing AI to new operations in the future, without needing to rebuild the entire system from scratch.