Report on the Achievements of NEDO Project "Research and Development for Ensuring AI Safety / Research and Development for Strengthening AI Safety"

Citadel AI Co., Ltd. (Representative Director: Hironori Kobayashi, hereinafter "Citadel AI") has been commissioned by the New Energy and Industrial Technology Development Organization (NEDO) and has been engaged in the "Project for Promoting Research and Development and Verification of AI Safety / Research and Development for Strengthening AI Safety" since fiscal year 2025.

The objective of this project is to promote research, development, and verification related to ensuring AI safety, aiming to integrally carry out the formulation and dissemination of AI safety standards and the development of AI safety evaluation and management technologies necessary for proper management and utilization of generative AI.

We are pleased to report one of the achievements: the compilation of the "Generative AI Practical Guide and Corporate Case Studies ~ A Framework for Integrating Quality, Safety, and Governance to Lead to Production Operation ~."

Based on interviews with over 20 companies, we have systematized challenges and countermeasures related to the utilization and governance of generative AI in a practical manner, while respecting existing concepts and frameworks. This document presents these as an implementation guide for companies, reproducible by other organizations, interspersed with individual case studies.

1. Background

Generative AI has become a familiar technology, and examples of AI agents being used in actual business are increasing. On the other hand, risks related to security and reputation, such as hallucinations and information leakage, have also become apparent.

Many organizations are currently grappling with these challenges in a trial-and-error manner, facing issues such as "I don't know where to start," "Considering the risks, it's hard to commit to production deployment," and "Guidelines exist, but I don't know how to specifically apply them to my company's services."

The lack of established "specific methods for quality assurance and risk management," which are indispensable for companies to responsibly provide services, has become a major barrier to full-scale adoption. There are many cases where companies cannot proceed to production operation beyond PoC (Proof of Concept).

2. Overview of Achievements

This report, "Generative AI Practical Guide and Corporate Case Studies ~ A Framework for Integrating Quality, Safety, and Governance to Lead to Production Operation ~," aims to serve as a guideline for organizational AI utilization by sharing challenges faced by companies in the field and concrete solutions and operational methods for them.

Regarding initiatives for utilizing generative AI, including AI agents, we have organized specific knowledge and practical know-how obtained through interviews with over 20 companies that are actually developing and operating AI agents, from the perspectives of "technology," "process," and "organizational culture." This has been systematized into a framework and reflected in this report.

Citadel AI itself also undertook the development of an AI agent, consistently carrying out everything from setting the problem to be solved, selecting technologies, collecting necessary data, identifying anticipated risks, and evaluating the system, thereby verifying the effectiveness of the framework.

The Appendix includes numerous examples of what purposes each interviewed company uses AI agents for and what measures they are taking against what challenges. Additionally, it shares considerations from when Citadel AI actually implemented a "chatbot designed for foreign resident support" and verified the framework's effectiveness.

For details of this report, please download it here.

3. How to Use This Report

This report, "Generative AI Practical Guide and Corporate Case Studies ~ A Framework for Integrating Quality, Safety, and Governance to Lead to Production Operation ~," systematizes and introduces methodologies as a framework for initiatives to utilize generative AI and to achieve both AI safety and AI governance, and is intended to be a reference material for organizations promoting AI utilization.

This document is not a guideline that sets rules or policies to be followed, nor a checklist that exhaustively lists initiatives to be executed. Rather, it presents a "methodology" for how to specifically achieve the "ideal state" indicated by existing guidelines such as the "AI Business Operator Guidelines," and the two are mutually complementary.

AI governance is a management issue. It is extremely important for management to recognize the significance of AI governance, exercise leadership, and work on it organizationally, while simultaneously advocating for promotion of utilization. We hope this report will be helpful when companies, organizations, or teams within companies introduce and utilize AI in response to business conditions.

Acknowledgments

This achievement is the result of work commissioned by the New Energy and Industrial Technology Development Organization (NEDO).

【About Citadel AI Co., Ltd.】

Citadel AI is a global startup from Japan that realizes the social implementation of "trustworthy AI." Development is led by an international team including engineers from Google's US headquarters. We provide our unique "AI Governance Tool" that supports the entire lifecycle of generative AI, from risk assessment at the introduction stage to guardrails during operation. It has been adopted by leading international standard organizations such as BSI and is highly acclaimed in both domestic and international markets.

Company Name: Citadel AI Co., Ltd. Location: 2-3-10 Mukogaoka, Bunkyo-ku, Tokyo, Todaimae HiRAKU GATE 3-1 Representative: Hironori Kobayashi, Representative Director Company Established: December 10, 2020 URL: https://citadel-ai.com/ja/ X(Twitter): https://twitter.com/CitadelAI