This page has been translated using TexTra by NICT. Please note that the translation may not be completely accurate.If you find any mistranslations, we appreciate your feedback on the "Request form for improving the automatic translation ".

The 5th Advanced AI Utilization Advisory Board

Last Updated:

Outline

For an overview of the Committee, please see Advanced AI Utilization Advisory Board .

Event Information

Date:
Tuesday, July 14, 2026 from 13:00 to 15:00
Location:
Office of the Prime Minister's Office - Online

Proceedings

  • 1. Opening
  • 2. Proceedings
    • 1. Outline of Periodic Report on generative AI System by Each Ministry and Agency
    • 2. Trends in generative AI in Japan and other countries
    • 3. Examples of generative AI utilization by national governments
    • 4. Explanation from Sakuma members
    • 5. Future Plans
  • 3) Closing

Data

Reference Documents

Summary of proceedings

(1) Outline of Periodic Reports on generative AI System by Each Ministry and Agency

The Secretariat made a report in Appendix 1 on the results of an analysis of the state of generative AI utilization by the various ministries and agencies.

Questions and Opinions

The main questions and opinions of the attendees are as follows.

  • In the note on page 2, it is stated that "two cases where the scope of users is not stated are excluded." Please tell us, if you know, for what purpose these two cases assume a AI in which they are used.
    • (Secretariat) In some cases, the scope of users was not described, but this does not mean that the scope was intentionally omitted.
  • (Member SAKUMA) In practice in the private sector, there are many patterns in which the introduction of AI tools, such as "shadow AI," is not understood. According to a questionnaire survey of about 50 companies conducted by the AI Governance Association, due to circumstances such as a AI function being incorporated as a new update into a system that is not procured in AI, there were very many cases in which their AI use cases were not fully understood, and 64% of them recognized an oversight. In addition, regarding the status of AI usage by external partners, approximately 87% of all companies indicated an oversight. Although it is a dark figure, I would like to ask if there have been any near-misses such as an oversight. In addition, I think it is very important to build a system to check for an oversight in the future.
    • (Secretariat) Since the Periodic Report focuses on information systems, the generative AI System has been covered in this report. On the other hand, for services subject to agreement on terms and conditions, the bureau in charge of information systems at the Cabinet Office and each Ministry and Agency supervises the use of such services and monitors service trends. Under the current situation, Confidentiality class-1 information can also be used by external services if procedures are completed. In Document 1, it is reported that the number of cases has decreased by division or office. However, these are free AI, etc. used for purposes such as investigating and organizing information on the Internet, and some of them are relatively easy to access. In addition, even for business systems, there are reports on those using external products. Based on these facts, it is considered that the AI used by the Cabinet Office and each Ministry is recognized. It is necessary to continue to thoroughly confirm whether there are any omissions in understanding.
    • (Chairman Kadobayashi) For the management of invisible AI function, such as the management of suppliers, a CAIO should be established to appropriately manage risks, but attention should be paid in the future.

(2) Trends in generative AI in Japan and other countries

The Secretariat has reported on the trend of generative AI in Japan and other countries from Appendix 2.

Questions and Opinions

The main questions and opinions of the attendees are as follows.

  • Regarding page 9, I think it is a very important point of view what definitions of AI agent and agent AI are discussed in documents of other countries. Are the definitions consistent among the respective documents? If you do not understand at this meeting, I would like you to touch on the details of the definitions in the future.
    • (Secretariat) This year's Advisory Board is an introduction, so it is only a summary introduction. In the next and subsequent Advisory Board meetings, in order to make the discussions more concrete, we would like to create and introduce a detailed version comparing individual items.
  • (Member YOSHINAGA) In the comparison of the introduction of AI agents in companies and organizations on page 12, the Japanese introduction rate is slightly lower than that of other countries. My understanding is that this indicates a cautious attitude. When we interviewed pharmaceutical companies, they said they are concerned about cyber-security risks such as information leakage due to the introduction of AI agents. If such concerns are dispelled, I think the introduction of AI agents will progress further. For that purpose, support from the central government is necessary. I think it is necessary to consider the extent of cybersecurity that needs to be implemented and policies that can reduce risks as much as possible. In particular, the healthcare sector is an area with the greatest potential for AI. Since it can be utilized for the development of new drugs and the discovery of treatment methods, it is necessary to promote the use of AI agents. On the other hand, data is an asset, so it is siloed. As a solution to this situation, I think it is necessary to newly consider a mechanism for aggregating data in the central government, like the National Institutes of Health (NIH) in the United States.
    • (Secretariat) First, it is the situation that OWASP and others also recognize that the risk of leakage increases when information that can be referenced by AI agents and operations from the outside are meshed. In our Guidelines, from the viewpoint of governance of AI agents in governments, whether or not measures can be uniquely defined in the Normative is one of the topics for this year, and we believe that we will aim to embody Normative elements in guidelines through guidance and others. Rather than discussing from scratch, we will refer to discussions in the OECD, AISI, AI Guideline for Business and others, and we will work from the perspective of how to apply them in governments. Second, in the AI Basic Plan as well, from the viewpoint of "vertical AI," policies have been indicated for the nation to develop and support AI specialized in each field. First of all, based on these policies, we believe that efforts will be materialized by the entire administration.
  • (Chai Shan Member) In introducing AI agents, from the perspective of governance, it is important to control the scope of authority and the scope of information to be referenced. From the perspective of promoting the use of AI agents, from the perspective of a human-in-the-loop, I believe it is also important not to excessively request human confirmation. In this revision, please be careful not to interpret that there is no problem if the reader first adds human confirmation.
    • (Secretariat) Since there are concepts such as on-the-loop and out-of-loop as variations of the human-in-the-loop, as you mentioned, we believe it is important to establish stages according to tasks, etc.
    • (Chairman Kadobayashi) In the context of AI for Science, there are initiatives by MEXT, etc., but those initiatives are considering minimizing human intervention. On the other hand, since the scope of these guidelines is the utilization of AI by the government, infallibility is central. Therefore, it is considered that human intervention will be induced. The intensity of human intervention needs to be discussed in the future.
  • On page 12, it is said that 30% of Japanese companies have introduced AI agents, but there is a gap from the actual feeling. When checking the original document, the impression is that it is limited only to large companies. Please describe such restrictive conditions for parameters.
    • (Secretariat) Understood. We will reflect your advice.

(3) Examples of generative AI utilization by governments

A presenter from Ministry of Defense reported on examples of advanced generative AI use and application in Ministry of Defense, including the use of AI agents, in Appendix 3, and members asked questions and offered opinions.

(4) Explanation from Sakuma members

Material 4 reports on the governance of AI agents in the private sector from member Sakuma.

Questions and Opinions

The main questions and opinions of the attendees are as follows.

  • (Member NAGANUMA) On the whole, it is consistent with the views of the Japan Business Federation. I understand that the governance of AI agents is as stated in the contents of the announcement. In this context, I believe that "how people are involved in decision making" is the biggest challenge for private companies. As shown in Exhibit 2, overseas, information on "what should be" of providing companies is emerging, while in Japan, each company in the private sector is searching for its direction. Overseas companies are moving quickly, and some of them have already started discussions on international standards, forums, summits, etc. There are concerns that definitions and frameworks are formed overseas ahead of others, and we believe that early action is necessary. I would like to ask what kind of activities the AI Governance Association is conducting internationally. In addition, I have the impression that there are still few domestic Japanese companies that have provided detailed guidelines for agent-type AI or stipulated human involvement, but I would like to ask if there are any companies that are taking concrete measures.
    • (Member Sakuma) Regarding the first point, international efforts, we recognize that it is still a future challenge as a whole. Regarding standardization, our organization also has members who operate ISO and business operators who are involved in international standardization at the company level, and we are cooperating as appropriate. We believe that we must be firmly involved in the board of international standardization in the future. We also intend to advance efforts such as entering into discussions with international organizations at the working level. Since we have already been cooperating at the industry level, such as our partnership with GARP, a financial risk management organization that operates globally, we would like to further communicate our ideas as a Japanese company. The second is about what kind of efforts are being implemented by companies that are currently starting. For example, as shown on page 20 of Exhibit 4, there are cases where efforts such as dividing the level of autonomy are being implemented. In the AI governance process in the past, attention was focused on how to use the output of AI, and there are cases where the concept of autonomy is added as a multiplication formula to create rules. On the other hand, there are still cases where human involvement is excessively required and efficiency is not achieved. It is necessary to continue to examine to what extent human oversight is necessary.
  • (Member YOSHINAGA) Recently, we discussed AI governance in East Asia and support systems for small and medium-sized enterprises at an international conference. Regarding human resource development at small and medium-sized enterprises, it was pointed out that general security training does not sufficiently cover AI risks, and there was an issue of how much money should be spent. It was also pointed out that there are almost no engineers who understand the vulnerabilities unique to AI and can write safe code. The AI Governance Association believes that returning the knowledge acquired by large enterprises to small and medium-sized enterprises and promoting smooth human resource exchanges will lead to the promotion of AI agent utilization in Japan as a whole.
    • (Member SAKUMA) The reality is that many of the members of the AI Governance Association are large companies. As you pointed out, we believe that it is very important to disseminate knowledge from large companies. As shown on page 13 of the material, we published a report aimed at developing professionals specializing in AI governance, "Toward the Establishment of Job Functions for AI Governance Professionals (AIGist) - Primary Report: Current Situation Analysis and Clarification of Points at Issue -," but the current situation is that the personnel who are responsible for AI governance in-house are insufficient regardless of whether they are large companies or small and medium-sized enterprises. It is important to organize core skill sets, collaborate with experts in various fields, and catch up on new technologies, and we would like to discover and introduce examples of collaboration among AI governance personnel in the future.
    • (Member NAGANUMA) Personnel is mentioned in the Basic Plan for AI, and the Task Force is also considering what kind of personnel are needed in which areas. It is desirable that efforts in the private sector be coordinated and aligned with governmental actions such as the Basic Plan for AI. I would like you to consider this so that it will eventually have a positive impact on SMEs as well.
  • (Nabatame member) Recently, there have been more opportunities to discuss corporate risks associated with AI agents, especially in dialogues with overseas companies. It goes without saying that management is important when using AI at one's own company, but at the same time, a society in which customers use AI agents has also progressed. There are cases where general consumers make purchases through AI agents at supermarkets, and cases where economic activities are conducted on an Agent-to-Agent basis. It is necessary to discuss the responsibility of companies for heteronomous AI agents in the future.
  • (Members of the Chai Shan) The JDLA is also promoting initiatives related to governance. There are some overlaps with the initiatives of the AI Governance Institute, and it seems that a certain market outlook has become clear regarding the minimum measures that should be taken. In government procurement, it is important to confirm to what extent the outsourcee is committed to safety and governance as a company. It is considered that the contents can be reflected in the revision of this guideline.
    • (Sakuma) It is important to secure trust as a company. Since it is difficult for the government to conduct detailed surveys, we would like to consider a reliable framework that can be used as a certain level of heuristics, while taking into account the degree of risk, while also introducing it from the private sector.
    • (Chairman Kadobayashi) Governance is often talked about as an organizational theory, but it is necessary to proceed with the study based on the technical aspect as in the case of cloud governance. Reference 2 showed that governance is incorporated into function for enterprises, such as AI management infrastructure. Human resources who can appropriately control such technology will be required. I recognize that systems such as licenses and qualifications related to governance will be developed in the future. It would be good to have a dialogue with cloud service providers and consider expanding the qualifications system for AI agents.
    • As pointed out, it is necessary for companies to develop personnel with knowledge of governance who can play a central role in AI governance.
  • On page 28 of the material, the action items of the AI Governance Navigation are arranged. Are these measures required to be taken by the company or organization, or are they required to be taken under the assumption of a specific generative AI system or AI agent?
    • (Sakuma, a member of the Council) Regarding the item "Measures for Each Risk Area," the items that should be addressed differ depending on the use case, so the responses were based on the representative use cases of each company. Regarding the items such as rule clarification, human resources development, organizational structure development, and transparency, it is possible to make decisions by organization, so we asked whether initiatives are being implemented as an organization. Regarding the activities of the entire organization, group governance (subsidiaries, etc.) is included, and the results are summarized on page 35 of the materials.
    • (Okada) In the self-diagnosis by companies in each area, the score for ensuring transparency and accountability was relatively low.
    • (Member: Sakuma) In the first place, there was an internal discussion that the action items themselves might be difficult. For example, regarding "ensuring clarity," the content required by the EU AI Act was taken into consideration, but the methodology has not yet been sufficiently established, so the score was not that high. The score for the item "whether risk management is evaluated from a third party perspective" tends to be particularly low, and it may be difficult for SMEs to respond to it. On the other hand, there are cases where alternative responses can be made based on continuous feedback from users, etc., so there is room for improvement in the future.
    • What initiatives is the AI Governance Association (JIIA) undertaking in relation to the AI alignment issue? For example, is JIIA undertaking initiatives in relation to constitutional AI, or are any initiatives currently underway?
    • (Member SAKUMA) It is necessary to consider approaches to reflect human intentions in AI by dividing them into two layers. The first is direct alignment, in which corporate policies are directly incorporated into AI, and the second is social alignment, in which these policies are aligned with the rights and interests of external stakeholders. Regarding the latter, social alignment, we believe that capturing stakeholders' risk awareness is an important part of transparency and accountability. On this point, although each company is developing policies, it is still in the process of development, and the Association is also making improvements through study meetings between companies and stakeholders. On the other hand, for direct alignment, measures for each risk area shown on page 42 of the materials fall under this category. Specifically, individual measures are being promoted according to the risks of use cases, such as the development of guardrails for personal information protection and the implementation of red teaming tests as a measure against bias. On the other hand, alignment in the development of the foundation model needs to be further discussed in the future.
    • We recognize that the demarcation of responsibilities between AI providers and user companies is also important.

(5) Future plans

The Secretariat reported the future schedule in Appendix 5.

Questions and Opinions

There were no questions or opinions from the participants regarding future plans.

Before the closing session, Mr. Parliamentary Vice-Minister for Digital Transformation KAWASAKI stated that he would like to have in-depth discussions on AI governance and data management in governments, and to have in-depth discussions based on de-facto standards while promoting the use of AI agents in Digital Agency.

Greater Than