Japan sets strict rules for AI agents in public administration

Japan launched on October 20, 2026, an experiment with AI agents in its ministries under the name Gennai, governed by a framework published the same day. According to the Japanese Digital Agency, this framework imposes seven safeguards: limiting agent rights, human approval for critical processing, log retention, and administrative responsibility for outcomes. Access is conditional on users—civil servants or parliamentary staff—accepting this responsibility.

The Gennai framework distinguishes two risk levels: internal tasks deemed "relatively low risk" and administrative acts or external decisions, subject to human validation and enhanced verification. Skills, the procedures executed by agents, must be assessed and controlled before being shared across ministries. Level 3 confidential information, equivalent to state secrets, is banned from Gennai, as are sensitive data or behavioral predictions.

This initiative, developed with the Japanese AI Safety Institute, aims to secure AI agent use in administration before a full rollout planned for 2027. It follows countries like France, which drastically restrict these tools in production, while Japan opts for progressive oversight and ISMAP-certified cloud services.