The Ministry of Economy, Trade and Industry released a procurement checklist Friday that municipalities must reference when renewing generative-AI chatbots, tightening how local governments evaluate vendors as hundreds of citizen-facing bots approach contract end dates this fiscal year.
What the checklist requires
The document translates national-government gen-AI procurement rules into municipal bid language. Planners must score suppliers on whether training data includes citizen PII, how prompts and outputs are logged for hallucination review, and what happens to fine-tuned weights if a contract ends. It mirrors Digital Agency “procurement” and “contract” check sheets version 2.0, but adds METI-specific prompts on industrial-data spillover when bots integrate with local SME support desks.
Municipal chief information officers said the checklist is not optional guidance for entities that receive METI smart-city subsidies; auditors will expect completed scorecards in renewal files starting October 1.
Why chatbot renewals triggered action
Japan’s 2023–2025 wave of experimental LINE and web chatbots—many built quickly with offshore large language models—now faces second-round procurement. Early contracts emphasized uptime and Japanese language quality, not model cards or retrieval grounding. Incidents in which bots invented welfare deadlines or misquoted ordinance numbers pushed the Digital Agency to harden central-government rules in May 2025; METI’s municipal addendum closes a gap for cities that fund bots through regional revitalization grants rather than internal affairs channels.
Internal affairs ministry materials updated in December 2025 already urged municipalities to adopt governance structures and confidentiality handling; METI’s checklist operationalizes those themes into pass-fail bid criteria.
Evaluation categories vendors will see
Bidders must disclose whether inference runs in Japan-hosted environments, whether sub-processors can change foundation models mid-contract, and how citizens opt out of transcript retention. The checklist asks municipalities to require incident notification within seventy-two hours for prompt-injection breaches and to cap auto-learning on live citizen chats without human review.
For chatbots tied to tourism or foreign-resident services—common in Kyushu and Hokkaido pilots—METI adds questions on multilingual hallucination rates and escalation paths to human phone desks during disasters.
Relationship to other ministries
METI does not supersede Digital Agency templates; planners are instructed to attach both national contract sheets and METI’s municipal annex. Chief AI officers at prefectural capitals said the duplication is intentional: METI wants visibility into bots that touch commerce and manufacturing helplines, not only resident records held by internal affairs systems.
InfoHandle’s editorial desk argued separately that guidelines need enforcement teeth; this checklist is the procurement-side response, focused on what evaluators can score before a vendor wins a renewal rather than criminal penalties after failure.
Vendor and city reactions
Systems integrators that sold first-generation bots are rewriting proposals to include retrieval-augmented generation with citation links to ordinance PDFs—a feature few 2024 pilots shipped. Smaller startups without Japan-region hosting face joint bids with domestic cloud providers to meet data-residency rows.
Cities with tight fiscal years welcome standardized scoring because it shortens council debate; others worry that strict model-change clauses will lock them into aging APIs. METI officials briefing local IT associations said waivers are possible for pilots under fifty million yen if councils document risk acceptance—a path several rural towns are considering.
What happens next
Training modules for municipal evaluators will roll out through METI’s regional bureaus in October, paired with sample scoring sheets for a fictional garbage-collection chatbot and a manufacturing subsidy Q&A bot. Vendors that already serve national agencies are adapting their model cards to the municipal annex to avoid rewriting proposals for each city.
Privacy commissioners in two prefectures asked whether checklist logging requirements conflict with local ordinances limiting retention of resident questions; METI’s FAQ says municipalities may shorten retention if councils publish the shorter period in bid documents, but hallucination audits still need thirty-day minimum logs.
METI will collect anonymized scorecards in March 2027 to report renewal outcomes to the Council for the Promotion of a Digital Society. Municipalities that skip the checklist risk losing smart-city matching funds. For residents, the practical effect is slower bot relaunches with more paperwork—and, planners hope, fewer confident wrong answers about taxes, childcare slots, and disaster shelters.
