Fujitsu began routing select property-tax inquiries on Hokkaido and Okinawa prefectural hotlines through its Takane large language model Sunday, testing whether a domestically hosted chat stack can cut hold times without misstating levy rules.

What the pilot covers

Callers asking about payment deadlines, installment plans, and late penalties first hear a recorded menu, then can opt into a voice bot backed by Takane. The model pulls answers from prefecture-published FAQs and ordinance summaries stored on Fujitsu’s private cloud in Japan, not from the open internet.

Complex cases—appeals, liens, or questions about inherited property—still transfer to human staff. Supervisors can listen to live calls and override the bot if it hesitates twice on the same topic.

Why Hokkaido and Okinawa

Hokkaido’s tax office sees seasonal spikes when second-home owners file from Tokyo, while Okinawa handles bilingual questions from U.S. base communities. Both prefectures already run Fujitsu case-management software, simplifying integration.

Digital Agency officials said the pair offers geographic spread without exposing the largest cities to untested automation on day one. If metrics hold through December, Osaka and Fukuoka may join a phase-two list.

Accuracy controls

Fujitsu trained Takane adapters on 14,000 anonymized call transcripts vetted by prefecture lawyers. Each answer includes a citation code clerks can verify on an internal dashboard. The system refuses to estimate individual tax bills from voice descriptions of land size, directing callers to online calculators instead.

Weekly red-team exercises inject trick questions about special agricultural exemptions. Failure rates above two percent trigger a model rollback, according to contract language shared with municipal unions.

Privacy and records

Voice audio is deleted within seventy-two hours unless a caller opts into a callback. Metadata—wait time, topic tags, satisfaction scores—feeds a prefecture analytics portal with personal identifiers stripped.

Privacy advocates asked whether LLM logs could be subject to information-disclosure requests. Both prefectures said they classified bot transcripts as administrative drafts exempt from blanket release, mirroring treatment of human note-taking during calls.

Cost and staffing

Fujitsu priced the six-month pilot at 480 million yen combined for both prefectures, including GPU hosting and on-site trainers. Union representatives secured a clause that no full-time phone staff will be cut during the trial; any savings must fund digital literacy classes for seniors.

Other vendors, including NEC and NTT Data, watch closely because local governments plan bundled AI procurements next fiscal year. A successful Takane demo would strengthen Fujitsu’s case after it absorbed heavy mainframe-to-cloud migration costs.

What success looks like

Officials will publish monthly scorecards on resolution rate, transfer rate, and complaint volume. They will also track whether online payment completion rises when callers get faster answers about barcode slips.

If hold times fall without accuracy incidents, Fujitsu intends to offer the stack to all forty-seven prefectures through a catalog contract Digital Agency maintains for standardized municipal IT. Failure would push policymakers back toward simpler retrieval bots without generative phrasing.

Metrics dashboard

Prefecture IT directors will see live dashboards comparing bot-handled calls with human-handled baselines, including average handle time and callback requests. Unions secured the right to pause the pilot if satisfaction scores fall below seventy percent for two consecutive weeks.

Fujitsu competitors may bid on phase two, but early integration work gives Takane an incumbency edge unless accuracy slips draw fire from opposition lawmakers during Diet questioning.

Language access

Okinawa call centers will keep English-speaking staff on every shift while the bot handles Japanese queries. Hokkaido plans Ainu-language brochures at counter desks even though the voice bot does not yet support indigenous languages.