TOKYO — Japan's Digital Agency put its generative-AI procurement guideline, DS-920, into force in revised form on September 1, moving the document out of pilot guidance and into standing rule. The version now in effect is v2.0, and it lands in the same fiscal year as a GENAI deployment programme the agency describes as reaching roughly 180,000 officials.
That pairing — a binding procurement document plus a six-figure headcount — is the September story. It is a different story from the on-premises municipal model deployments and the logging template filings that surfaced in recent days; those concern specific vendors and specific local rollouts. DS-920 concerns how central government buys and runs generative AI at all.
The September 1 switch
The guideline's title covers both procurement and utilisation of generative AI for government administration, and those two halves matter separately. Procurement governs which models and services ministries may buy, and under what contract terms. Utilisation governs what officials may do with them once installed — including where outputs may and may not feed into an administrative decision.
The Digital Policy Alert tracker logs the effective date as September 1, 2026, matching the agency's own notice of the revision. What changed between the prior version and v2.0 is not laid out, in the English material, at the level of a clause-by-clause comparison. That gap is worth flagging first: the rule that governs a 180,000-person rollout is published, but a full diff of the drafting is not something a reader can pull from the agency's public pages.
Governance sits inside each ministry
The governance question sits one level above the text. In the structure the agency describes, responsibility for generative-AI use sits with the organisations that deploy it rather than in a single central approval queue. Chief AI officers inside ministries, not a central gatekeeper, are the accountable layer in that framing.
The practical consequence is divergence. Two ministries can run the same DS-920-compliant stack and evaluate it differently. Nothing in a procurement guideline forces a shared evaluation harness across agencies, and no shared benchmark is described on the pages reviewed. What was evaluated, by whom, and against which set of tasks therefore stays organisation-by-organisation — which is exactly the question a procurement rule cannot answer on its own.
What nobody can verify yet
Three things are unverifiable as of this weekend. First, the model roster: DS-920 v2.0 governs procurement, but a consultable list of approved models and services is not part of what the agency's English pages present. Second, evaluation results: if aggregate quality numbers from the pilot exist, they are not published where a reader can check them. Third, incident handling: if a ministry deployment produces a bad output inside an administrative process, the public trail for how that is escalated and recorded is unclear.
The logging question runs on a separate track. The logging template filing that circulated earlier in the month concerns a distinct instrument from a different body. DS-920's utilisation rules and that template are related in spirit but not identical in scope, and treating them as one thing would overstate what took effect on September 1.
What to watch next
Two checkpoints will test whether the September 1 date changes behaviour. The first is budget: FY2026 ministry requests for the pilot will show how much is being spent, and whether model-level detail is attached is a separate disclosure choice. The second is the agency's own programme page, which will need updating if the roughly 180,000 figure shifts as ministries onboard through the fiscal year.
For Japanese ministries, DS-920 v2.0 is now the default route. For everyone outside the building, the number that matters is not 180,000. It is how many of those deployments anyone on the outside is permitted to inspect.