The Ministry of Digital Affairs published a generative-AI procurement checklist Monday that Taiwan’s central agencies must attach to service tenders closing after Nov. 1, tightening disclosure rules on training data, inference locations, and benchmark scores before a year-end rush of chatbot and document-automation contracts.
Checklist contents
Agencies must require vendors to state whether models run on sovereign cloud regions, list subprocessors with access to citizen data, and submit Taiwan AI Evaluation Center benchmark results for tasks named in each statement of work. Contracts above NT$10 million need independent red-team summaries and explicit human-review steps for outputs used in enforcement or benefits decisions.
Why timing matters
Digital Affairs Minister Herming Chiou told reporters that ministries accelerated AI pilots after Mid-Autumn staffing returned, creating uneven guardrails. The checklist harmonizes language with the AI Basic Act’s implementing rules still awaiting final Cabinet sign-off. Procurement officials said the document stops short of banning foreign models but makes residency exceptions visible to auditors.
Vendor field
Taiwan system integrators welcomed standardized security questionnaires, which cut bespoke paperwork per agency. Global cloud providers must pair with local partners to meet data-sovereignty attestations; at least three majors updated Taiwan region documentation within 48 hours of the checklist release. Startups argued benchmark fees should be subsidized for pilots under NT$3 million, a request MODA said it would review in November office hours.
Civil society input
Transparency advocates praised mandatory publication of model cards on agency websites after award, though they wanted stronger limits on law-enforcement use cases. Labor unions representing government call-center staff asked for retraining funds when chatbots absorb routine inquiries; the checklist encourages but does not mandate workforce plans.
Implementation
The Public Construction Commission will embed checklist references in its e-tender template by Oct. 15. Agencies with active pilots must file gap analyses by Dec. 1 or pause expansions. MODA will host weekly clinics in Taipei and Kaohsiung for procurement officers rewriting RFP scopes for code assistants and meeting summarizers.
What agencies should do now
Officers drafting year-end chatbot renewals should map each user journey to a risk tier in the checklist, attach benchmark IDs, and budget red-team time before award rather than after go-live. Missing elements will trigger return reviews from MODA’s digital governance office, which already flagged two September tenders for vague data-retention clauses. The checklist does not solve every AI ethics debate, but it gives civil servants a shared scorecard before budgets commit.