Summary

  • Government organisations reported 545 AI use cases in 2026, against 272 in 2025, even though fewer organisations took part in the latest survey.
  • Data sovereignty remains a stated challenge, but the published survey does not identify each case’s data, supplier, processing location or control terms.

New Zealand’s public-service AI count more than doubled in a year. The Government Digital Delivery Agency says 59 organisations reported 545 use cases in 2026: 42 public-service organisations accounted for 432, and 17 wider public-sector bodies for 113. The previous survey recorded 272 cases from 70 organisations. The latest count grew while participation fell, so it signals more reported activity among respondents, not a census or a direct measure of how many people are affected.

The activity is moving beyond planning. The 2026 report counts 167 cases in operational phases, three times the 2025 number. It says AI is being used most often for administration and digital functions, followed by communications, policy, project management, corporate work and human resources. Agencies reported efficiency and service-delivery benefits. The survey also says public acceptance, reliability and data sovereignty remain complex challenges.

That pairing gives the survey its significance. It counts activity and names a control problem, but the public report does not map individual use cases to the information they process, their providers, the location of processing, retention and training terms, or the official who can authorize or stop each use. Without those links, readers cannot tell whether a locally stored record remains within practical government control when it is sent to an AI service, retained by a supplier or used to shape a decision.

New Zealand has invested in one visible layer of control. The Government opened a NZ$326 million all-of-government data centre in June 2025. The GCSB operates it at RNZAF Base Auckland; the government said it would hold some of agencies’ most sensitive information for 25 years. That is a substantial storage capability. The public announcement does not say that the 2026 AI cases run there, however, and a data centre’s location alone cannot establish which provider sees a prompt, how long inputs persist or whether an output can be challenged.

The government’s Public Service AI Framework calls for traceable data, human accountability, transparency, risk management and public disclosure of AI use where relevant. Yet the same page says agencies are encouraged to align with the framework; it is not binding. The Government Chief Digital Officer’s work programme to 2027 lists an AI Assurance Model and Toolkit as planned work. The Privacy Commissioner separately expects agencies to assess privacy impacts, conduct provider due diligence, engage Māori where affected, and ensure personal or confidential information is not retained or disclosed by a tool.

Those instruments point toward practical questions, but the survey does not show whether each of its 545 cases answers them.

This is not proof that New Zealand’s agencies lack safeguards. Privacy law and existing public-service duties still apply, and individual agencies may hold detailed assessments that the aggregate survey does not publish. The narrower finding is that adoption is now countable at system level, while public evidence of case-level data control is not visible in the same report.

The distinction matters for ordinary users. A person applying for help, paying a tax bill or seeking a public service needs to know whether an AI tool only drafts an internal summary or shapes a decision about them; what information it receives; which agency remains responsible; and how the person can correct an error or ask for human review. Counting use cases cannot answer those questions.

The next assurance step should connect a public inventory to these control facts without exposing sensitive security details. For each deployable case, agencies could disclose its purpose, stage, data sensitivity, accountable owner, provider-retention rule, human review route, privacy assessment status and incident or appeal path. Exceptions for security-sensitive uses can be explained, but a count without a testable control record cannot settle the sovereignty concern that the survey itself identifies.

Sources