AI in Public Sector in Aotearoa New Zealand: A Living Whitepaper

New Zealand’s public sector has doubled reported AI use cases and tripled operational cases, but scaling remains constrained by weak central visibility, uneven assurance, and unresolved questions about cost, data sovereignty and workforce impact.

Executive Summary

AI adoption across Aotearoa New Zealand’s public sector has entered a more operational phase. The most important new evidence is the 2026 cross-agency AI survey, published on 19 August 2026.

The survey recorded:

• 545 reported AI use cases across 59 organisations.
• 432 use cases across 42 public-service organisations.
• 113 use cases across 17 wider public-sector organisations.
• 167 cases in operational phases, approximately three times the 2025 figure.
• 31% of all reported cases classified as operational.

The evidence suggests a shift from isolated experimentation toward routine use in administration, communications, policy, project management, health, justice, social services and finance. Generative AI is the most common technology, followed by natural-language processing, agentic AI and machine learning. (digital.govt.nz)

However, adoption data remains weaker than adoption rhetoric. The survey counts reported use cases and describes reported benefits, but it does not provide a system-wide independent evaluation of financial returns, service quality, error rates, equity outcomes or public trust. The public sector is therefore becoming more active in using AI without yet having a sufficiently mature public evidence base for judging whether it is creating durable public value.

The central tension is now clear:

• Agencies are being encouraged to use AI to improve productivity and modernise services.
• Central government is building shared guidance, training and infrastructure.
• At the same time, the Government’s public-service reform agenda is linking AI with agency consolidation and workforce reduction.
• Yet the Government Digital Delivery Agency does not hold a complete register of AI tools, automated decision-making systems or related impact assessments across the public service.

The current state is best described as rapidly expanding, mostly assistive adoption under incomplete assurance.

What Has Changed Since the Last Update

The quantitative baseline has been superseded

The previous edition identified the 2025 survey as the latest system-wide baseline: 272 use cases across 70 agencies. That position changed on 19 August 2026, when the 2026 survey was published.

The headline growth is substantial: reported use cases doubled from 272 to 545, while the number of cases in operational phases increased to 167. The survey also reports that more than half of participating organisations had six or more use cases, with an average of approximately nine cases per organisation. (digital.govt.nz)

The comparison needs care. The number of participating organisations fell from 70 in 2025 to 59 in 2026, and the composition of the participating wider public sector may have changed. The results therefore show a strong increase in reported activity, but not necessarily a precisely measured increase in the total amount of AI used across every public body.

Official information releases expose a visibility gap

A newly published Official Information Act response states that the GDDA does not hold:

• A register of generative-AI or automated decision-making tools in use or procurement across the public service.
• A list of agencies using those tools.
• Privacy impact assessments, algorithm impact assessments or AI impact assessments for AI tools across the public service.
• A definitive list of enterprise generative-AI tools licensed or made available to public servants.

The response does not establish that individual agencies lack these records. It does establish, however, that central government does not yet possess a complete system-wide view of deployed or procured AI. It also says the GDDA does not routinely disclose specific use of AI in advice to Ministers, on the expectation that AI is becoming widely used. (publicservice.govt.nz)

This is an important qualification to the 2026 survey. The public service can report hundreds of use cases while still lacking the central inventory needed to determine where AI is operating, what information it processes, which systems affect individuals, and what assurance has been completed.

Workforce capability has moved from guidance to infrastructure

On 18 August, the Government Digital Delivery Agency and Leadership Development Centre launched an AI Development Series.

The free, self-paced programme covers:

• AI fundamentals.
• Safe and responsible use.
• Prompting and checking AI-generated content.
• The implications of AI for public-sector work.

The training emphasises human judgement, responsibility and oversight. It is designed to give agencies a shared baseline rather than leave each organisation to develop introductory material independently. (digital.govt.nz)

This is a practical response to the workforce gap identified in the 2025 State of the Public Service briefing, which found that approximately one-third of public servants had tried AI at work but only 14% used it regularly. The next challenge is to determine whether general literacy training translates into safe use in real workflows.

AI is being connected more explicitly to structural reform

On 31 August 2026, media reports described a previously unseen letter from Public Service Commissioner Sir Brian Roche proposing a reduction in the number of agencies from 42 to between 15 and 20, supported by common technology and substantially more AI. The Commissioner and Minister for the Public Service subsequently clarified that no final decisions had been made and that any changes would require further work and Cabinet approval. (reported account) (z-umbraco-nzb-frontend-lin-as-ae-pr.azurewebsites.net)

The proposal is not evidence of an AI deployment. It is evidence of how AI is now being positioned in the public-service operating model: not only as a productivity tool, but as part of a broader redesign involving agency structures, shared platforms and workforce composition.

That creates a higher evidential burden. AI may support consolidation, but the public record does not yet demonstrate that AI alone can deliver the claimed service improvements, savings or capability retention.

Current State of AI Adoption

Adoption is broadening across functions

The 2026 survey identifies the largest concentrations of use cases in the social, justice, central-agency, health and finance sectors. The most common functional areas are administration, digital and technology, communications, policy, project management, and corporate and human-resources activities. More than half of reported use cases also directly or indirectly support public-facing services. (digital.govt.nz)

This indicates that public-sector AI is still concentrated in knowledge work and administrative processes rather than autonomous decision-making. Typical applications include:

• Drafting and summarisation.
• Information retrieval and search.
• Document classification.
• Transcription and meeting support.
• Case preparation.
• Regulatory analysis.
• Customer-service assistance.
• Data and consultation analysis.
• Workflow automation.

Agentic AI is now visible in the survey, but there is insufficient public evidence to conclude that autonomous agents are operating widely in high-consequence government decisions.

Operational maturity is improving, but “operational” is not the same as “effective”

The 2026 survey’s 167 operational cases are a significant maturity signal. They suggest that agencies are moving beyond design and proof-of-concept work.

The category still needs interpretation. “Operational” can include a system being used in a limited business process, not necessarily a nationally scaled or independently evaluated service. The survey does not publish a detailed maturity breakdown showing:

• Number of users.
• Transaction volumes.
• Production availability.
• Error and escalation rates.
• Human-review rates.
• Costs avoided.
• Benefits realised.
• Equity or accessibility impacts.

Consequently, the public sector can reasonably claim that more AI systems are operating, but not yet that the systems are delivering measured system-wide productivity gains.

The dominant model remains bounded augmentation

Current deployments generally follow a common pattern:

1. AI processes, classifies or summarises information.
2. A public servant reviews or adapts the output.
3. An accountable official remains responsible for the decision or communication.
4. The system is introduced into a defined workflow rather than exposed as a general autonomous decision-maker.

This model is visible in Health New Zealand’s restrictions on clinical use, the Biosecurity New Zealand standards pilot, local-government information-management tools, and the Government’s own AI guidance. The model is comparatively defensible because it limits the authority delegated to the system while testing whether administrative effort can be reduced.

Health New Zealand is one of the clearest examples of scale

Health New Zealand has moved beyond isolated experimentation in several areas.

AI scribe technology was reported as live in every emergency department by February 2026, with access extended to approximately 1,250 doctors and frontline staff. The Minister of Health reported that pilot users saw, on average, one additional patient per shift as a result of time saved on documentation. Health New Zealand also reported that 80% of surveyed Middlemore staff considered the tool to improve productivity or efficiency after one month of use. These are government-reported results rather than an independently published evaluation. (beehive.govt.nz)

Health New Zealand has also established a more formal adoption and governance environment:

• Its AI and large-language-model guidance prohibits staff from entering personal, confidential or sensitive information into unapproved tools.
• Staff must not use generative AI for clinical decisions or personalised advice to patients.
• Users remain responsible for checking outputs and acknowledging AI use.
• Proposed use cases are referred to the National Artificial Intelligence and Algorithm Expert Advisory Group. (healthnz.govt.nz)

HealthX, Health New Zealand’s national digital and AI innovation programme, is also coordinating wider experimentation. Health New Zealand reports that Microsoft Copilot access had expanded to more than 2,050 licences by early 2026. Its BroPilot initiative adapts Copilot use around Māori values, tikanga and collective responsibility. These figures and descriptions are self-reported by Health New Zealand and should not be treated as evidence that all licence holders use the technology regularly or that benefits have been independently measured. (healthnz.govt.nz)

Local government shows more visible variation

Local authorities are developing some of the most transparent public-sector AI practices, but adoption remains uneven.

Hutt City Council reports 300 secure AI licences, an internal AI governance group, an AI risk-management framework and a public AI register. It has developed assistants for meeting minutes, project reporting, consultation analysis, building-consent work, traffic-management-plan review and other administrative tasks.

The Council is also testing AI for emergency communications, land-information memoranda and resource-consent workflows. It reports that invoice automation can save three to five minutes per invoice and that AI-supported consultation analysis reduced a task that might have taken weeks to two days. These are council-reported results, and several initiatives remain under development. (huttcity.govt.nz)

Tauranga City Council has separately reported that its LGOIMA AI proof of concept was productionised to process new official-information and privacy requests, including sorting and responding to email-related parts of requests. Earlier council reporting said duplicate-email processing that previously took several days could be completed in seconds. This is a local-government operational example, but the available evidence is council reporting rather than an independent audit. (infocouncil.tauranga.govt.nz)

The contrast between these councils and the central OIA response is notable. Some councils are publishing public registers and detailed use cases, while the central digital authority does not yet hold a complete cross-government inventory.

Governance, Policy and Regulation

The framework is principles-led and largely non-binding

The Public Service AI Framework remains the central policy instrument. It applies to all forms of AI used in New Zealand public services and is organised around principles including:

• Inclusive and sustainable development.
• Human-centred values.
• Transparency and explainability.
• Safety and security.
• Accountability.

The framework supports agencies in making case-by-case decisions, but it is not binding. Agencies are encouraged to align with it rather than compelled through a single statutory approval process. (digital.govt.nz)

The Public Service AI Work Programme to 2027 aims to address this distributed model through 15 initiatives across four areas:

• Common-use tools.
• Safe and responsible AI.
• Customer and partnerships.
• AI workforce.

Planned deliverables include a central AI hub, an innovation and accelerator lab, an AI sandbox, an assurance model, standardised safety mechanisms, an AI marketplace, the Govt.nz AI assistant, and expanded training. (digital.govt.nz)

The OIA response shows that some of these capabilities are still being built. The proposed central repository and assurance model should be understood as work-programme objectives, not proof that a complete governance system is already operating.

Public-facing AI remains planned in several areas

The Govt.nz AI assistant is intended to help people find government information and services through a conversational interface. The original pilot found that 85% of participants considered it more efficient than their previous way of searching, while also identifying risks involving incorrect answers, made-up links, overconfidence and confusion between navigation and transactions. (digital.govt.nz)

A May 2026 parliamentary statement said the assistant was expected by the end of September 2026. As at 1 September 2026, the available official material continues to describe it as a planned public tool and does not establish that a national production launch has occurred. It should therefore be treated as an imminent intended deployment, not current evidence of live nationwide operation. (digital.govt.nz)

Security guidance is becoming more operational

The National Cyber Security Centre published Opportunities for AI in Cyber Defence on 12 August 2026. The guidance covers potential AI use across governance, identification, protection, detection, response and recovery, while recognising that malicious actors are using AI to increase the speed and scale of attacks. (ncsc.govt.nz)

This complements earlier warnings about frontier-AI threats. At the international level, New Zealand joined Australia, Canada, the United Kingdom and the United States in discussing the national-security implications of advanced AI models, including characteristics that may require additional scrutiny and the use of national AI tabletop exercises. (dpmc.govt.nz)

The practical implication for agencies is that AI security is not a separate project. It depends on established controls for identity, access, data classification, logging, supplier management, incident response and recovery.

Privacy and data sovereignty remain unresolved pressure points

Health New Zealand’s guidance explicitly identifies privacy breaches, model-provider data retention, bias, limited te reo Māori support, cultural misrepresentation and Māori data-sovereignty concerns. It also notes that publicly available large language models may not provide adequate protections for Māori data. (healthnz.govt.nz)

The wider public service framework similarly expects agencies to consider Māori views, transparency, fairness, accessibility and the effects of AI across its lifecycle. Yet the absence of a central inventory makes it difficult for the public, Parliament or oversight bodies to see where those expectations are being applied.

This is particularly important where AI intersects with:

• Health and disability information.
• Social-development and welfare data.
• Immigration and identity systems.
• Biometrics.
• Justice and corrections.
• Māori and environmental data.
• Public-facing eligibility or service-navigation tools.

Case Studies

Case Study 1: The 2026 cross-agency survey

Status: Reported operational and planned use cases across government.

The survey is the strongest evidence that AI adoption is broadening and becoming more operational. It also shows the limits of current measurement. The survey provides counts, sectors, technologies and reported benefits, but not an independent assessment of results.

Its main value is directional: agencies are no longer only discussing whether to use AI. More systems are entering deployment, and more use cases are connected to public-service delivery. Its main limitation is that the public still cannot readily distinguish between a small internal tool, a production workflow, a limited pilot and a nationally scaled service. (digital.govt.nz)

Case Study 2: Health New Zealand’s AI scribe and HealthX programme

Status: AI scribe reported live nationwide in emergency departments; broader HealthX initiatives at different stages of testing and scaling.

The AI scribe is one of the clearest public-sector examples of movement from pilot to frontline deployment. Its use is bounded: it supports documentation but does not replace clinical judgement.

HealthX provides a more systematic model for adoption. It combines enterprise licensing, clinical leadership, evaluation, international benchmarking, cultural capability and staged implementation. Its partnership with UCLPartners is intended to draw on NHS experience in AI scribes, mental-health applications, AI-enabled diagnostics and information integration. That partnership is a capability and learning arrangement, not evidence that all those applications are deployed in New Zealand. (beehive.govt.nz)

Case Study 3: Hutt City Council’s public AI register and assistants

Status: Multiple internal tools in use; several service-facing applications under development or testing.

Hutt City Council is notable for publishing a public description of its AI programme, including licences, governance, use cases and human-review expectations. It states that staff check and amend AI outputs and that AI-generated material is disclosed where appropriate.

The Council’s approach demonstrates a possible model for local-government transparency: a public register, explicit risk-management structures and a separation between administrative assistance and final decision-making. It also demonstrates the limits of self-reporting: claimed time savings and service improvements have not been independently validated in the published material. (huttcity.govt.nz)

Case Study 4: Govt.nz AI assistant

Status: Planned national public-facing service; launch not confirmed as at 1 September 2026.

The Govt.nz assistant is strategically important because it would be among the first shared, citizen-facing AI services operating across government information. Its pilot showed that users value conversational navigation, particularly for complex or stressful interactions.

The pilot also established several design requirements:

• Show sources and links.
• Acknowledge uncertainty.
• Ask clarifying questions.
• Avoid taking over transactional systems.
• Protect users who disclose personal or sensitive information.
• Provide human and non-digital alternatives.

The project is therefore less a chatbot deployment than a test of whether government can maintain accuracy and trust while presenting information through an AI interface. (digital.govt.nz)

Trends

1. AI adoption is moving from experimentation to workflow integration

The rise in operational cases is the clearest trend. Agencies are selecting processes where AI can assist with repetitive information work and where human review remains feasible.

The next maturity question is not whether more systems enter production, but whether agencies redesign workflows around them. Simply adding AI to existing processes may increase checking, duplication and risk rather than reduce workload.

2. Shared infrastructure is becoming the preferred policy response to fragmentation

The GDDA, AI Work Programme, common-use tools, proposed AI hub, marketplace and shared training all reflect a move away from entirely agency-by-agency adoption. This is intended to reduce duplication and improve procurement, security and reuse. (digital.govt.nz)

The difficulty is that central coordination is developing faster than central visibility. A shared operating model cannot be fully effective without reliable registers, common maturity definitions, standard evaluation and consistent publication practices.

3. AI is increasingly associated with workforce redesign

Government messaging now links AI with productivity, agency consolidation and public-service reform. This makes workforce trust a material adoption issue.

If AI is presented primarily as a way to reduce staff numbers, agencies may face:

• Lower employee willingness to experiment.
• Loss of institutional knowledge.
• Pressure to deploy before assurance is complete.
• Underinvestment in training and workflow redesign.
• Difficulty distinguishing genuine productivity from workload transfer.

The stronger case for AI is not automatic job substitution. It is increasing frontline capacity, reducing administrative burden, improving access to information and allowing scarce expertise to focus on complex work.

4. Cultural and language capability are becoming practical adoption requirements

Health New Zealand’s BroPilot initiative and its national AI guidance show that cultural safety is moving from general principle into implementation practice. AI systems need to work appropriately for Māori, Pacific peoples, disabled people, people using te reo Māori and people whose circumstances are poorly represented in global training data.

This requires more than translating interfaces. It involves data governance, user research, human escalation, content ownership and decisions about which tasks should not be automated.

5. Operational evidence remains thinner than adoption claims

The sector now has better information about the number of use cases but still limited information about outcomes. Public reporting should increasingly distinguish among:

• Use-case count.
• Number of licensed users.
• Active users.
• Production availability.
• Transaction volume.
• Human-review rates.
• Error and escalation rates.
• Cost per transaction.
• Time saved.
• Service-quality changes.
• Equity and accessibility effects.
• Public complaints and corrections.

Without these measures, “AI adoption” risks becoming a count of activity rather than an assessment of public value.

Outlook

Over the next 12 months, five developments are likely to matter most.

Shared assurance and inventory

The proposed AI hub, assurance model and safety mechanisms will need to become practical operating tools. A credible system should allow agencies and the public to understand which AI systems are being used, their purposes, their data inputs, their degree of autonomy and the controls applied.

Public-facing deployment

The Govt.nz AI assistant is expected to be a significant test. Its success will depend less on the conversational interface than on content quality, source traceability, escalation design, accessibility and the ability to correct errors quickly.

Health-sector scaling

Health New Zealand is likely to remain one of the largest public-sector adopters because of its workforce scale and administrative burden. The key risks will be clinical boundaries, privacy, safety evaluation, cultural appropriateness and ensuring that AI-generated documentation does not introduce new risks into patient records.

Local-government diffusion

Councils with visible programmes may become practical sources of reusable patterns for LGOIMA processing, consent administration, consultation analysis, emergency communications and meeting support. Smaller councils may need shared procurement, common templates and regional capability because they lack specialist AI, privacy and assurance staff.

Stronger scrutiny of the AI-and-workforce relationship

As agency consolidation and staffing changes proceed, stakeholders will require evidence that AI is improving public services rather than being used to justify predetermined reductions. The central test will be whether agencies retain enough domain expertise to supervise systems and handle cases that do not fit standard patterns.

Overall Assessment

As of 1 September 2026, AI adoption in Aotearoa New Zealand’s public sector is best described as operational expansion without full system-wide assurance.

The sector has crossed an important threshold. The 2026 survey reports twice as many use cases as the previous year and three times as many operational cases. Health New Zealand has moved AI scribes into nationwide emergency-department use, local councils are deploying administrative assistants, and central agencies are building shared training, procurement and assurance mechanisms. (digital.govt.nz)

But the evidence does not yet support claims that AI is delivering large, independently verified productivity gains across the public sector. Much of the available material remains self-reported, pilot-based or descriptive. The GDDA’s lack of a complete central register and impact-assessment library is a significant governance limitation, particularly as AI becomes connected to agency consolidation and workforce reform. (publicservice.govt.nz)

The most credible path remains bounded augmentation:

• AI searches, classifies, drafts and summarises.
• AI supports public servants and clinicians.
• AI improves access to government information.
• Humans retain responsibility for consequential decisions.
• Agencies use staged deployment, monitoring and review.

New Zealand’s next challenge is to make that model measurable and durable. Progress will depend on transparent inventories, consistent assurance, total-cost accounting, robust evaluation, culturally safe design, secure data practices and sufficient human capability to supervise systems that are increasingly embedded in public-service work.