AI in Public Sector

Executive Summary

The latest system-wide evidence remains the 2026 cross-agency AI survey, published on 19 August 2026. It recorded:

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

No newer cross-agency adoption survey has been published. The 2026 results therefore remain the best quantitative baseline, but new Official Information Act releases provide a more concrete view of the technology footprint: the Government Digital Delivery Agency held information on 28,728 Microsoft 365 Copilot licences as at 13 July 2026, alongside 99 Copilot Studio subscriptions and 937 legacy Copilot Studio licences. The figures do not cover all agencies, all AI products or actual usage levels. 2026 cross-agency survey AI licence OIA response digital.govt.nz

The public sector is therefore moving from experimentation to workflow integration, especially in administration, policy, communications, customer support, document processing, tax administration, health and local-government services. The dominant model remains bounded augmentation: AI drafts, summarises, classifies or retrieves information while people retain responsibility for decisions.

The principal weakness is not lack of activity but lack of system-wide evidence. The Public Service Commission does not hold a consolidated view of agency AI policies, impact assessments, operational audits, reported incidents or assurance gaps. A cross-government AI Assurance Model is expected to be published by the end of October 2026, but it is not yet evidence of a completed operating system. publicservice.govt.nz

The current assessment is rapidly expanding adoption with incomplete central assurance and limited independent measurement of public value.

What Has Changed Since the Last Update

Licence scale is now visible, although not adoption

The most significant new quantitative evidence since the 1 September 2026 edition is the release of OIA 2026-0217.

As at 13 July 2026, the Government Digital Delivery Agency held information on:

• 28,728 Microsoft 365 Copilot licences.
• 99 Copilot Studio subscriptions.
• 937 legacy Copilot Studio licences.

The figures relate to licences held through the Microsoft Cloud, Software and Services Agreement. They do not include all basic Copilot functionality available through other Microsoft products, other AI suppliers, or every public-service department. The Commission explicitly advised that a more accurate picture would require information from individual agencies. publicservice.govt.nz

This changes the interpretation of the 2026 survey. The survey showed the number of use cases; the licence disclosure shows that a substantial shared-tool infrastructure is also being funded. Neither measure establishes:

• How many licences are actively used.
• How frequently they are used.
• Which workflows they support.
• Whether staff use approved or unapproved tools.
• What productivity or service benefits have been realised.

The sector should therefore distinguish between licensed capacity, active usage, operational deployment and demonstrated outcomes.

The assurance model has moved closer, but is not yet operating as a central control

A September OIA response provides a clearer timetable for the proposed AI Assurance Model.

The response says the model is intended to support a consistent and proportionate approach to AI risk management and adoption across the public service. It is expected to be publicly available by the end of October 2026. The same response says the GDDA does not hold a central dashboard showing, across agencies:

• Adoption of agency AI policies.
• Responsible-AI officials or governance groups.
• AI registers and impact assessments.
• Pre-deployment assurance or independent review.
• Operational monitoring or audit.
• AI incidents or material control failures.
• Known assurance capability or evidence gaps. publicservice.govt.nz

This is a material qualification to the previous edition’s description of a visibility gap. The gap is not merely that information has not been published. In several areas, the central system lead does not hold the information in consolidated form.

The national AI assistant has missed the point at which a launch should have been observable

On 21 May 2026, the Minister for the Public Service and Digitising Government told Parliament that an AI assistant for Government New Zealand would be available by the end of September 2026. hansard.parliament.nz

As at 1 October 2026, the official Govt.nz app page lists emergency alerts, government services, home-screen tools and an upcoming digital wallet. It does not list an AI assistant among the app’s current features. The page was last updated on 14 September 2026 and says digital credentials are coming in October 2026, but it does not confirm a public AI assistant release. govt.nz

The Public Service AI Work Programme still describes the Govt.nz assistant as a planned deliverable. The available evidence therefore supports a more cautious conclusion than the previous edition: the assistant’s public launch is not confirmed and appears not to have been publicly evidenced by the stated end-September target. This may represent a delay, a limited release not yet documented in the main public channels, or a change in scope.

The distinction matters because the assistant is intended to be a shared, citizen-facing service rather than an internal productivity tool. Its deployment would require stronger evidence of accuracy, source traceability, accessibility, privacy protections, escalation and ongoing monitoring than a low-risk drafting assistant.

More operational case studies are now visible

The latest evidence adds several concrete examples beyond the 2026 survey.

• Inland Revenue reports that AI and optical character recognition helped process 70% of FamilyBoost claims, while an AI-powered call-summarisation tool delivered a reported 17% reduction in after-call time and was being rolled out more widely. These are agency-reported results, not independent evaluations. ird.govt.nz
• Auckland Council launched Ask Auckland Council, an AI assistant for reporting issues such as graffiti, illegal dumping, roaming dogs and missed bin collections. The council says an earlier test with 1,000 Aucklanders found that 80% considered it easy to use, compared with 42% for existing channels. These results are council-reported. ourauckland.aucklandcouncil.govt.nz
• Biosecurity New Zealand began a four-month pilot of a generative-AI tool to support development of import health standards. The announcement explicitly says key decisions remain with biosecurity experts. This is evidence of a pilot, not scaled operational use. beehive.govt.nz

Current State of AI Adoption

Adoption is broad, but concentrated in information work

The 2026 cross-agency survey remains the strongest cross-system source. It identifies the highest concentrations of activity in the social, justice, central-agency, health and finance sectors. Common functional areas include administration, digital and technology, communications, policy, project management, corporate services and human resources. Generative AI is the most common technology, followed by natural-language processing, agentic AI and machine learning. digital.govt.nz

The pattern is consistent with adoption in other public administrations: agencies are starting with work where:

• Inputs are primarily documents or text.
• Outputs can be checked by a skilled employee.
• The system does not itself determine eligibility or entitlement.
• The process can be bounded by approved data sources.
• Human escalation remains practical.

Common applications include:

• Drafting and editing.
• Summarising meetings, calls and documents.
• Information retrieval.
• Classification and triage.
• Consultation analysis.
• Customer-service assistance.
• Policy and legal research.
• Document and claims processing.
• Workflow and project reporting.

There is still insufficient public evidence that autonomous agents are operating widely in high-consequence decisions involving benefits, immigration, justice, health treatment or enforcement.

Operational maturity is increasing, but the category remains broad

The 167 operational cases reported in 2026 are an important maturity signal. They indicate that more systems have moved beyond design, exploration or proof-of-concept stages. digital.govt.nz

However, “operational” does not necessarily mean:

• Nationally scaled.
• Used by large numbers of staff or customers.
• Integrated into core systems.
• Independently evaluated.
• Reliable without extensive human checking.
• Cost-effective after licensing, integration and assurance costs.

The public evidence still lacks a common maturity framework that distinguishes a small internal assistant from a production service handling thousands of transactions. The next improvement in sector reporting should be a shift from use-case counts to operational measures such as active users, transaction volumes, human-review rates, error rates, escalation rates, cost per transaction and service-quality outcomes.

Enterprise tooling is becoming the default route to adoption

The Copilot licence disclosure suggests that much public-sector adoption is occurring through enterprise productivity suites rather than bespoke national AI systems. This approach has practical advantages:

• Procurement and security arrangements can be shared.
• Agencies can begin with familiar tools.
• Data-access controls and identity management are inherited from existing platforms.
• Staff can use AI in existing applications such as email, documents, meetings and spreadsheets.

It also creates risks. A high licence count can produce the appearance of adoption without corresponding workflow redesign or measurable value. It can also increase dependence on a small number of vendors and make it difficult to identify where AI is being used when functionality is embedded in ordinary software.

Inland Revenue is a clearer example of workflow integration

Inland Revenue provides one of the strongest public examples of AI moving beyond general experimentation.

Its 2025 annual reporting describes:

• AI and optical character recognition supporting FamilyBoost claim processing.
• AI-powered call summarisation in customer support.
• An AI fluency programme completed by approximately 2,000 staff.
• Pilots in policy, legal, software testing and staff-support functions.
• Work to consider whether tax legislation remains fit for purpose as AI changes how taxpayers obtain information and services. ird.govt.nz

The agency’s public AI page says AI is used to support productivity, decision-making, compliance-risk management and intervention design, while consequential activity remains subject to human oversight and decision-making. ird.govt.nz

This is a useful model of bounded adoption: AI is applied to high-volume administrative work and decision support, while legal accountability remains with the agency. The evidence is nevertheless principally self-reported, and the published material does not provide a full cost-benefit analysis or independent assessment of error and equity impacts.

Local government is producing some of the most visible citizen-facing deployments

Auckland Council’s Ask Auckland Council is more advanced publicly than the national Govt.nz assistant. It is available through a web interface and supports problem reporting across several categories. The council says it can interpret everyday language, understand people and place names in te reo Māori, ask follow-up questions and analyse uploaded photographs. ourauckland.aucklandcouncil.govt.nz

This deployment is significant for three reasons:

1. It connects AI to a defined service transaction rather than general conversation.
2. It uses multimodal input, including photographs.
3. It has been tested with members of the public before wider release.

The council’s reported user results are encouraging, but they should not be treated as independent evidence of long-term service improvement. The published material does not report sustained usage, correction rates, complaint rates, cost per report or whether faster reporting leads to faster resolution.

Public-sector AI is also creating new procurement and professional-assurance risks

September reporting on Wellington City Council’s $435,000 Deloitte staffing report alleged that large portions of the report were written using AI and identified serious underlying data and calculation errors. The council decided not to seek a refund, while Deloitte maintained that AI-enabled tools were used with human sign-off. This is a media-reported case rather than an audited finding about the exact extent of AI use. RNZ report unmasked.goodfriends.workers.dev

The case is relevant even though it concerns consultancy work rather than a government-built AI system. It illustrates a wider assurance problem: human sign-off is not sufficient if reviewers do not independently verify source data, calculations, assumptions and conclusions. Public agencies procuring AI-enabled professional services will need to specify:

• When AI may be used.
• What must be disclosed.
• What evidence must be retained.
• Who is responsible for checking outputs.
• What quality controls apply to quantitative analysis.
• How errors and remediation will be handled.

Governance, Policy and Regulation

The operating model remains principles-led

The Public Service AI Framework and Responsible AI Guidance for the Public Service remain the central instruments for public-sector AI governance.

They emphasise human-centred values, transparency, accountability, safety, security, fairness and responsible use. The framework is designed to guide agency decisions rather than create a single statutory approval regime. The Government’s wider AI strategy also relies primarily on existing law, including privacy, consumer-protection and human-rights obligations, rather than a dedicated AI Act. beehive.govt.nz

This flexible approach is proportionate for low-risk experimentation, but it places more responsibility on agencies to identify risk, document controls and publish information. It also makes central visibility more important. Without consistent registers and assurance evidence, principles can remain aspirational.

The AI Assurance Model is the key near-term governance test

The proposed AI Assurance Model is intended to turn broad principles into practical evidence requirements. According to the September OIA response, it is expected to support consistent risk management and proportionate adoption across the public service, with publication due by the end of October 2026. publicservice.govt.nz

A credible model should clarify:

• Which use cases require pre-deployment approval.
• How AI systems are risk-tiered.
• When an algorithm or AI impact assessment is required.
• Minimum requirements for human oversight and contestability.
• Monitoring expectations after deployment.
• How agencies should record incidents, exceptions and risk acceptance.
• What evidence should be made public.
• How suppliers and contractors are held accountable.

The model’s success will depend less on its wording than on whether agencies use it consistently and whether the GDDA can identify when controls are absent.

Central governance responsibilities remain incompletely documented

A separate OIA response found that the Public Service Commission did not hold information detailing cross-government AI governance structures, the division of responsibilities among agencies, external advisory mechanisms, or the staff and budget dedicated specifically to cross-government AI governance and risk management. publicservice.govt.nz

This does not mean that no agency has governance arrangements. It means that the central system does not hold a complete, consolidated description of them. For stakeholders, the distinction is important:

• Agency-level governance may exist.
• Cross-government coordination may be developing.
• Central oversight and reporting are not yet fully visible.

The result is a distributed model in which capability varies by agency. Larger organisations can support specialist privacy, security, data, legal and procurement functions; smaller councils and agencies may need shared tools and templates to reach the same standard.

Regulators are being encouraged to use AI in lower-risk work

In May 2026, the Government issued guidance to regulators on using AI to increase productivity. The guidance identifies applications such as detecting non-compliance, prioritising monitoring, surfacing evidence and identifying patterns in large datasets, while stating that regulatory decisions remain dependent on human judgement, legal interpretation and accountability. beehive.govt.nz

This expands the public-sector AI agenda beyond administrative assistance. Risk-based monitoring and inspection prioritisation can materially affect who receives scrutiny and how public resources are allocated. The relevant governance questions are therefore not only whether a model is accurate, but also:

• Which groups may be over-selected for intervention.
• Whether people can challenge an AI-informed prioritisation.
• Whether proxy variables reproduce historic bias.
• Whether staff understand the difference between a risk signal and a finding.
• Whether the model is monitored for drift.

Case Studies

Case Study 1: Microsoft Copilot across the public sector

Status: Large shared licence footprint; actual usage and outcomes not centrally measured.

The September OIA release provides the clearest available indication of enterprise AI capacity across government. It recorded 28,728 Microsoft 365 Copilot licences as at 13 July 2026, but did not cover every department or every AI product. publicservice.govt.nz

The case illustrates the difference between procurement scale and operational maturity. Licences may support experimentation, training, occasional use or regular workflow integration. Until agencies publish active-use and outcome data, the number should be treated as an infrastructure indicator rather than a productivity result.

Case Study 2: Inland Revenue’s administrative and customer-support use

Status: Several AI-enabled functions reported in operation or wider rollout; results primarily self-reported.

Inland Revenue reports that AI and optical character recognition support processing of 70% of FamilyBoost claims. It also reports a 17% reduction in post-call time from an AI call-summarisation tool that was being rolled out across frontline services. ird.govt.nz

The agency has combined deployment with capability-building: approximately 2,000 staff completed its AI fluency programme, and the training materials are being adapted for wider public-sector use. It also maintains public information on its AI governance approach, including human oversight and obligations under the Revenue Acts, Privacy Act and Algorithm Charter. ird.govt.nz

The main evidence gap is independent validation. The published figures do not show whether the time savings persist at scale, whether quality improves, or whether the systems affect different customer groups differently.

Case Study 3: Ask Auckland Council

Status: Public-facing service launched; reported usability testing completed.

Ask Auckland Council was launched in April 2026 to help residents report problems including graffiti, illegal dumping, roaming dogs, missed bin collections and party noise. The council says the assistant can interpret natural language, understand te reo Māori place and people names, ask follow-up questions and analyse photographs. ourauckland.aucklandcouncil.govt.nz

Before launch, the council made the assistant available to 1,000 Aucklanders reflecting the city’s demographic profile. It reports that 80% of users found the tool easy to use, compared with 42% for existing channels. ourauckland.aucklandcouncil.govt.nz

This is one of the clearest examples of a New Zealand public body moving from internal experimentation to a defined citizen-facing workflow. Its longer-term value will depend on whether reports are correctly classified, whether the assistant reduces incomplete submissions, and whether council response times improve.

Case Study 4: Biosecurity New Zealand’s import-health-standard pilot

Status: Four-month pilot announced and under development; no public evidence of scaled operation or final results.

Biosecurity New Zealand began developing and testing a generative-AI tool to support the document-heavy process of developing import health standards. The Government said the tool would assist with the work while key decisions remained with Biosecurity New Zealand experts. beehive.govt.nz

This is a representative use case for public-sector AI: large volumes of technical material, a need for evidence retrieval and summarisation, and a clear requirement for expert accountability. As at 1 October 2026, no public evaluation of the pilot’s accuracy, time savings, error rates or safety controls has been identified.

Case Study 5: Govt.nz AI assistant

Status: Planned or unconfirmed national deployment.

The 2025 pilot tested an AI assistant across government websites and found that 85% of participants considered it more efficient than their previous search methods. It also identified risks involving incorrect information, invented links, overconfidence, personal-data disclosure and confusion between navigation and transactions. digital.govt.nz

The pilot’s findings support a cautious design: the assistant should guide users to authoritative information, show sources, acknowledge uncertainty and avoid replacing transactional systems. However, the publicly available Govt.nz app information as at 14 September 2026 does not confirm that the assistant has launched. govt.nz

The project has consequently become a test not only of conversational AI, but of whether government can maintain reliable content, source ownership and rapid correction across multiple agencies.

Trends

1. The licence economy is expanding faster than the measurement system

The public service now has evidence of a large enterprise-AI licence footprint, but not of active usage or realised benefits. This creates a risk that adoption will be reported through procurement and availability rather than through actual service performance.

Future reporting should separate:

• Licences purchased.
• Active users.
• Regular users.
• Operational workflows.
• Transactions assisted.
• Human-review rates.
• Errors and escalations.
• Costs avoided or incurred.
• Service and equity outcomes.

2. Central assurance is catching up with decentralised experimentation

The AI Assurance Model, AI policy template and Public Service AI Framework indicate that central government is building common controls. The OIA releases show that central visibility is still incomplete and that the assurance model has not yet become a consolidated monitoring system. publicservice.govt.nz

The next stage will require more than guidance. Agencies need reusable assessment templates, a common risk taxonomy, supplier controls, monitoring practices and a way to report incidents and lessons across organisational boundaries.

3. Public-facing deployment is slower than internal productivity use

Internal assistants can be deployed with limited exposure if staff are expected to check outputs. Public-facing tools create a different risk profile. Errors may affect people who lack the time, confidence or knowledge to challenge them.

Auckland Council has moved ahead with a constrained service workflow. The national Govt.nz assistant, by contrast, had a stated end-September target that is not publicly confirmed as delivered. The contrast suggests that focused, bounded services may be easier to operationalise than a general cross-government interface.

4. AI is being connected to workforce and structural reform

Government policy continues to associate AI with productivity, agency consolidation and public-service transformation. The May 2026 public-service overhaul announcement explicitly linked reducing the number of departments with increasing use of AI and other digital tools. beehive.govt.nz

This creates a higher burden of proof. AI-related savings should account for:

• Licences and model usage.
• Integration and data-cleaning work.
• Cybersecurity and privacy controls.
• Human review.
• Training and change management.
• Error correction and public appeals.
• Retention of specialist knowledge.

Without this accounting, claimed productivity gains may represent work transferred to employees, contractors or customers rather than work removed from the system.

5. Cultural capability is becoming a practical implementation issue

The public service cannot treat responsible AI as a generic compliance exercise. Systems used in health, social development, justice, biosecurity and local government may affect Māori data, te reo Māori, tikanga, collective interests and relationships with iwi and communities.

The relevant question is not only whether an interface supports te reo Māori. It is whether the system’s data, assumptions, escalation routes and decision boundaries are appropriate for the people and communities affected.

6. Procurement assurance is becoming as important as model assurance

The Wellington consultancy controversy, whether or not every reported detail is ultimately substantiated, illustrates the difficulty of supervising AI-enabled professional work. A public agency may not operate the model itself, but it remains responsible for the quality of work it buys and relies on.

Contracts for AI-enabled services should address disclosure, data handling, verification, audit trails, source checking, model changes and liability. Human sign-off should be treated as a control that requires evidence, not as a substitute for evidence.

Outlook

Publication of the AI Assurance Model

The most immediate development to watch is the promised publication of the AI Assurance Model by the end of October 2026. Its practical value will depend on whether it specifies minimum evidence requirements and whether agencies are expected to report compliance or exceptions. publicservice.govt.nz

Confirmation or revision of the Govt.nz assistant

The national AI assistant now requires a clear status update. Government should distinguish among:

• A public production launch.
• A limited or invite-only release.
• A technical deployment not yet promoted.
• A delayed project.
• A revised scope within the Govt.nz app programme.

The absence of a clear public status creates avoidable uncertainty about delivery, cost and accountability.

More agency-level transparency

The central system may not hold every agency’s AI information, but agencies can publish their own registers and summaries. Useful disclosures would include purpose, data sources, degree of autonomy, human controls, supplier, lifecycle stage, evaluation results and public contact points.

Local-government examples show that public registers and plain-language descriptions are feasible, even when they are not yet standard practice.

Evidence on benefits and workforce effects

The next cycle of reporting should move beyond claims of productivity and capture whether AI changes service quality, processing times, customer effort and employee workload. Agencies should also report where pilots have been stopped or scaled back. Negative results are valuable evidence for a sector learning how to deploy safely.

More scrutiny of regulatory and high-consequence uses

As regulators and agencies apply AI to compliance, inspections, prioritisation and risk detection, the need for contestability and impact assessment will increase. Low-risk administrative assistance and high-impact prioritisation should not be governed as if they were the same category of use.

Overall Assessment

As of 1 October 2026, AI adoption in Aotearoa New Zealand’s public sector is best described as operational expansion with an emerging but incomplete assurance infrastructure.

The 2026 survey shows a substantial increase in reported use cases and operational activity. New licence data indicates that enterprise AI capacity is being distributed at considerable scale. Inland Revenue reports AI-enabled claims processing and customer-support gains, Auckland Council has launched a bounded public-facing assistant, and Biosecurity New Zealand is testing AI in technical regulatory work. digital.govt.nz

The central weakness remains measurement. There is no complete public register of deployed AI, no consolidated view of agency assurance, and limited independent evidence of financial, service-quality, equity or workforce outcomes. The planned AI Assurance Model may address part of this gap, but its publication and adoption are still future milestones. publicservice.govt.nz

The national Govt.nz AI assistant is also no longer appropriately described as an imminent confirmed launch. A ministerial deadline of the end of September 2026 has passed, while the current Govt.nz app information does not identify the assistant as an available feature. hansard.parliament.nz

The strongest path remains bounded augmentation:

• AI processes, searches, drafts and summarises.
• Public servants retain decision authority.
• High-impact uses receive stronger assurance.
• Public-facing systems show sources and uncertainty.
• Agencies monitor errors, appeals and unequal impacts.
• Procurement contracts preserve accountability.
• Māori perspectives and data sovereignty are addressed at design and deployment stages.

The public sector has moved beyond asking whether AI should be used. The more important questions are now where it is operating, what it is allowed to do, who is accountable when it fails, and whether the claimed benefits survive independent scrutiny.