AI Adoption in Healthcare in New Zealand: A Living Whitepaper

Executive Summary

AI adoption in New Zealand healthcare is moving from isolated pilots toward structured, supervised deployment, but it remains far from autonomous clinical care.

The most significant change since the previous update is the emergence of a more formal national adoption model:

• Health New Zealand introduced a nationally consistent AI pre-implementation evaluation framework in July 2026.
• A new Centre for Digital Modernisation of Health is intended to reduce procurement, funding and governance barriers that have slowed digital and AI projects.
• Budget 2026 allocated $300 million to the Health Digital Investment Plan and $153.6 million for health-sector cyber-security capability.
• AI is now being extended beyond documentation into patient coaching, mental health triage, and symptom-checker evaluation.
• A July 2026 red-team exercise involving the national AI scribe deployment highlighted the need for adversarial testing and ongoing monitoring, even where no patient data was exposed.

The overall adoption pattern remains:

1. Documentation and workflow assistance: most mature and already operating at national scale in public emergency departments.
2. Primary-care patient engagement: early deployment through AI health coaching and integrated practice-management tools.
3. Screening and imaging: progressing through procurement, evaluation and local validation.
4. Clinical decision support: still primarily in research and trial settings.
5. Autonomous diagnosis or treatment decisions: not supported by Health NZ’s current operating policy.

The emerging New Zealand model is therefore best described as clinician-supervised augmentation, with increasing emphasis on evaluation, equity, interoperability and cyber resilience. (beehive.govt.nz)

---

What Has Changed Since 13 July 2026

1. Mental-health helplines received funding for AI-enabled triage

On 13 July 2026, the Government announced $3.35 million over four years to develop and implement an AI-enabled triage and referral tool across publicly funded mental-health and addiction helplines.

The proposed tool is intended to:

• Assess people’s support needs consistently.
• Identify the urgency of assistance required.
• Improve referrals between helplines and services.
• Reduce delays caused by rising call volumes and workforce pressure.

This is a funded development and implementation programme, not evidence that the tool is already operating nationwide. Its significance is that AI is being considered at the mental-health system front door, where inappropriate escalation, missed risk and inequitable access carry substantial consequences. (beehive.govt.nz)

2. Health NZ announced a national AI evaluation framework

On 28 July 2026, Health NZ publicised a new national framework intended to create a consistent approach to assessing AI tools before deployment.

The framework is associated with the work of the University of Auckland’s TRANSFORM programme and is being used by the National Artificial Intelligence and Algorithm Expert Advisory Group. It is designed to support structured review of proposed AI tools rather than allowing adoption to depend solely on individual hospitals, vendors or clinical champions. (healthnz.govt.nz)

This is one of the most important governance developments of the year. It suggests that Health NZ is attempting to turn AI governance into a repeatable institutional process covering:

• Clinical safety.
• Privacy and security.
• Equity and bias.
• Workflow and interoperability.
• Evidence of benefit.
• Implementation readiness.
• Post-deployment monitoring.

3. Health NZ created a digital-modernisation function to address adoption barriers

A July report on Health NZ’s new Centre for Digital Modernisation of Health described its role in reducing funding, procurement and governance bottlenecks affecting digital and AI innovation.

The Centre is intended to coordinate projects under the 10-year Health Digital Investment Plan, develop business cases, improve procurement pathways and provide specialist capability in digital delivery and innovation. The initiative reflects a recognition that AI adoption has been constrained not only by technical limitations, but also by fragmented systems and slow institutional processes. (healthcareitnews.com)

4. A national AI-scribe deployment faced a public red-team challenge

A security-testing exercise involving Heidi, the AI scribe deployed across public emergency departments, found that prompt-based manipulation could induce the system to produce content outside its intended documentation role.

Health NZ and Heidi said:

• No patient data was exposed.
• The issue was contained within an isolated test session.
• Remedial changes had been made.
• The exercise did not cause patient harm.

The incident nevertheless exposed a material governance lesson: a tool approved for administrative documentation must be tested not only for its intended use, but also for misuse, scope creep and adversarial prompting. The episode is particularly relevant because the system is used in a high-pressure environment by approximately 1,250 emergency-department doctors and frontline staff. (odt.co.nz)

5. Patient-facing AI moved into routine primary-care testing

On 12 August 2026, Tāmaki Health began rolling out an AI Health Coach at selected clinics. The tool, developed with New Zealand digital-health company Groov, is designed to provide support before, between and after appointments.

The service focuses on behaviour change and self-management for issues including:

• Diabetes.
• Gout.
• Weight management.
• Anxiety.
• Other lifestyle-related health goals.

Patients can be referred to human health coaches or clinical teams when the system identifies a need for additional support. Tāmaki Health and Groov describe the tool as an extension of the existing workforce rather than a replacement for clinicians. (healthcareitnews.com)

The early results cited by the companies are promising but should be treated as preliminary and vendor-reported. Across more than 4,500 earlier conversations, users reported improved confidence and capability in managing their health, with an average helpfulness rating of 8.5 out of 10. Independent clinical-outcome evaluation is still required. (healthcareitnews.com)

6. Health NZ progressed a symptom-checker proof of concept

Health NZ’s July 2026 symptom-checker update said international evidence justified further exploration, but not immediate procurement or national deployment.

The next stage is a controlled proof of concept examining:

• Accuracy and effectiveness.
• User experience.
• Patient behaviour.
• Effects on care pathways.
• Effects on demand across health services.
• Safety, quality and equity implications.

Health NZ expects the findings to inform a decision in mid-to-late 2027 on whether to proceed to formal procurement. This is a useful example of the country’s cautious approach: patient-facing AI is being tested, but the system is not treating international vendor evidence as sufficient for national implementation. (healthnz.govt.nz)

---

Current State of AI Adoption

1. Emergency departments: the most mature national deployment

AI scribes remain the clearest example of AI adoption at scale in New Zealand healthcare.

Health NZ announced in February 2026 that AI scribe technology was available in every emergency department, reaching approximately 1,250 doctors and frontline staff. The original rollout followed trials in which clinicians reported substantial reductions in documentation time and the ability to see additional patients during a shift. (beehive.govt.nz)

However, the deployment experience also shows that adoption is not equivalent to full automation:

• Clinicians must review and approve generated notes.
• AI output can contain errors or hallucinations.
• Patients are generally asked for consent before recording.
• Clinicians remain accountable for the final clinical record.
• The system’s intended role is documentation, not diagnosis or treatment recommendation.

The national rollout is therefore operationally significant, but it remains a human-in-the-loop documentation system, not autonomous clinical AI. (rnz.co.nz)

2. Primary care: rapid but uneven adoption

Primary care remains the most dynamic setting outside hospitals.

The Royal New Zealand College of General Practitioners reports that 41% of GPs and 16% of rural hospital doctors had either used or intended to use AI. Note-taking and scribe tools remain the dominant application. (rnzcgp.org.nz)

A 2026 New Zealand survey of 197 primary-care respondents found that 40% had experience with AI scribes. Respondents reported benefits including:

• Less multitasking during consultations.
• Time savings.
• Better patient rapport.
• Reduced documentation pressure.

The same research identified continuing gaps in:

• Patient consent.
• Understanding of vendor terms and conditions.
• Confidence about legal and ethical compliance.
• Consistency of local governance.

The evidence suggests that adoption is expanding faster than standardisation. Primary-care practices are increasingly using AI, but the quality of implementation depends on the practice-management system, vendor controls, clinician training and local leadership. (rnzcgp.org.nz)

Medtech’s February 2026 launch of Medtech AI illustrates the direction of travel. Its platform is integrated into Medtech Evolution, uses patient-record context, generates structured notes and correspondence, and requires clinician review before information is written back to the record. (medtechglobal.com)

3. Patient engagement and health coaching: an emerging category

The Tāmaki Health and Groov deployment represents a shift from AI used behind the scenes to AI interacting directly with patients.

The model is deliberately bounded:

• The AI provides coaching rather than diagnosis.
• It supports existing care plans.
• It can escalate to human staff.
• It is being rolled out initially in selected clinics.
• Expansion depends on uptake, clinical feedback and outcomes evaluation.

This category may become important because it addresses a structural problem in New Zealand healthcare: much of chronic-disease management occurs outside clinical appointments. However, patient-facing AI creates additional requirements around safety-netting, language, cultural appropriateness, digital access and escalation reliability. (healthcareitnews.com)

4. Mental health: high potential, high governance risk

Mental health is becoming a significant AI adoption area, with three distinct pathways now visible:

• AI scribes for mental-health clinicians and crisis teams.
• AI-enabled triage and referral for helplines.
• Patient-facing digital coaching and wellbeing support.

The potential benefits include faster navigation, improved consistency and additional support between appointments. The risks are equally significant:

• Misclassification of urgency.
• Failure to detect suicidality or acute distress.
• Inappropriate reassurance.
• Unequal performance across cultural and linguistic groups.
• Confusion between wellbeing coaching and clinical treatment.

The funded helpline triage programme should therefore be viewed as a major test of whether New Zealand can apply AI safely in a high-risk, high-volume environment without allowing the tool to replace clinical judgement. (beehive.govt.nz)

5. Screening and diagnostics: the leading clinical scale-up pipeline

Breast screening remains the most visible national clinical AI programme. Health NZ is pursuing procurement and validation of a mammogram-reading tool, with a proposed model in which AI would perform one of the independent reads while clinicians retain responsibility for diagnosis, follow-up and treatment decisions. A rollout from early 2027 has been proposed, subject to testing and validation. (beehive.govt.nz)

Diabetic retinal screening provides a more cautious local lesson. A New Zealand Medical Journal paper described a proof of concept designed to improve access to screening for Pacific peoples. The project encountered barriers involving:

• Digital-system integration.
• Changes to models of care.
• Clinician readiness.
• Suitability of the AI tools.
• Resourcing and organisational support.

The central conclusion was that apparently simple AI use cases can become complex once they meet real clinical workflows and equity requirements. (nzmj.org.nz)

6. Clinical decision support: research is advancing faster than deployment

The strongest example is the REVOLUTION trial, announced by the Health Research Council in June 2026.

The New Zealand-led trial has received nearly $5 million and is expected to involve:

• 50 intensive-care units across New Zealand and Australia.
• More than 24,000 patients.
• Machine-learning-guided oxygen therapy for critically ill patients on life support.

The research will test whether AI-derived treatment targets can improve survival, rather than merely whether a model can make accurate predictions. This distinction is important: New Zealand’s research pipeline is increasingly focused on clinical outcomes, implementation and safety, not just algorithmic performance. (hrc.govt.nz)

---

Governance, Privacy and Digital Foundations

Human accountability remains the official operating model

Health NZ’s current privacy statement says AI may be used for transcription, consultation summaries, document generation, staff knowledge tools and preliminary review of images and scans. It also states that AI-generated information affecting patient records or clinical decisions must be reviewed by the responsible clinician.

Health NZ does not use AI to make automated decisions about a person’s healthcare. Its guidance also prohibits staff from entering sensitive patient information into unapproved generative-AI tools or using such tools for clinical decisions or personalised patient advice. (healthnz.govt.nz)

Governance is becoming more operational

Health NZ’s National Artificial Intelligence and Algorithm Expert Advisory Group provides advice on AI development and implementation. The addition of a national pre-implementation framework should make governance more consistent across:

• Public hospitals.
• Research projects.
• Commercial vendors.
• Digital-health programmes.
• Patient-facing tools.

The July 2026 Health NZ OIA publication concerning the use and evaluation of Heidi also indicates a growing level of public documentation around AI procurement and assessment. (healthnz.govt.nz)

Cybersecurity is now a prerequisite for AI scale

Budget 2026 allocated:

• $153.6 million for national health-sector cyber-security monitoring, expertise, data-security processes and critical IT upgrades.
• $300 million to support the first three years of the Health Digital Investment Plan, including core-platform upgrades and modernisation of radiology systems.

These investments are not all AI-specific, but they are directly relevant to AI adoption. AI systems depend on connected records, cloud services, identity controls, secure interfaces and reliable data pipelines. Cyber incidents and weak vendor controls can therefore slow or undermine AI deployment even when the underlying model performs well. (beehive.govt.nz)

Interoperability remains a major constraint

Health NZ’s Shared Digital Health Record programme expects immunisation and medication data to be integrated later in 2026, while access to primary-care data is expected from mid-2027. The July 2026 SNOMED CT New Zealand release also added and updated reference sets supporting coded clinical information and interoperability. (healthnz.govt.nz)

The implication is that more advanced AI use cases—population-risk prediction, longitudinal decision support and cross-setting care coordination—will remain constrained until the underlying data environment becomes more complete, standardised and secure.

---

Case Studies

Case Study 1: Heidi AI scribe in emergency departments

Use case: Ambient transcription and automated clinical documentation.

Scale: National access across emergency departments, reaching approximately 1,250 frontline staff.

Reported benefits:

• Reduced documentation burden.
• More time available for patient care.
• Faster creation of notes, handovers and discharge documentation.
• Potential improvement in clinician wellbeing.

Implementation lessons:

• Clinician review remains essential.
• Output can contain inaccurate or fabricated details.
• Patient consent and transparency are important.
• Red-team testing should be part of approval and monitoring.
• A tool’s intended scope can be exceeded through misuse or prompt manipulation.

The case demonstrates both the practical value and the governance complexity of scaling generative AI into frontline care. (beehive.govt.nz)

Case Study 2: Tāmaki Health and Groov AI Health Coach

Use case: Patient self-management and behaviour-change support between appointments.

Initial deployment: Selected Tāmaki Health clinics.

Conditions targeted: Diabetes, gout, weight management, anxiety and related health goals.

Operating model:

• 24/7 digital access.
• Personalised coaching and action plans.
• Escalation to human health coaches or clinical teams.
• Co-design input from patients, clinicians and health coaches.

Assessment: Early engagement results are encouraging, but the available evidence remains self-reported and provider-generated. The next test will be whether the tool improves sustained engagement, service access and measurable health outcomes without creating new safety or equity risks. (healthcareitnews.com)

Case Study 3: AI-enabled diabetic retinal screening

Use case: Expanding access to diabetic retinopathy screening, particularly for Pacific communities.

Approach: Mobile ophthalmic cameras were taken into Pacific primary-care practices, with kaiāwhina trained to capture images.

Key lesson: AI could not simply be inserted into the existing service. The project encountered challenges in digital integration, workflow redesign, clinician readiness and tool suitability.

Assessment: This is an important implementation case because it shows that equity-oriented AI requires service redesign and community fit, not only a technically accurate model. (pubmed.ncbi.nlm.nih.gov)

Case Study 4: Metlifecare HEIDI rollout

Use case: Ambient documentation for registered nurses in aged residential care.

Progress: Following a four-month trial involving 12 registered nurses at Parkside Village, Metlifecare announced a rollout to three additional care homes.

Reported benefits:

• Less time spent on administrative documentation.
• More time for resident care and whānau conversations.
• Improved handover and documentation quality.

All AI-generated notes remain subject to staff review and approval. The case confirms that documentation AI is spreading beyond hospitals and GP practices into aged care, where staffing pressure and continuity of care are significant concerns. (metlifecare.co.nz)

---

Trust, Equity and Social Licence

AI adoption in Aotearoa New Zealand is being shaped by concerns that are specifically connected to Māori, Pacific and other underserved communities.

Researchers at the University of Auckland have warned that many AI systems are trained on datasets that do not adequately represent Māori and Pacific populations. This creates uncertainty about whether tools developed overseas will perform reliably or equitably in New Zealand. (auckland.ac.nz)

The main social-licence requirements emerging from the research and policy record are:

• Clear disclosure when AI is being used.
• Meaningful consent where conversations or health data are processed.
• Human accountability for clinical decisions.
• Māori and Pacific participation in design and governance.
• Protection of Māori data sovereignty.
• Testing across relevant demographic and linguistic groups.
• Accessible alternatives for people who cannot or do not want to use digital tools.
• Evidence that public benefit, rather than private commercial gain, is the primary objective.

These requirements are not peripheral ethical considerations. They are becoming practical conditions for whether AI projects can achieve patient acceptance and scale. (auckland.ac.nz)

---

Key Trends

1. AI adoption is becoming layered rather than singular

New Zealand is not pursuing one national “AI in healthcare” solution. Adoption is emerging across several layers:

• Clinician documentation.
• Practice administration.
• Patient coaching.
• Mental-health navigation.
• Imaging and screening.
• ICU treatment research.
• Population and service analytics.

2. Documentation remains the clear entry point

Scribes offer a comparatively bounded value proposition: reduce administrative burden while leaving diagnosis and treatment decisions with clinicians. This explains why they have scaled faster than autonomous clinical systems.

3. Patient-facing AI is the next major test

AI coaching, symptom checking and mental-health triage all move AI closer to direct patient interaction. These tools may improve access, but they also require stronger safety-netting, escalation and equity controls than back-office documentation tools.

4. Implementation science is becoming as important as model accuracy

The diabetic-retinal-screening experience and the national evaluation framework both show that New Zealand is increasingly asking:

• Does the tool work in local populations?
• Can it fit existing workflows?
• Is the data environment ready?
• Will clinicians use it correctly?
• Can the benefits be measured?
• Can the system be monitored after deployment?

5. Adversarial testing is now a visible governance requirement

The Heidi red-team incident shows why paper-based approvals and vendor assurances are insufficient on their own. Future evaluations will need to test not just intended functionality, but also:

• Prompt injection.
• Misuse.
• Scope expansion.
• Data leakage.
• Unsafe recommendations.
• User over-reliance.
• Failure under degraded connectivity or system outages.

6. Digital infrastructure will determine the speed of scale

AI adoption will remain limited if clinical data is fragmented, records are not interoperable, and cyber controls are weak. The Health Digital Investment Plan, Shared Digital Health Record and national terminology work are therefore foundational AI infrastructure, even when they are not marketed as AI projects.

7. New Zealand is building a cautious but increasingly coherent model

The country remains slower than some international markets in deploying autonomous clinical AI. However, its direction is becoming clearer: national governance, local validation, clinician accountability, staged procurement and explicit attention to equity.

---

Outlook for the Next 12–18 Months

The most likely areas of progress are:

• Further expansion of AI scribes into mental health, primary care and aged care.
• Evaluation of the Tāmaki Health AI coaching model.
• Development of AI triage and referral for mental-health helplines.
• Controlled testing of symptom-checker tools.
• Breast-screening AI validation ahead of a possible 2027 rollout.
• Results and implementation activity from the REVOLUTION ICU trial.
• Greater use of the national AI evaluation framework.
• More public reporting of approved or deployed Health NZ AI tools.
• Increased integration between AI products and national health-information infrastructure.

The least likely near-term development is widespread autonomous diagnosis or treatment. Health NZ’s policy settings, professional guidance and current deployment pattern all continue to favour human-supervised use. (healthnz.govt.nz)

---

Overall Assessment

As of 18 August 2026, AI adoption in New Zealand healthcare has entered a new phase: from experimentation toward controlled institutionalisation.

The country now has:

• A nationally scaled AI documentation deployment.
• Active primary-care and aged-care adoption.
• Patient-facing AI coaching in early use.
• Funded AI development for mental-health triage.
• A symptom-checker proof of concept.
• Major clinical AI research entering trial phases.
• A national evaluation framework.
• Substantial new investment in digital foundations and cyber resilience.

At the same time, the evidence does not support a conclusion that New Zealand healthcare has become broadly automated. The dominant operating model remains:

AI assists; clinicians review, decide and remain accountable.

The most important strategic shift is not simply that more AI tools are being introduced. It is that Health NZ and its research partners are developing the institutional machinery to decide which tools should be adopted, under what conditions, with what safeguards and against which measurable outcomes.

New Zealand’s healthcare AI landscape is therefore best characterised as selective, supervised and increasingly governed adoption. The next stage will depend less on whether AI can generate impressive outputs, and more on whether the health system can integrate those tools safely, equitably and sustainably into real-world care.