AI in Healthcare

Executive Summary

Since the 1 September 2026 edition, the sector has seen meaningful movement in four areas:

• Health New Zealand has moved its radiology modernisation programme from advance notice to a live request for proposals. The procurement includes an imaging-AI orchestrator, but no national imaging-AI service is operating yet.
• Tāmaki Health has published early results from its AI Health Coach pilot, reporting greater outreach to high-need patients and more patient-reported information. The results are provider-reported and not from a controlled evaluation.
• Health NZ has begun a HealthX pilot of AI-assisted rostering for medical radiation technologists, extending AI adoption into workforce operations.
• Privacy enforcement and patient communication have become more prominent. The Privacy Commissioner issued compliance notices to Health NZ and Manage My Health after the 2025 portal breach, while RANZCR published new patient resources explaining AI in imaging and radiotherapy.

The overall pattern remains selective and supervised:

1. Documentation: operating at national scale in public emergency departments and expanding through primary care, mental health and aged care.
2. Patient engagement: moving from demonstration to early operational use, but with limited independent evidence of health outcomes.
3. Workforce operations: entering pilots through rostering and scheduling.
4. Imaging: progressing through national infrastructure procurement and local equipment capability.
5. Clinical decision support: concentrated in trials and research.
6. Autonomous diagnosis or treatment: no evidence of broad deployment in New Zealand.

The most important change is not a breakthrough model. It is the gradual construction of the conditions under which AI can be used safely: integrated systems, privacy controls, patient communication, procurement standards and evaluation capacity.

What Has Changed Since the Last Update

Radiology AI has entered formal procurement

Health NZ opened the Radiology Digital Modernisation Programme request for proposals on 25 September 2026. The RFP is open until 19 November 2026 and covers four capabilities:

• An integrated reporting environment.
• An image archive and enterprise viewer.
• An enterprise imaging integration layer.
• An imaging AI orchestrator.

Health NZ describes the programme as a national digital investment intended to replace fragmented legacy systems and create a single connected radiology service across Aotearoa. gets.govt.nz

This is a material progression from the future procurement opportunity described in the previous edition. It demonstrates that Health NZ has moved from market engagement and architectural planning to a competitive procurement process.

It is still not evidence of deployment. The RFP does not establish that a supplier has been selected, that the platform is funded through contract award, or that any AI application is operating in clinical practice. The immediate significance is strategic: imaging AI is being treated as a capability that depends on national workflow, archive, integration and governance infrastructure.

The procurement also creates a future marketplace question. An orchestrator could allow Health NZ to connect and assess multiple AI applications rather than committing the entire system to one diagnostic model. That approach may support competition and model replacement, but it will require robust validation, monitoring for performance drift, cybersecurity controls and clear accountability for errors.

Primary-care AI coaching has produced its first local operating results

Tāmaki Health published early results from its AI Health Coach rollout with Groov on 14 September 2026. The six-week data covered six clinics and six human Health Coaches. Tāmaki Health reported that AI-supported outreach engaged seven times more patients than a comparable previous human-only text campaign, increased collection of patient-reported health information by 8%, and generated 105 additional patient touchpoints without additional workforce time. tamakihealth.co.nz

The pilot targeted people with poorly controlled diabetes who had not attended a clinic for at least six months. Tāmaki Health reported that Māori and Pacific peoples made up 56% of enrolments, above their representation in the participating clinics. Patients rated the tool’s helpfulness at an average of 8.9 out of 10.

These findings are relevant because they move patient-facing AI beyond a product demonstration into routine primary-care activity. They also suggest a model in which AI extends the reach of a human workforce rather than attempting to replace it.

The evidence remains limited. The comparison with previous human-only outreach was not a controlled trial, the cohort was relatively small, and the results were published by the provider and technology partner. The data show engagement and workflow signals, not sustained improvement in diabetes control, reduced hospital use or improved equity. Tāmaki Health also says that multilingual use remains in beta and is undergoing further safety review. tamakihealth.co.nz

AI has entered a Health NZ workforce-operations pilot

A HealthX contract with University of Auckland spin-out RosterLab is being piloted with medical radiation technologist teams at Counties Manukau, Waikato, Starship and Waitematā.

RosterLab’s system uses optimisation and AI-assisted workforce-management tools to balance staffing levels, skills, fatigue, employment requirements and staff preferences. The company reported that earlier work with Middlemore’s medical radiation technologists reduced roster-production and management time from about 120 hours per month to less than 20 hours. Those figures are company-reported and relate to previous work rather than the current HealthX pilot. auckland.ac.nz

This is not clinical AI, but it is healthcare-sector adoption. Rostering affects service capacity, staff fatigue, continuity and fairness. It also illustrates a lower-risk route into operational AI: the system recommends or generates options, while human coordinators remain responsible for decisions and exceptions.

The current status should be described as a pilot, not a national deployment. Public information does not yet establish the pilot’s impact on staffing gaps, overtime, fatigue incidents, patient waiting times or workforce satisfaction.

Patient communication has become part of AI implementation

On 10 September 2026, the Royal Australian and New Zealand College of Radiologists launched patient resources explaining how AI may be used before, during and after imaging and radiotherapy.

The resources address potential benefits, consent, privacy, protection of health information and the continuing role of clinicians in decision-making. Their publication reflects a practical shift: patient understanding is being treated as an implementation requirement rather than a communications exercise after deployment. ranzcr.com

This is especially relevant to radiology, where AI may influence image acquisition, triage, reconstruction, reporting or treatment planning without being visible to the patient. Explaining which role the system plays—and which decisions remain human—is important for informed consent and trust.

Privacy enforcement has raised the bar for digital-health suppliers

On 23 September 2026, the Privacy Commissioner issued compliance notices to both Manage My Health and Health NZ following the cyberattack on the patient portal in December 2025. The Commissioner identified seven areas where security protections were ineffective. Health NZ must complete its required changes by 29 January 2027, while Manage My Health has until 31 August 2027. privacy.org.nz

The notices are not specific to AI, but they directly affect AI procurement. AI tools often add another supplier, cloud environment, data flow, integration point or subcontractor to an already complex health-information ecosystem. A security failure in a connected portal demonstrates why vendor assurances cannot substitute for organisational due diligence, access controls, monitoring and contractual accountability.

The practical consequence is likely to be more scrutiny of:

• Identity and access management.
• Multifactor authentication.
• Data retention and deletion.
• Supplier and subcontractor controls.
• Audit logging.
• Incident response.
• Data residency and cross-border disclosure.
• The ability to suspend or disable an AI system safely.

A previous regulatory assumption has been superseded

The previous edition referred to the Therapeutic Products Act 2023 as if its main provisions, including regulation of software as a medical device, were coming into force in September 2026. That is no longer the correct position.

The Act was repealed in December 2024. The Ministry of Health is developing a Medical Products Bill, which is intended to replace the Medicines Act 1981 and provide a modern framework for medicines and medical devices. Cabinet decisions have addressed innovative medical products, clinical trials and software as a medical device, but the new Bill is not yet a fully operative AI-specific regulatory regime. health.govt.nz

This leaves current healthcare AI adoption governed primarily through existing privacy law, professional obligations, procurement controls, clinical governance and, where applicable, medical-device requirements.

By contrast, the previous edition’s discussion of Information Privacy Principle 3A has progressed as described. The Health Information Privacy Code 2020 was amended from 1 May 2026 to reflect the new obligation concerning indirect collection of personal information. privacy.org.nz

Current State of AI Adoption

Documentation remains the only clearly scaled national use case

The strongest operating deployment remains Heidi’s AI scribe across public emergency departments. Health NZ reported in February that access had been completed across all emergency departments, reaching about 1,250 doctors and frontline staff. It also said more than 1,000 additional licences were being progressed, predominantly for mental-health teams. beehive.govt.nz

The operating model is bounded:

• The system transcribes or processes a clinical conversation.
• It generates draft notes or related documentation.
• A clinician reviews and approves the output.
• The clinician remains responsible for the clinical record.
• The tool is not authorised to make autonomous diagnostic or treatment decisions.

The Government reported that doctors using the tool saw, on average, one additional patient per shift during the pilot. It also reported that 80% of surveyed Middlemore staff said the tool improved productivity or efficiency, and 84% said it positively affected their experience and wellbeing. These are government- and staff-reported findings, not an independent national evaluation of patient outcomes, documentation errors or long-term workload effects. beehive.govt.nz

The evidence therefore supports a conclusion of national operating access, not yet proven national clinical benefit.

Primary-care adoption is becoming more integrated but remains fragmented

Primary-care organisations are now promoting several AI-scribe and workflow models, including Heidi, Medtech AI and other ambient documentation tools.

Medtech AI is marketed as an integrated layer within Medtech Evolution, combining real-time transcription, structured notes, patient summaries, correspondence and write-back into the practice-management system. Medtech says its outputs require clinician approval and that audio is not retained. These are vendor-reported product and security claims rather than independently audited performance results. medtechglobal.com

The important change is that competition is moving from standalone transcription to workflow integration. A tool that simply produces text may leave clinicians with copying, checking and filing tasks. Products that connect to practice-management systems can reduce that friction, but they also increase the consequences of incorrect data matching, inappropriate write-back or unauthorised access.

National adoption rates remain unavailable. There is no public, comprehensive dataset showing how many New Zealand practices actively use AI scribes, how frequently they are used, which vendors are involved, or how performance differs by specialty, language, rurality or patient group.

Patient-facing AI is moving into supervised service delivery

The Tāmaki Health and Groov pilot is the clearest current example of patient-facing AI in routine primary care. The system supports health coaching between appointments and is designed to direct patients back to human health coaches or clinical teams when additional support is needed. tamakihealth.co.nz

This is a different risk category from documentation. A scribe can produce a flawed draft that a clinician corrects before it enters the record. A patient-facing coach may influence behaviour before a clinician sees the interaction. It therefore requires:

• Clear scope and safety-netting.
• Reliable escalation.
• Monitoring of conversations.
• Appropriate handling of mental-health or acute symptoms.
• Accessible non-AI alternatives.
• Evaluation across language, ethnicity, age and digital-access groups.

Tāmaki Health reports that all AI Health Coach conversations in the pilot were anonymised and clinically reviewed, with no unsafe responses identified to date. That is a useful operating control, but it is still an internal monitoring result over a short period, not evidence of safety at scale. tamakihealth.co.nz

Imaging is now the principal infrastructure-scale pathway

The national radiology procurement gives imaging AI a clearer route to system-wide adoption than existed in the previous edition. The programme is designed to connect image archives, viewers, reporting, enterprise imaging and AI applications across all regions. gets.govt.nz

This may address several barriers identified in earlier work:

• Fragmented regional systems.
• Difficulty moving images between services.
• Duplicate examinations.
• Inconsistent reporting workflows.
• Lack of a common pathway for testing AI tools.
• Weak visibility of model performance across populations and sites.

However, the procurement process is still at the supplier-selection stage. No national AI diagnostic deployment should be inferred from the RFP.

The sector is also developing stronger professional safeguards. RANZCR and the Australian and New Zealand Digital Health Institute warned in August about concerns involving AI tools in Australia’s National Lung Cancer Screening Program, reinforcing the importance of clinical governance and oversight when AI contributes to image analysis or final outputs. Although that programme is Australian, the issue is relevant to New Zealand’s planned imaging-AI architecture. ranzcr.com

Screening illustrates the gap between technical promise and implementation readiness

A 2026 New Zealand Medical Journal viewpoint by Health NZ and University of Auckland authors reviewed the local experience of an AI-integrated diabetic-retinal-screening proof of concept.

The authors found that what appeared to be a straightforward use case was complicated by digital-system limitations, changes to models of care, variable clinician readiness and questions about whether the AI tools were appropriate. They concluded that careful planning, adequate resources and organisational support were necessary for successful implementation. pubmed.ncbi.nlm.nih.gov

This is one of the most useful pieces of evidence about New Zealand’s AI adoption because it describes implementation difficulty rather than only model performance. It suggests that even a clinically attractive screening use case can stall when the surrounding workflow, data, staffing and service design are not ready.

The Ministry of Health continues to describe New Zealand as being in the early stages of assessing how AI can be adopted safely. Its precision-health principles require technologies to be accessible, safe and effective, accountable, equitable and good value. health.govt.nz

Clinical decision support remains research-led

The REVOLUTION trial remains the most substantial New Zealand-linked effort to evaluate AI-guided treatment rather than administrative assistance. The Health Research Council is funding a five-year trial across 50 intensive-care units in New Zealand and Australia, recruiting more than 24,000 patients to test whether machine-learning guidance can improve personalised oxygen therapy for patients on life support. hrc.govt.nz

The trial is important because it is designed to test patient outcomes in a randomised clinical setting. A model can predict treatment response retrospectively without improving care prospectively; clinicians may not follow its recommendations, or the recommendation may create new workflow and safety problems. The trial’s design addresses some of those concerns, but results are years away.

Other research remains earlier-stage. University of Auckland researchers are using MRI scans from hundreds of stroke patients to develop models that may help predict recovery and personalise rehabilitation. The team says future studies will need New Zealand imaging data and further validation before clinical use. auckland.ac.nz

The distinction between these activities matters:

• A funded trial is not a live clinical service.
• A research model is not a validated decision-support tool.
• A vendor demonstration is not evidence of patient benefit.
• A procurement process is not deployment.

Data infrastructure is improving, but the data problem remains material

The Shared Digital Health Record and related primary-care data initiatives continue to move through integration and readiness work. Health NZ has reported engagement with early-adopter practices, testing with practice-management-system vendors, patient communication work and strengthened privacy and cybersecurity arrangements. healthnz.govt.nz

New research illustrates why this foundation matters. University of Auckland researchers linked GP and hospital records for more than 374,000 adults and found that 30.8% had two or more long-term conditions. Of 115,320 people identified with multimorbidity, 62,453 were recorded only in GP data and would not have been identified through hospital data alone. auckland.ac.nz

AI systems trained or evaluated mainly on hospital data may therefore misrepresent the health needs of people whose care is delivered primarily in the community. The implications are particularly important for chronic disease management, population-risk modelling, care coordination and resource planning.

Better connectivity alone will not solve this problem. Data quality, coding consistency, consent, Māori data governance, representation and the ability to evaluate performance across settings remain necessary.

Governance, Policy and Regulation

Professional accountability remains the main clinical safeguard

The Medical Council’s March 2026 guidance requires doctors to use AI only with appropriate oversight, assess whether tools are safe and suitable, protect patient privacy, consider bias and retain responsibility for clinical decisions. mcnz.org.nz

The same practical principle is appearing across other professions: AI may assist practice, but responsibility cannot be delegated to a system or supplier.

For healthcare organisations, this means governance must cover the full operating lifecycle:

• Selection and procurement.
• Clinical validation.
• Patient information and consent.
• Staff training.
• Human review.
• Incident reporting.
• Performance monitoring.
• Model or vendor changes.
• Decommissioning.

An approval at the point of purchase is not enough. AI systems can change through model updates, workflow changes, new data sources or altered vendor terms.

Privacy obligations are becoming more specific

The Office of the Privacy Commissioner states that the Privacy Act applies to the use of AI tools in New Zealand and recommends a Privacy Impact Assessment before handling personal information through AI. The assessment should be updated as the system, data or use case changes. privacy.org.nz

The amended Health Information Privacy Code now includes Rule 3A, requiring health agencies in relevant circumstances to notify people when health information is collected indirectly from a third party. This is relevant to AI systems that combine information from practice-management systems, portals, national datasets or other providers. privacy.org.nz

The Manage My Health compliance notices give these obligations practical weight. Health agencies cannot assume that a contracted supplier’s security controls are sufficient. They must exercise reasonable control over the information they provide to service providers and maintain safeguards against unauthorised use or disclosure. privacy.org.nz

Medical-device regulation remains unsettled

The repeal of the Therapeutic Products Act means that New Zealand does not yet have the comprehensive new framework for software as a medical device that the previous edition implied.

The proposed Medical Products Bill is intended to address innovative medical products and software as a medical device, including AI used for therapeutic purposes. The Ministry has published policy material, but the Bill’s final scope, passage and commencement remain future matters. health.govt.nz

In the interim, organisations deploying clinical AI must work through existing legal and professional frameworks. This can create uncertainty for products that sit between administrative software, clinical decision support and regulated medical devices.

The absence of a single AI-health statute does not mean there are no controls. It means responsibility is distributed across privacy law, professional standards, clinical governance, procurement, consumer rights and any applicable device regulation.

Patient choice and explanation are becoming operational requirements

The publication of RANZCR’s patient resources signals a broader move toward explaining AI use at the point of care. Patients need to know whether AI is:

• Recording or transcribing a consultation.
• Producing a draft clinical note.
• Prioritising an image or referral.
• Reconstructing or enhancing an image.
• Supporting a clinician’s interpretation.
• Making a recommendation.
• Communicating directly with the patient.

The appropriate consent process may differ by use case. A patient may reasonably expect a clinician to use speech-recognition software for documentation, while an AI system that influences diagnosis or treatment requires a more substantive explanation of its role, limitations and human oversight.

Case Studies

Case Study 1: Heidi in emergency departments

Use case: Ambient clinical documentation.

Status: Operating nationally across public emergency departments.

Reported benefits: Reduced documentation time, possible additional patient capacity and improved staff experience.

Evidence status: Government and provider-reported pilot and rollout findings. No public independent national evaluation of patient outcomes, error rates or long-term workforce effects.

Implementation lesson: National scale is achievable when the use case is bounded, the clinician remains the editor and the value proposition is directly linked to administrative workload. beehive.govt.nz

Case Study 2: Health NZ Radiology Digital Modernisation Programme

Use case: National imaging integration, archiving, reporting and orchestration of imaging-AI applications.

Status: Live RFP opened on 25 September 2026; supplier selection and implementation are still ahead.

Potential value: More consistent imaging workflows, improved access to images, less duplication and a platform for evaluating multiple AI applications.

Evidence status: Procurement intent, not operating deployment.

Implementation lesson: The health system is treating AI as dependent on architecture, interoperability and governance rather than as a standalone software purchase. gets.govt.nz

Case Study 3: Tāmaki Health and Groov AI Health Coach

Use case: Patient engagement and chronic-disease coaching between appointments.

Status: Early rollout across selected primary-care clinics.

Reported benefits: Sevenfold higher engagement than a previous human-only outreach benchmark, increased patient-reported information and additional between-appointment touchpoints.

Evidence status: Provider- and vendor-reported early results from a small, uncontrolled pilot.

Implementation lesson: Patient-facing AI may extend human services, but engagement results must be followed by evidence of safety, sustained behaviour change, clinical outcomes and equity. tamakihealth.co.nz

Case Study 4: RosterLab and HealthX

Use case: Workforce rostering for medical radiation technologists.

Status: Initial HealthX pilot across four Health NZ services.

Reported benefits: Earlier company work indicated a major reduction in roster-administration time.

Evidence status: The current pilot has not yet produced publicly reported outcome data.

Implementation lesson: Workforce optimisation is an important adoption pathway because it can improve service capacity and staff experience without placing an algorithm directly in the clinical decision chain. auckland.ac.nz

Case Study 5: AI-integrated diabetic-retinal screening

Use case: AI-supported screening for diabetic retinopathy, with a focus on improving access for Pacific peoples.

Status: New Zealand proof of concept; not established as a national operating service.

Key finding: Implementation was constrained by digital systems, model-of-care changes, clinician readiness and questions about tool suitability.

Implementation lesson: The difficult part of clinical AI is often not model accuracy. It is redesigning the service around the model while preserving safety, equity and accountability. pubmed.ncbi.nlm.nih.gov

Case Study 6: AI-assisted stroke-recovery research

Use case: MRI-based prediction of stroke recovery and rehabilitation needs.

Status: Research and model development.

Potential value: Earlier prediction of recovery trajectories and more personalised rehabilitation planning.

Evidence status: Further validation and New Zealand data are required before clinical use.

Implementation lesson: New Zealand research capability is advancing, but translation into clinical tools will depend on external validation, representative data and prospective evaluation. auckland.ac.nz

Trends

1. Adoption is widening horizontally before it deepens clinically

AI is appearing in more parts of healthcare—documentation, coaching, rostering, imaging infrastructure and research—but most use cases remain shallow in clinical authority.

The sector is adding more tools without yet transferring major diagnostic or treatment decisions to them. This reflects a rational risk gradient: administrative and operational applications are easier to supervise than systems that alter clinical care directly.

2. Workflow integration is becoming more important than model novelty

The radiology RFP, integrated primary-care scribes and Shared Digital Health Record work all point to the same conclusion: AI value depends on where information enters and leaves the workflow.

A highly capable model that requires manual copying, fragmented login processes or incomplete data may generate little net benefit. Conversely, a less ambitious model embedded into a trusted system may produce measurable value.

3. Patient-facing AI is the next major governance test

The Tāmaki Health pilot shows why patient-facing applications are attractive: they may provide support between appointments and reach people who are not engaging with standard services.

They also create more complex safety questions. Organisations must manage inappropriate reassurance, missed escalation, cultural and linguistic fit, vulnerable users, digital exclusion and the possibility that patients treat the tool as a substitute for clinical care.

The evidence base is currently thinner than the level of commercial interest.

4. Equity is becoming an implementation variable

The sector is increasingly moving beyond general statements that AI should be equitable.

The diabetic-retinal-screening experience shows that equity goals can be undermined by workflow and system barriers. The multimorbidity research shows that incomplete primary-care data can distort population need. The Tāmaki Health pilot suggests that targeted outreach can reach Māori and Pacific patients, but that result requires independent and longer-term assessment.

Important questions remain unresolved:

• Does performance differ across Māori, Pacific and other populations?
• Does the system work with te reo Māori and Pacific languages?
• Are older people and disabled people able to use it?
• Does digital access determine who benefits?
• Can rural services operate it reliably?
• Who controls and benefits from data generated through the system?

5. Security incidents will influence procurement more than abstract AI principles

The Manage My Health compliance notices provide a concrete example of the consequences of weak supplier controls. Healthcare organisations will increasingly assess AI vendors through the same lens as other sensitive-data providers, but with additional questions about model training, retention, subcontractors and data flows.

Security is not a separate technical workstream. It is part of clinical safety when a system handles consultation audio, patient histories, diagnostic images or treatment recommendations.

6. National infrastructure is progressing, but implementation capacity is the constraint

The radiology RFP demonstrates stronger national coordination. The Shared Digital Health Record work demonstrates progress on data-sharing infrastructure. Health NZ’s AI and digital guidance provides a basis for more consistent evaluation.

Yet infrastructure procurement does not automatically create implementation capacity. Health services still need clinical informatics expertise, data engineering, cybersecurity, change management, procurement capability and time for frontline staff to test and adapt systems.

7. The evidence gap is now more visible

There is more public information about AI deployments than a year ago, but much of it remains self-reported.

The evidence categories should continue to be separated:

• Product availability.
• Announced partnership.
• Proof of concept.
• Controlled pilot.
• Routine operating use.
• Scaled deployment.
• Independent evidence of patient or system benefit.

New Zealand has examples in most categories, but relatively few in the final two.

Outlook

Over the next 12 to 18 months, the most plausible developments are:

• Supplier selection and planning for the national radiology digital platform.
• Further AI-scribe adoption in primary care, mental health, aged care and allied health.
• Expansion of patient-facing coaching models beyond initial clinics.
• Evaluation of HealthX workforce-operations pilots.
• Continued integration of primary-care data into national services.
• More patient-information material from professional colleges and providers.
• Broader use of Health NZ’s AI evaluation and governance frameworks.
• Stronger contractual and cybersecurity requirements for health technology suppliers.
• Further policy development for software as a medical device under the Medical Products Bill.
• Progression of AI-guided treatment trials, including REVOLUTION.

The least likely near-term outcome remains widespread autonomous diagnosis or treatment. Current deployments and guidance favour clinician-supervised augmentation, and the most ambitious clinical applications are still being studied prospectively.

The main strategic risk is not that New Zealand will fail to access AI technology. It is that organisations will deploy disconnected tools faster than they can establish shared standards, reliable data, security controls and meaningful evaluation.

Overall Assessment

At 1 October 2026, AI adoption in New Zealand healthcare is becoming more operational, but it is not yet clinically transformative.

The sector now has:

• A nationally operating AI documentation service in emergency departments.
• Expanding primary-care access to AI scribes and workflow tools.
• An early patient-facing AI coaching deployment with reported engagement gains.
• A national radiology procurement that explicitly includes an imaging-AI orchestrator.
• A Health NZ workforce-rostering pilot.
• Research into AI-guided critical-care treatment and stroke recovery.
• A documented implementation case showing why diabetic-retinal-screening AI is difficult to scale.
• More explicit patient communication about AI.
• Stronger privacy enforcement and clearer obligations for indirect data collection.
• A developing, but not yet complete, regulatory pathway for software as a medical device.

The previous edition’s central assessment remains sound, but the emphasis has shifted. The question is no longer simply whether AI can be introduced into healthcare. It is whether New Zealand can integrate it into services that are secure, interoperable, culturally appropriate, clinically accountable and demonstrably beneficial.

The strongest evidence is still for AI that assists people with documentation, coordination and workflow. The evidence is thinner for AI that changes patient behaviour, prioritises care or influences clinical decisions.

The sector should therefore measure progress by:

• Reliable operating use rather than announcements.
• Independent evaluation rather than vendor-reported satisfaction.
• Patient outcomes rather than administrative activity alone.
• Equity across populations and settings.
• Stronger security and supplier accountability.
• The ability to monitor, correct and stop systems when they fail.

New Zealand’s healthcare AI trajectory is best described as supervised expansion under infrastructural and governance constraints. The next phase will be determined less by access to better models than by whether the health system can build the conditions necessary to trust them.