Digital Health AI Summit 2025 - Table Summaries

- 1 -

IN THIS REPORT 3 Foreword 4 AI in Smarter Patient Flow 5 Upskilling the Workforce in AI 7 AI in Clinical Decision Making 8

AI Scribing Tools – Safe and Effective Use

AI – Alleviating Administrative Burden for Clinicians

AI – Leveraging Data for Impact

AI at Work – Supercharging Productivity

AI in Telehealth

AI Driven Mental Health Support Tools

AI-Powered Efficiencies in Aged Care

AI in Health Prevention

AI Evaluation – Measuring Success to Inform Decision Making

AI Agents as Trusted Digital Labour

Developing Partnerships and Collaborations to Advance AI

AI-Powered Operational Efficiency

Mapping and Navigating the Health AI Ecosystem

21 AI-Powered Interoperability, Driving Efficiency Optimisation, and Preventing Readmission 22 AI in Health Equity and Access 23 AI Scribes are Here – What AI Tools Are Next for Primary Care? 24 AI – Optimising Business Processes

- 2 -

FROM THE CEO

Kia ora,

Scott Arrol HiNZ CEO

The Digital Health AI Summit in 2025 brought together health leaders from across the sector to collaborate to make AI driven change in healthcare. Held in Wellington on March 20-21, attendees had the opportunity to contribute and learn at 20 ‘action stations’ (tables) each with a different discussion topic. The outcomes of those discussions – totalling more than 140 hours over two days – are summarised in this report and have been created with the assistance of our AI tool sponsor, Contented AI.

They are also published under each table topic on eHealthForum.nz.

We encourage you to visit the forum to view the detailed summaries and continue the conversation by posting a topic or replying to a post. Summit attendees have told us that the connections made across sector were one of the best features of the summit, but now it’s up to you to stay connected! Please also share this report with anyone who wants to join the conversation – these topic discussions are not limited to just those who attended the summit.

Ngā mihi,

Scott

- 3 -

AI IN SMARTER PATIENT FLOW

Discussion Points ▪ Patient flow encompasses movement through healthcare systems from pre-assessment to discharge, including information flow ▪ Current challenges include fragmented systems, lack of data standardisation, and limited interoperability between public/private sectors ▪ Hegemonic application of healthcare systems often fails to incorporate Treaty of Waitangi principles and cultural needs ▪ Data sharing barriers exist between regions (e.g., South Island vs North Island patient information) ▪ Opportunities exist for AI to assist with referral management, discharge planning, bed management, and resource allocation ▪ Potential for AI to help patients navigate complex healthcare systems through "AI navigators" or agents Key Actions ▪ Explore AI applications for triage and referral management to direct patients to appropriate services ▪ Consider AI tools for predicting patient acuity and resource needs to improve bed management ▪ Investigate ambient AI scribes to reduce documentation burden on clinicians ▪ Look into AI-supported discharge planning to speed up patient flow and reduce bottlenecks ▪ Develop standardised definitions for data collection to enable effective AI implementation Need for balance between AI assistance and human clinical judgment ▪ Importance of co-design with patients and communities to ensure culturally appropriate solutions ▪ Consideration of privacy, security and consent frameworks for data sharing ▪ Workforce development and training required for effective AI implementation ▪ Potential for ACC framework to address liability concerns with AI clinical decision support ▪ Opportunity for New Zealand to lead in healthcare AI due to unique NHI system and ACC model Additional Notes ▪

Table sponsored by

Supported by St Georges Hospital

- 4 -

UPSKILLING THE WORKFORCE IN AI

Discussion Points ▪

Key capabilities needed for healthcare workforce to work effectively with AI: • Critical thinking and ability to analyse AI outputs • Trust in data and governance frameworks • Leadership at policy and governance levels • Innate curiosity and ability to think critically • AI literacy and familiarity with available tools • Equitable access to AI tools relevant to New Zealand context ▪ Distinction between general AI and generative AI: • Recognition that AI is broader than just generative AI • Acknowledgement that AI has been used in healthcare for decades • Need for workforce to understand different types of AI applications ▪ Skills that will become more important: • Decision-making skills • Critical thinking and problem solving • Communication skills for explaining AI use to patients • Ability to validate AI outputs and identify hallucinations ▪ Skills that may become less relevant: • Operational skills involving manual repetitive tasks • Relying solely on memory knowledge • Administrative documentation that can be automated ▪ Future roles and workforce planning: • Need for AI governance specialists • Data scientists and AI curators who can validate outputs • Potential for AI to handle triage and routine patient education • Concern about maintaining human connection in healthcare delivery ▪ Leadership requirements: • CEOs and leaders should demonstrate AI literacy • Leaders should use AI tools themselves to understand capabilities • Need for clear governance frameworks and guardrails • Importance of creating safe environments for experimentation Develop frameworks for AI governance in healthcare settings ▪ Create training programmes that build AI literacy across all workforce levels ▪ Establish guardrails for responsible AI use in clinical settings ▪ Consider how to maintain human connection while implementing AI tools ▪ Develop strategies for upskilling current workforce rather than complete replacement ▪ Address infrastructure gaps that prevent effective AI implementation

Key Actions ▪

- 5 -

Additional Notes ▪

Tension exists between efficiency gains and maintaining quality patient care ▪ Concerns raised about whether AI will actually save time or just shift workload ▪ Cultural considerations for AI implementation in New Zealand context ▪ Need to consider both regulated and non-regulated healthcare workers ▪ Patients increasingly using AI tools (like ChatGPT) for self-diagnosis before seeing clinicians ▪ Importance of addressing digital literacy gaps across different generations of healthcare workers ▪ Recognition that AI implementation requires fundamental changes to healthcare delivery models

Table sponsored by

- 6 -

AI IN CLINICAL DECISION MAKING

Discussion Points ▪ Distinction between AI for clinical decision support versus AI making clinical decisions ▪ Clinicians should maintain final accountability for decisions, with AI serving as a tool ▪ Concerns about potential deskilling of healthcare workforce if over-reliant on AI ▪ Data quality, privacy, and sovereignty issues when training AI models ▪ Need for proper guardrails and frameworks for AI use in healthcare ▪ Challenges with explaining "black box" AI decision-making processes ▪ Varying levels of trust in AI across different clinical applications ▪ Potential for AI to both address and worsen health inequities Key Actions ▪ Develop education programmes on effective AI prompt creation and critical assessment ▪ Create clear policies on AI use that maintain clinician accountability ▪ Establish frameworks for measuring AI impact on clinical outcomes ▪ Consider sustainable funding models for AI implementation in healthcare ▪ Explore options for New Zealand-specific data training while maintaining privacy ▪ Develop strategies to improve health literacy alongside AI literacy Additional Notes ▪ Different types of AI have different applications (rule-based algorithms vs large language models) ▪ Patients are already using AI tools for health information regardless of clinical guidance ▪ Cultural considerations around data use differ between individual and community-focused societies ▪ Economic sustainability of AI in healthcare remains a significant challenge ▪ Existing regulatory frameworks may be adaptable rather than creating entirely new ones ▪ Offline AI systems may offer better privacy protection than cloud-based options

Table sponsored by

Supported by NAHSTIG

- 7 -

AI SCRIBING TOOLS – SAFE AND EFFECTIVE USE

Discussion Points ▪ AI scribing tools reduce administrative burden, allowing clinicians to maintain eye contact and be more present with patients ▪ Benefits include improved patient experience, reduced cognitive load, and potential for better patient education ▪ Concerns raised about privacy, data security, informed consent processes, and integration with existing systems ▪ Challenges with implementation include training needs for clinicians, especially those less comfortable with technology ▪ Potential for AI tools to improve equity by supporting multilingual consultations and helping patients with communication barriers Key Actions ▪ Develop clear, accessible informed consent processes that can be implemented at reception rather than during consultations ▪ Consider multiple education approaches for both patients and clinicians based on different learning styles ▪ Ensure proper review of AI-generated notes before finalising in patient management systems ▪ Address integration issues with practice management systems like MedTech and Indici ▪ Establish protocols for pausing recording when sensitive information is shared Additional Notes ▪ Current adoption shows "first adopter" clinicians leading implementation with others following ▪ Privacy impact assessments are crucial before widespread implementation ▪ Potential for AI tools to support whānau-based consultations where multiple family members are present ▪ Need for standardised language and coding to ensure data quality and consistency ▪ Consideration of how AI tools might specifically benefit patients with communication barriers or disabilities

Table sponsored by

- 8 -

AI – ALLEVIATING ADMINISTRATIVE BURDEN FOR CLINICIANS

Discussion Points ▪ Explored how AI could assist clinicians by identifying symptoms during patient encounters, particularly for mental health conditions ▪ Discussed administrative burden reduction through AI tools like scribes, automated coding, and document generation ▪ Considered data integration challenges across healthcare systems and the need for standardised approaches ▪ Examined privacy concerns and data sovereignty issues with AI tools ▪ Reviewed potential AI applications including discharge summary generation, patient record summarisation, and clinical coding Key Actions ▪ Investigate AI tools that can identify residual symptoms in major depressive disorder ▪ Explore opportunities for AI to assist with referral processes and triage ▪ Consider developing guidelines for safe AI implementation in clinical settings ▪ Evaluate AI scribes like Heidi and Tuhi for wider implementation ▪ Address data integration challenges before implementing advanced AI solutions Additional Notes ▪ Participants noted that clinicians spend up to 50% of their time on administrative tasks ▪ Concerns raised about automation bias and the need for human oversight of AI outputs ▪ Discussion of cultural considerations for AI implementation in Aotearoa, including Te Reo Māori integration ▪ Noted the importance of education and training for effective AI adoption ▪ Highlighted the need for balance between innovation and regulation in healthcare AI

Table sponsored by

- 9 -

AI – LEVERAGING DATA FOR IMPACT

Discussion Points ▪ Data quality and consistency challenges across healthcare systems in New Zealand ▪ Difficulties with data silos and integration between 42 different data warehouses nationwide ▪ Concerns about data privacy, sovereignty and sharing between organisations ▪ Bias in healthcare data, particularly regarding gender and ethnicity ▪ Challenges with standardisation of clinical coding and terminology ▪ Importance of localised data for AI models rather than importing overseas models ▪ Considerations around building, buying, or adapting AI solutions ▪ Workforce readiness and training needs for AI implementation ▪ Funding constraints for AI initiatives in healthcare Key Actions ▪ Consider starting small with AI implementations and iterating based on results ▪ Focus on explainable AI to build trust with clinicians and patients ▪ Investigate natural language processing for unstructured clinical notes ▪ Explore data disaggregation to address bias in AI models ▪ Develop governance frameworks for AI use that balance innovation with privacy ▪ Consider adapting existing AI models with local data rather than building from scratch ▪ Investigate standardisation of ethnicity data collection (minimum Level 2) ▪ Look into creating "sandbox" environments for researchers to access de-identified data Additional Notes ▪ Dedalus highlighted their AI product (Clinalytics Medical AI) which is embedded in workflow ▪ Several organisations are at early stages of AI implementation ("phase zero") ▪ Breast Cancer Foundation NZ is working on implementing AI for pathology report reading ▪ Concerns raised about the cost of AI implementation versus available funding ▪ Discussion about the need for continuous model retraining and validation ▪ Emphasis on AI as augmentation rather than replacement for clinical decision-making ▪ Noted that AI models need to be registered as medical devices in many jurisdictions

Table sponsored by

- 10 -

AI AT WORK – SUPERCHARGING PRODUCTIVITY

Discussion Points ▪ New Zealand ranks low globally in AI enthusiasm and confidence, particularly in healthcare and public sectors ▪ Participants discussed using basic off-the-shelf AI tools to improve productivity in healthcare settings ▪ Several use cases were explored including clinical documentation, patient communication, language translation, data analysis, and administrative tasks ▪ Participants mapped various AI applications on an impact vs feasibility matrix ▪ Most participants reported using AI tools daily in their work Explore using AI for summarising patient notes to improve clinical efficiency ▪ Consider implementing AI tools for translating languages to better serve diverse patient populations ▪ Investigate AI applications for improving rostering and resource allocation ▪ Look into AI-powered tools for clinical documentation to reduce administrative burden ▪ Develop strategies to address the digital divide between healthcare organisations Key Actions ▪ Additional Notes ▪ Major blockers identified: Health NZ policy restrictions, funding limitations, paper-based systems, and risk aversion ▪ Key enablers identified: leadership support, knowledge sharing, cross-sector collaboration, and allowing time for staff to experiment with tools ▪ Participants noted the inequity in digital maturity across the health system ▪ Private sector organisations appear to have more flexibility to implement AI solutions ▪ Concerns raised about data privacy, clinical risk, and the need for appropriate guardrails

Table sponsored by

- 11 -

AI IN TELEHEALTH

Discussion Points ▪ Workshop used nominal group technique to explore the future of telehealth with AI in 10 years ▪ Participants envisioned telehealth services utilising AI from various perspectives ▪ Multiple case examples were used to stimulate thinking about telehealth's broad definition ▪ Telehealth defined as healthcare using digital technology where provider and recipient are separated by time and/or distance

Key Actions Participants identified steps needed to achieve their envisioned future, including: ▪ Developing shared electronic health records with individual ownership ▪ Creating standards for data interoperability ▪ Addressing the digital divide and ensuring equitable access ▪ Establishing clear regulatory frameworks for AI in healthcare ▪ Implementing proper funding models for telehealth innovation ▪ Building trust through transparency and patient control of data

Additional Notes ▪ Common themes across groups included data standards, patient-centred care, and regulatory concerns ▪ Participants expressed mixed views on regulation—some advocated for new regulatory bodies while others preferred more agile approaches to existing frameworks ▪ Many groups prioritised human factors over technological solutions ▪ Concerns raised about AI bias, privacy, and the need for healthcare workforce support ▪ International data sharing emerged as an important consideration for future telehealth systems

Table sponsored by

- 12 -

AI DRIVEN MENTAL HEALTH SUPPORT TOOLS

Discussion Points ▪ Current AI adoption varies widely across organisations, from daily use to minimal engagement ▪ Mental health practitioners show more resistance to AI tools compared to other healthcare sectors ▪ Privacy and data security concerns are heightened for mental health information ▪ Participants discussed the need for clear governance frameworks and ethical guidelines ▪ The Minister of Mental Health is actively seeking digital solutions to address service gaps ▪ Existing AI applications include ambient scribes, chatbots, and administrative tools ▪ Cultural considerations and equity issues must be addressed in AI development Explore 90-day trial approaches to test AI tools in controlled environments ▪ Develop clearer guidelines for informed consent when using AI in clinical settings ▪ Investigate successful international models that could be adapted for New Zealand ▪ Consider establishing AI committees within healthcare organisations ▪ Improve interoperability between different health information systems ▪ Identify low-risk use cases to build confidence and demonstrate value Key Actions ▪ Additional Notes ▪ Many young people already use AI tools like ChatGPT for mental health support ▪ Digital tools could help bridge gaps while people wait for in-person services ▪ AI could assist with early intervention and identifying patterns that humans might miss ▪ Wearable technology integration with AI shows promise for monitoring mental wellbeing ▪ Voice analysis technology could help detect depression and other conditions ▪ The tension between efficiency and therapeutic relationships needs careful consideration ▪ Māori data sovereignty principles must be incorporated into AI governance

Table sponsored by

- 13 -

AI POWERED EFFICIENCIES IN AGED CARE

Discussion Points ▪ Current state of aged care technology in New Zealand varies widely - from pen and paper systems to sophisticated digital platforms ▪ Approximately 40,000 people in residential aged care facilities and another 40,000 receiving home care support ▪ Clinical safety opportunities through AI include medication management, falls prevention, and pain assessment ▪ Workforce challenges include staff shortages and the need for better training on new technologies ▪ Data interoperability issues between healthcare systems (hospitals, GPs, aged care facilities) ▪ Privacy concerns balanced against clinical safety benefits ▪ Potential for AI to support ageing in place and reduce loneliness Explore AI scribe tools like Heidi to reduce administrative burden for nurses ▪ Investigate facial recognition technology for pain assessment (Pain Check) ▪ Consider voice recognition software for detecting distress in non-verbal residents ▪ Look into developing digital navigator roles to support technology adoption ▪ Engage with Te Patu Aora (Health NZ) about improved data sharing between systems ▪ Develop guidelines for AI implementation that balance privacy with clinical benefits Key Actions ▪ Aged care is predominantly nurse-led with doctors visiting periodically ▪ InterRAI assessment tool is mandatory, but most providers use duplicate systems ▪ Funding challenges impact technology adoption - capital investment not covered by government funding ▪ Cultural considerations important for technology acceptance ▪ Potential for public-private partnerships to drive innovation ▪ Need for better integration between retirement villages and higher-care facilities ▪ Growing demand with ageing population will require technological solutions Additional Notes ▪

Table sponsored by

- 14 -

AI IN HEALTH PREVENTION

Discussion Points ▪ Explored the role of AI in health prevention at both individual and population levels ▪ Discussed the importance of personalisation in health prevention approaches ▪ Examined how AI could help identify at-risk populations and target interventions ▪ Considered the challenges of data quality, privacy, and interoperability ▪ Debated the effectiveness of wearable technology (like Oura rings) for health monitoring ▪ Explored how AI could help identify personal motivators for behaviour change ▪ Discussed the potential for AI to analyse social media data to understand health behaviours Investigate how AI can meet people where they are to tailor health messages ▪ Consider how to make preventative health tools accessible to all populations ▪ Explore ways to use AI to identify what motivates individuals to change behaviour ▪ Look into how AI can help identify high-risk individuals earlier for intervention ▪ Research the return on investment for preventative health interventions Key Actions ▪ Participants shared personal experiences with health monitoring technology ▪ Discussed the challenge of engaging young people (17-24) who prefer in-person interactions ▪ Noted the equity challenges in health technology access ▪ Highlighted the importance of trust and transparency in health data collection ▪ Acknowledged the tension between commercial interests and public health goals ▪ Discussed the Saudi Arabian and Finnish approaches to preventative health Additional Notes ▪

Table sponsored by

- 15 -

AI EVALUATION – MEASURING SUCCESS TO INFORM DECISION MAKING

Discussion Points ▪ Evaluation frameworks for AI in healthcare need to be tailored for Aotearoa New Zealand context ▪ Current AI implementation timeline (months) differs significantly from traditional health innovations (15-17 years) ▪ Key measurement categories identified: • Clinical measures (productivity, staff satisfaction) • Patient-related measures (access, engagement, satisfaction) • Administrative measures • System-level measures (integration with workflows) • AI-specific measures (explainability, reliability) ▪ Equity considerations must be incorporated into evaluation frameworks ▪ Explainability of AI remains a contested area with varying perspectives on its importance ▪ Cost vs value calculations need to be distinguished Key Actions ▪ Develop a pragmatic evaluation approach suitable for the rapid implementation timeline of AI ▪ Establish baseline measurements for comparison where possible ▪ Consider adopting elements from international frameworks (e.g., UK's Digital Technology Assessment Criteria) with NZ-specific modifications ▪ Incorporate ongoing monitoring into implementation plans with clear responsibility assignment ▪ Define success measures specific to each AI use case and problem being addressed Additional Notes ▪ Current regulatory framework in NZ lacks specific provisions for AI as medical devices ▪ Risk tolerance may vary between public and private organisations ▪ Benefits realisation should be clearly defined before implementation ▪ Evaluation should consider the entire model of care, not just the AI tool itself ▪ Stakeholder involvement (clinicians, patients, system leaders) is essential in defining success measures

Table sponsored by

- 16 -

AI AGENTS AS TRUSTED DIGITAL LABOUR

Discussion Points ▪ AI agents differ from large language models by having access to internal data, ability to execute workflows, and potential to act autonomously ▪ Trust is a significant concern when deploying AI agents in healthcare settings ▪ Debate around whether AI should augment human capabilities rather than replace humans entirely ▪ Data quality, governance and sovereignty are critical challenges for effective AI implementation ▪ Importance of defining clear scope and purpose for AI agents to build trust and manage expectations ▪ Need for human oversight in clinical decision-making contexts ▪ Potential for AI to help with administrative tasks, contact centres, and back-office functions Key Actions ▪ Start with small, low-risk use cases to build confidence before expanding to more complex applications ▪ Implement proper governance frameworks for both structured and unstructured data ▪ Ensure transparency in how AI agents make decisions and what data they access ▪ Consider cultural appropriateness when implementing AI in New Zealand context, particularly for Māori and Pacific communities ▪ Establish clear monitoring and audit capabilities for AI agent activities Additional Notes ▪ Healthcare organisations face resource constraints that AI could potentially address ▪ Significant variation in AI readiness across different healthcare organisations ▪ Current regulatory frameworks may not be well-suited for AI agent implementation ▪ Potential for agents to work together in complementary roles, similar to multidisciplinary teams ▪ Need to balance innovation with appropriate caution in clinical settings ▪ Importance of considering Te Ao Māori perspectives in AI implementation

Table sponsored by

- 17 -

DEVELOPING PARTNERSHIPS AND COLLABORATIONS TO ADVANCE AI

Discussion Points ▪ Barriers to partnerships and collaborations in AI health research, including data sovereignty, governance, and access issues ▪ Challenges in transitioning from research to implementation, with many projects stalling at prototype stage ▪ Need for synthetic data to enable innovation while preserving privacy ▪ Importance of building trust with patients and communities regarding data use ▪ Funding gaps between research and commercialisation phases ▪ Challenges of siloed systems and lack of interoperability across health sector ▪ Value of person-centred approaches when developing AI solutions ▪ Potential for AI to reduce administrative burden and improve efficiency Key Actions ▪ Explore creation of a project registry/library to track AI initiatives and prevent duplication ▪ Consider developing a framework for evaluating AI tools and solutions ▪ Investigate synthetic data approaches to enable innovation while protecting privacy ▪ Establish clearer pathways from research to implementation with defined stage gates ▪ Develop better methods for quantifying the value of AI implementations to justify investment ▪ Engage clinicians and end-users earlier in the development process Additional Notes ▪ Success stories shared included Microsoft Copilot implementation at Ministry of Health ▪ Breast Cancer Foundation's approach of bringing researchers in-house rather than sharing data externally ▪ Discussion of international examples and opportunities for cross-border collaboration ▪ Recognition that AI tools should augment rather than replace clinical decision-making ▪ Acknowledgement that cultural considerations and Te Ao Māori perspectives must be integrated into AI development

Table sponsored by

- 18 -

AI-POWERED OPERATIONAL EFFICIENCY

Discussion Points ▪ The meeting focused on eight core domains for AI implementation in healthcare, with emphasis on operational efficiency ▪ Participants shared experiences from various healthcare organisations including ACC, Stroke Foundation, Tōpūtanga Tapuhi Kaitiaki o Aotearoa (NZ College of Midwives), and Ministry of Health ▪ Key metaphor discussed: healthcare organisations are transitioning from "steamships to spaceships" - requiring fundamental operational changes rather than incremental improvements ▪ Data sovereignty was highlighted as critical, particularly for Māori data and ensuring cultural integrity ▪ Participants noted the importance of starting with clear problem statements rather than technology-first approaches ▪ Clinical transcription tools were discussed as a practical example of AI implementation, with time savings noted but also potential unintended consequences Key Actions ▪ Organisations should develop clear AI implementation frameworks covering problem definition through to scaling ▪ Establish proper data governance and privacy frameworks before implementing AI solutions ▪ Involve clinicians and end-users from the outset of AI implementation projects ▪ Consider piloting small, low-risk AI implementations before scaling to larger applications ▪ Develop strategies to bring both early adopters and technology-hesitant staff on implementation journeys ▪ Evaluate AI tools based on their ability to enhance rather than replace human capabilities Additional Notes ▪ ACC has implemented auto-summarisation tools that save frontline staff significant time ▪ Stroke Foundation is exploring AI to support prevention efforts and improve acute care diagnostics ▪ Ministry of Health is trialling Microsoft Copilot with positive results for breaking down information silos ▪ Concerns raised about New Zealand's health IT infrastructure readiness for AI implementation ▪ Participants discussed the balance between building local solutions versus adopting international platforms ▪ The potential for AI to reduce clinician burnout by handling administrative tasks was highlighted

Table sponsored by

- 19 -

MAPPING AND NAVIGATING THE HEALTH AI ECOSYSTEM

Discussion Points ▪ The AI Forum is a non-profit organisation founded in 2017 to connect AI innovators, end users, investors, regulators, researchers, educators and entrepreneurs ▪ The Health AI Working Group was established following the AI Forum's blueprint for AI in Aotearoa, which identified healthcare as a key sector for AI transformation ▪ Participants discussed barriers to AI adoption in healthcare including regulatory uncertainty, lack of clear governance frameworks, and challenges with data access ▪ Radiology was highlighted as a sector with established AI implementation that could provide learnings for other healthcare areas ▪ Concerns were raised about the sale of MedTech (used by 85% of GPs) to a Canadian company and implications for patient data ▪ Discussion of the need for risk-proportionate approaches to AI regulation, with different standards for clinical decision support versus productivity tools ▪ Participants noted the gap in medical device regulation in New Zealand, with no pre-market authorisation or post-market surveillance requirements Key Actions ▪ Attendees encouraged to sign up to the AI Forum register to build connections across the health AI ecosystem ▪ Proposal to develop interim guidance for healthcare-specific AI implementation while waiting for formal regulation ▪ Suggestion to create a framework for quality checks of AI tools (e.g., quarterly reviews of clinical scribes) ▪ Recommendation to establish a mechanism for sharing experiences with specific AI products and vendors ▪ Proposal to develop shared training resources for AI literacy across healthcare organisations ▪ Suggestion to create a classification framework that helps identify risk levels for different AI use cases Additional Notes ▪ Current regulatory timeline suggests medical products bill won't be in place until 2028 ▪ Distinction made between software as medical device (requiring higher regulation) versus productivity tools ▪ Participants discussed the challenge of balancing innovation with appropriate guardrails ▪ Concerns raised about automation bias, where clinicians may stop checking AI outputs after becoming comfortable with tools ▪ Discussion of data sovereignty issues and challenges with New Zealand's limited computing infrastructure ▪ Participants noted the need for better education about different types of AI and their appropriate applications

Table sponsored by

- 20 -

AI-POWERED INTEROPERABILITY, DRIVING EFFICIENCY OPTIMISATION, AND PREVENTING READMISSION

Summary ▪ Interoperability challenges in New Zealand healthcare systems, with fragmented data across primary care, hospitals, and community services ▪ Barriers to AI adoption including infrastructure limitations (poor WiFi, outdated systems), privacy concerns, and lack of standardised data formats ▪ Potential AI applications discussed: • Administrative task automation to free clinical staff for patient care • Discharge summary improvement and follow-up alerts • Patient recall systems and chronic disease management • Early intervention triggers based on known clinical pathways • Ambient AI for background monitoring without disrupting workflows Key Actions ▪ Need for centralised guidance on AI implementation while allowing bottom-up innovation ▪ Develop clear consent frameworks that balance patient privacy with clinical utility ▪ Improve vendor engagement through deeper workflow understanding before implementation ▪ Create multi-modal training approaches when introducing new technologies ▪ Consider a "constitution" of agreed standards for healthcare data and AI use Additional Notes ▪ Cultural resistance to standardisation identified as a significant barrier to healthcare integration in New Zealand ▪ Staff adoption requires transparent communication about implementation challenges ▪ Patient data sovereignty concerns particularly important for Māori and Pacific populations ▪ Successful AI implementation requires balancing immediate time-saving benefits with realistic expectations about transition periods ▪ Leadership support critical for creating organisational confidence in new technologies

Table sponsored by

- 21 -

AI IN HEALTH EQUITY AND ACCESS

Discussion Points ▪ Scope of discussion expanded beyond disability to include mental health, rainbow, migrant, Māori, Pasifika, rural and other vulnerable groups ▪ Digital poverty remains a significant barrier to equitable AI access - many communities lack basic hardware, internet connectivity, and digital literacy ▪ Data quality concerns - marginalised groups often have poor representation in datasets used to train AI systems ▪ Trust issues between vulnerable communities and health systems may extend to AI tools ▪ Privacy and data sovereignty concerns, particularly for Māori and vulnerable populations ▪ Potential for AI to reduce administrative burden on healthcare workers, allowing more time for patient care ▪ Need for New Zealand-specific AI solutions that reflect Te Tiriti o Waitangi principles and local context ▪ Concerns about AI being seen as a replacement for human services rather than augmentation Key Actions ▪ Focus on improving basic digital infrastructure and literacy before implementing advanced AI solutions ▪ Develop clear guidelines for AI use in healthcare settings, including transparency about when AI is being used ▪ Ensure AI tools are designed with accessibility features from the beginning, not as afterthoughts ▪ Involve people with lived experience in AI development from inception, with fair compensation for their expertise ▪ Create mechanisms for AI to help reduce repetitive trauma storytelling for patients ▪ Establish New Zealand-specific data repositories to improve AI training on local populations ▪ Develop education programmes about AI risks and benefits for both healthcare workers and patients Additional Notes ▪ Comparison made between experimental medicines (which carry clear warnings) and AI tools (which often don't) ▪ Discussion of urban narcissism/geographical narcissism affecting rural healthcare access ▪ Potential for AI to help with navigation of complex health systems, particularly for those with communication difficulties ▪ Opportunity for AI to improve continuity of care through better record-keeping ▪ Need to measure meaningful outcomes that matter to patients (relationships, quality of life) rather than just system metrics ▪ Concerns about commercial interests creating siloed systems that don't communicate with each other ▪ Recognition that younger generations have different perspectives on privacy and technology use

Table sponsored by

- 22 -

AI SCRIBES ARE HERE – WHAT TOOLS ARE NEXT FOR PRIMARY CARE?

Discussion Points ▪ AI scribes are already widely adopted in primary care; discussion focused on what tools should come next ▪ Data quality is essential for effective AI implementation; poor coding in general practice is a significant barrier ▪ Clinical decision support tools could benefit urgent care systems and rural practices with limited resources ▪ AI agents could improve triage and appointment booking, directing patients to the most appropriate care provider ▪ Potential for AI to analyse patient populations and help identify optimal workforce composition for community needs ▪ Opportunity for AI to streamline administrative processes like ACC claiming and inbox management ▪ Need for frameworks that help practices assess and implement AI tools with appropriate safeguards Key Actions ▪ Develop a simplified framework for practices to evaluate AI tools (current framework is 50 pages and too complex) ▪ Explore integration of AI tools with patient portals to provide NZ-specific health information ▪ Investigate "matching service" concept to connect patients with appropriate community providers ▪ Consider how AI can help identify missed revenue opportunities in practices (e.g., unclaimed ACC) ▪ Engage with vendors to develop AI-to-AI communication capabilities between systems Privacy concerns remain the biggest barrier to AI adoption in practices ▪ Tension exists between maintaining human connection in healthcare while leveraging AI efficiencies ▪ Different demographic groups have different preferences for digital vs human interaction ▪ Cost is a significant barrier for some AI tools (e.g., triage bots can cost $1-10 NZD per triage) ▪ Need to balance risk management with innovation when implementing new AI tools ▪ Māori data sovereignty and disability inclusion must be considered in AI tool development Additional Notes ▪

Table sponsored by

- 23 -

AI – OPTIMISING BUSINESS PROCESSES

Discussion Points ▪ John demonstrated a simple AI application built in-house for under $1,000 that generates meeting minutes from transcripts ▪ The team discussed using AI to draft media responses by loading previous media logs as context ▪ Participants explored the technical aspects of AI implementation including: • System prompts and instructions • Temperature and top P settings (randomness factors) • Vector stores for providing context • Azure OpenAI service configuration ▪ Data sovereignty concerns were raised regarding where transcription occurs and how data is stored ▪ The group discussed potential issues with AI-generated content including: • Inconsistent outputs when using the same input • Potential for perpetuating past approaches rather than supporting new directions • Challenges with mathematical calculations and numerical data • Limitations in accessing complete meeting transcripts Key Actions ▪ The team plans to open-source the code for their AI application to allow other organisations to implement similar solutions ▪ The media team will incorporate parliamentary questions (OPQs and WPQs) into their AI system to enhance media responses ▪ The organisation will continue evaluating the balance between building custom AI solutions versus using commercial products like Microsoft Copilot Additional Notes ▪ The cost comparison showed significant savings: custom solution costs approximately 10 cents per use versus $50 per month per user for Copilot ▪ Participants noted that even with imperfect transcription, AI can sometimes produce coherent meeting minutes ▪ The group discussed ethical considerations around recording meetings and potential changes in behaviour when people know they're being recorded ▪ Several participants shared concerns about the Public Records Act implications and whether transcripts would be discoverable under the Official Information Act ▪ The discussion highlighted that AI implementation requires careful consideration of governance, data quality, and organisational culture - even for seemingly simple applications like meeting minutes

Table sponsored by

- 24 -

Page 1 Page 2 Page 3 Page 4 Page 5 Page 6 Page 7 Page 8 Page 9 Page 10 Page 11 Page 12 Page 13 Page 14 Page 15 Page 16 Page 17 Page 18 Page 19 Page 20 Page 21 Page 22 Page 23 Page 24