The tool arrived before the rules did
Health is the most tightly regulated sector in the country, which makes the adoption curve remarkable. Newsroom reports that nearly 80% of New Zealand clinicians are now using AI tools to cut administration, with clinicians saying the tools free up roughly an extra hour a day for patients. In general practice the figure is higher still: Anna Campbell, chief support services officer at Whakarongorau Aotearoa/NZ Telehealth Services, told the NZ Herald that 84% of general practices are now using AI for note-taking and scribe functions.
The demand is easy to understand once you look at the time saved. In November 2025, RNZ reported that trials in Hawke’s Bay and Whanganui cut average note documentation time from 17 minutes to 4 minutes using the AI scribe Heidi, prompting Health NZ to buy 1,000 licences for a nationwide emergency department rollout. Emergency doctors now estimate the tool saves up to 10 minutes per patient, with Health NZ’s director of digital innovation and AI Sonny Taite saying clinicians consistently report it reduces documentation time and frees them for patient care.
Adoption ran ahead of governance
The adoption raced up the curve while the framework lagged. A round-two primary care survey in early 2025 found 68% of respondents using AI tools, up from 52%, with daily use climbing from 25% to 40%. By the Phase 3 survey in August 2025, 48% were using AI daily and 52% were using a free transcription tool. The Phase 3 report explicitly flagged an urgent need for sector-wide policy frameworks and workforce development.
That free-tool figure is the tell. Shadow IT was already entrenched before anyone drew a line. In March 2026, Health NZ issued a directive telling staff to stop using free tools such as ChatGPT and Gemini for clinical notes, stating their use was “strictly prohibited due to data security, privacy and accountability concerns.” The ban even covered drafting notes with AI and transcribing them manually, and required any tool to be registered with Health NZ’s national AI advisory group. In other words, the rules arrived after roughly half the sector had already built a workflow around unapproved software.
Consent is the specific exposure
The sharpest liability sits in a detail most clinics gloss over. The Medical Council’s March 2026 guidance now clearly defines informed consent requirements for AI in patient care, meaning ticking a box or a passing mention no longer clears the bar. True consent means the patient understands where their data is stored, whether it is anonymised, and how it is used.
Campbell frames the risk in language any business owner will recognise. “The people who buy and run these tools are telling us the thing that frightens them is not the price,” she told the Herald. “It is the fine print nobody reads, and what happens to a patient’s private information once it disappears inside a system they do not control.” Her conclusion cuts against the usual framing: “Governance is not the handbrake on AI in healthcare. It is what makes it safe to bring a tool near a vulnerable person at all.”
The clinical accuracy problem
Speed is not the same as safety. In November 2025, Otago University bioethics professor Angela Ballantyne told RNZ that the Hawke’s Bay trial reviewed only eight clinicians’ experiences before national rollout, and cited a case where an AI tool changed a patient describing a symptom “in the morning” to “every morning” – a small edit with potentially serious treatment consequences. In April 2026, specialist GP Karl Cole described the field to NZ Doctor as “the Wild West, with no dedicated regulator and no real safeguards.” But he also spotted the upside, arguing the gap gives primary care a chance to shape tools to suit the sector, with independent testing and strong vendor contracts as the safeguards.
What every white-collar sector should read into this
Health is the canary. If 80% of clinicians adopted AI before the rules were settled, and half were using unapproved free tools, the same dynamic is running through every professional services firm, law practice and accountancy in the country. The pattern is consistent: a genuinely useful tool gets adopted from the bottom up, shadow IT proliferates, and liability quietly compounds until the first complaint lands.
The fix is not to slow down. It is to build the scaffolding while adoption is still cheap to steer: approved tool lists, data processing agreements with vendors, staff training, and clear consent where customer or employee data is involved. The organisations that treat governance as the precondition for adoption rather than a brake on it will be the ones still standing after the first regulator or client comes knocking.
Sources
- Doctors adopting AI faster than rules can keep up (2026-07-15)
- What frightens GP clinics using AI and the fine print nobody reads – Anna Campbell (2026-07-16)
- Emergency doctors estimate AI scribe ‘Heidi’ saving up to 10 minutes per patient (2026-07-16)
- Health NZ staff told to stop using ChatGPT to write clinical notes (2026-03-26)
- Guidance on using artificial intelligence (AI) in patient care – Medical Council of NZ (2026-03-10)
- AI in healthcare ‘like the Wild West’, GP warns (2026-04-09)
- AI scribe tool rolled out to emergency departments, promises to slash clinicians’ admin (2025-11-20)
- Round two AI in Primary Care survey – Collaborative Aotearoa (2025-03-20)