AI for healthcare in NZ and Australia: data that stays onshore.
For practices, clinic networks and healthcare groups. Claude and ChatGPT for the work around the clinic, with patient data processed in Australia or New Zealand, what we never do, and how a rollout runs. Published 5 November 2026.
The short answer
A New Zealand or Australian practice can run Claude for clinical admin and operations without patient data leaving the region. Every request is routed through a gateway the practice owns to Amazon Bedrock in an Australian region, so nothing goes to a vendor's own cloud. That is how Sentry AI deployed Claude across 15 healthcare practices serving around 180,000 patients, from reception to the board. The AI drafts and a person approves; nothing writes to a patient record on its own, nothing is trained on patient data, and nothing decides how sick someone is.
Where AI for healthcare pays off: around the clinic, not in it
Most healthcare providers have met AI already: a dictation tool, Copilot in the practice manager's inbox, a vendor demo that promised to take the phones. Little of it reached production, because a practice cannot run patient data through a tool it does not control. Fix the routing and four places open up, none of them clinical decision making.
The front desk
Bookings, repeat prescription requests, opening hours, results enquiries: the same questions, hundreds of times a day. An assistant takes the repetitive load and hands anything it cannot resolve to a person.
Administrative drafting
Referral letters, recall letters, correspondence, replies to the practice inbox. Drafted, never sent: a clinician or administrator approves every one. The first draft is where the hours go, and it is the safest place for a model.
Staff questions
Policies, funding rules, enrolment procedures, standing instructions that three people know and everyone else asks about. An assistant that has read the same pages answers them, and the pages hold no patient data.
Privacy and compliance paperwork
Every AI workflow in a healthcare setting needs a privacy impact assessment before it touches a patient. Written in the house format from the process it describes, so the compliance work keeps pace with the build.
The proof
Claude across 15 practices, every inference in Australia
Caregroup runs 15 medical and emergency practices in East Auckland, around 180,000 patients between them. They had Claude and Copilot licences and almost nobody using them on patient-adjacent work, because no privacy officer would sign it off. We put Claude Desktop on every seat pointed at the group's own gateway on AWS, which sends every inference to Amazon Bedrock in Australia. The same frontier models, inside a boundary the practice already controls, under the same access rules and audit trail as the rest of their estate.
On top of that: Claude from reception to the board, a process library captured by recording clinicians and receptionists doing the real work once, privacy impact assessments written in the house format, and a voice receptionist for the whole network, connected to the Indici patient management system, that hands to a person when it should and to 111 in an emergency. The standing rule across all of it: AI drafts, a human approves.
Sovereign Claude across 15 healthcare practices
How inference was pinned to Australia, how Claude reached every seat, and what was built on top. Every figure on this page comes from there.
What we never do with patient data
These are the lines a privacy officer asks about first, and the answers do not change from one engagement to the next.
- Train a model on patient data
- Move patient data offshore, or to a vendor's own cloud
- Let an AI write to a patient record on its own
- Let an AI decide how sick someone is
- Put patient identifiers in a knowledge base, a wiki or an eval page
- Capture what staff type: the eval reads what was done, never what was said
The vendor side is simpler than it sounds. Claude's own apps do not list Australian residency and ChatGPT's own apps do not process in Australia, which is why the routing through Bedrock is the whole answer rather than a detail. Our Claude vs ChatGPT page holds the residency table, dated against each vendor's documentation.
How a rollout runs
The same way every team we move onto Claude is moved: one workflow at a time, taught in person, measured against the baseline timed in week one. Three parts.
- Step 1
The Claude training day
One day in your own rooms. Every role set up and using Claude on its own work before we leave, the first skills built from jobs your team names, and the baseline for each job timed with the person who does it. The goal is set in the room: by default, twice as fast as your own baseline.
- Step 2
Skills per role
Reception, the practice manager, the clinical leads, the board: each gets the skills for its own jobs, shipped to every seat, with access by role so finance pages are not open to everyone and clinical workflows reach only the people cleared for them.
- Step 3
The daily eval
Every morning a check reads what the AI did the day before: which skills were used and abandoned, which tools failed, what people accepted or threw away. It picks at most one thing to fix, waits to see it twice, then fixes the skill. The prompt text and the answers are never captured.
How the day itself runs is on the Claude Day page; how the morning check works is on AI evals. How Claude is connected to the systems a practice already runs, with permissions per person, is explained on our MCP page.
Where the rule is stricter
Where sovereign and on-premises models fit
Australian processing answers most New Zealand practices. Some rules say the data cannot leave New Zealand at all, and some groups run enough volume that owning hardware beats metering an API. For those, an open weight model in your building or a New Zealand data centre keeps everything onshore, with a frontier model kept for the hard cases. Our sovereign AI page sets out the three options and when each fits; private LLM infrastructure is the service.
We are also building our own open weight model stack, starting on a Mac mini in our office, to prove what a machine on a desk can carry. On-premises custom models are what that work builds towards, not something we sell today; the build log is public.
Governance a board and a privacy officer can read
Every AI workflow gets a privacy impact assessment before it touches a patient, a named owner, and a way to stop it. The register of what runs, who owns it and what it may do is the core of AI governance as we practise it, and it maps onto ISO 42001 if your board wants the standard. We run our own register in the open. We do not certify anyone and do not claim a certificate we do not hold.
Further reading: the longer 2026 guide for New Zealand healthcare providers, covering the Health Information Privacy Code, the practice management systems that can be connected, and where a single practice starts.
Frequently asked questions
Can a New Zealand or Australian practice use Claude or ChatGPT without patient data going offshore?
Yes, with the right routing. Claude through Amazon Bedrock in an Australian region processes every request onshore, inside an AWS boundary the practice controls. That is how a group of 15 practices in East Auckland runs Claude on every seat. Claude's own apps and ChatGPT's own apps do not process in Australia, so the routing is the whole answer, not a detail.
Does the AI train on our patient data?
No. Nothing we deploy trains on your data, the business plans and the Bedrock route keep your content out of model training by default, and we never move patient data to a vendor's own cloud, to a region you have not approved, or into a knowledge base. The documentation we build for a practice is hosted in region and holds no patient identifiers.
What does AI actually do in a practice?
The work around the clinic, not the clinic itself. Front-desk enquiries, bookings and repeat prescription requests; drafting referral and recall letters for a clinician to approve; answering staff questions about policies, funding rules and enrolment procedures; and privacy impact assessments written in the house format. Nothing decides how sick someone is.
How does a rollout run?
It starts with a training day in your own rooms: every role set up on Claude, the first skills built on your real work, and the baseline for each job timed with the person who does it. Then skills per role, from reception to the practice manager to the board, and a daily eval that reads what the AI did the day before and picks one thing to fix.
Where do sovereign or on-premises models fit?
Where the rule says the data cannot leave New Zealand at all, or where volume makes owning the hardware cheaper than metering. An open weight model in your building or in a New Zealand data centre keeps everything onshore. We are building our own model stack to prove what that can carry, and we are honest that on-premises custom models are what we are building towards, not something we sell today.
What about governance and ISO 42001?
Every AI workflow in a healthcare setting needs a privacy impact assessment before it touches a patient, a named owner, and a way to stop it. We help you keep that register, and we run our own ISO 42001 register in the open so you can see what one looks like. We do not certify anyone and we do not claim a certificate we do not hold.
Running a practice or a healthcare group?
In a short call we will look at where your patient data may be processed, which jobs around the clinic to start with, and what a training day in your rooms would cover.
Ask an AI about this page
Opens the assistant with this page loaded: read it, summarise it, cite it.