How we deployed sovereign Claude across 15 healthcare practices for Caregroup

A network of medical and emergency practices across East Auckland, 180,000 patients between them, needed frontier AI in front of clinical staff without a byte of patient data leaving the region. We pinned inference to Australia, put Claude on every seat from reception to the board, and started building the workflows on top.

What they were paying for

A network this size runs on Microsoft 365, SharePoint, the Indici patient management system and a lot of people stitching the three together by hand. The practices had Copilot and Claude licences. What they did not have was a way to use a frontier model on patient-adjacent work that a privacy officer would sign off, so the licences sat with the people who could avoid patient data and nobody else.

The first thirty days of usage data after the rollout showed 24 people active, from reception to the board, and nine of them going looking for a skills library on their own. The demand was there. What was missing was the sovereign path and the harness.

Pinning inference to Australia

A hyperscaler lets you choose where a model runs; a model vendor's own cloud does not. Caregroup's Claude Desktop seats route through a gateway they own on AWS, and that gateway sends every request to Amazon Bedrock in Sydney and Melbourne. The same frontier models, inside an Australian boundary the practice already controls, with the same identity and audit trail as the rest of their AWS estate.

That one decision is what made the rest possible. It is the difference between AI a medical network has to keep staff away from and AI a clinician can use on a Tuesday afternoon, and it is why we offer sovereign and private LLM infrastructure as a service in its own right. The wiki, the register and the eval pages follow the same rule: hosted in region, holding no patient identifiers.

What we installed

Claude Desktop on the practice's own organisation, for four groups who each need something different from it. Then the harness over the top: the skills for the programme's real jobs, a wiki of how the network runs that the assistant reads before it guesses, and the standing rule that a person approves anything that changes a record.

The skills

The library is written for the programme that is building the clinical workflows, and for the administrators and clinicians who feed it. A healthcare programme produces documents in a house format, needs a privacy impact assessment for every new workflow, and has to capture how a clinician actually does a job before anything can be automated. The skills do those three things.

process-capture

Clinical · the workflows

A clinician or receptionist records themselves doing the job once, screen and voice. The skill turns the recording into a written process: every click, every field, every decision. That process library is what the GUI automation is built from.

IndiciWordSharePoint

privacy-impact-assessment

Compliance · every new workflow

Every AI workflow in a healthcare setting needs a privacy impact assessment before it touches a patient. This skill writes it in the house format, from the process capture, so the compliance work keeps pace with the build instead of trailing it.

WordSharePoint

where-things-live

Everyone · the library map

The biggest failure in the first month of telemetry was people asking Claude to read files it could not find. Four in ten reads failed. This skill is the map of the programme library, so the assistant looks in the right place first.

SharePointTeams

The rest cover the programme office: meeting minutes, project requirements, the cost model, the risk register, slide packs, and a getting-started skill for a new seat. Each one reads from the systems it needs, and nothing else.

IndiciIntelliMedAWSMicrosoft 365SharePointWordExcelPowerPointTeamsOutlook

Five groups, one library

Each group got the skills for its own jobs and the connections those jobs need. The clinical leads reach the patient management system; the programme office does not.

Programme office	Meeting minutes, project requirements, the cost model, slide packs	WordPowerPointExcelSharePoint
Clinical leads	Process capture, the process library, desk research	IndiciWordSharePoint
Practice administration	House documents, where things live, getting started	SharePointTeamsOutlook
Privacy and compliance	Privacy impact assessments, the risk register	WordSharePoint
Leadership and board	The governance questions, the benefits framework, the steering pack	PowerPointTeams

The governance questions

Building this properly surfaces questions a practice cannot answer alone, and should not answer by default. Who is the privacy officer for an AI workflow. What happens to a clinician's screen recording once the process has been captured from it. Which data may cross the region boundary, and which never may. When the benefits framework is allowed to call a saving measured rather than observed.

We do not settle those quietly. Each one is written up with the evidence and put to the board, so the record of who decided what exists before the first patient-facing workflow goes live. That record is also what an ISO 42001 auditor asks for first, and it is the same discipline behind our AI governance work.

What the loop found

The usage data from a sovereign estate arrives de-identified, so the eval sees what was done and never who did it. Two findings from the first month.

Nine of the twenty-four active people called the skills tool on their own, looking for a library, and found only the built-ins. That is the clearest demand signal an eval can produce, and it set the order of the build.

Reads of the programme library failed four times in ten, because nobody had told the assistant where things live. That became the first skill everyone gets.

What Caregroup owns

The gateway, the skills library, the process library captured from their own clinicians, the wiki and the governance record, all in their name and inside their AWS boundary. Models will change and staff will move on. The record of how the network actually works stays, and it keeps compounding.

The custom development underneath

Clinical workflows need more than a skill file before AI can help with them, so alongside the library we are doing the custom development necessary to enable them. Two pieces carry most of it. A GUI interface over the patient management system, built by recording clinicians and receptionists doing the real work and understanding every click and decision in it, so the AI prepares the screen and a person confirms it. And an AI voice receptionist for the whole series of healthcare and emergency practices in East Auckland, 180,000 patients and counting, handling bookings, repeat prescriptions and general enquiries, escalating to a person or to 111 when it should. The patient forms that feed both are already live on their own URL, in English and Chinese.

Could this be your practice?

If you are holding AI licences your clinical staff are not allowed to use, or your front desk is buried under routine calls, this is the problem we solve. We start with a free AI Opportunity Audit: enter your company, connect the tools you already pay for, and see your three biggest opportunities mapped before you spend a dollar.

Frequently asked questions

More case studies