Health & life sciences / Industry brief 03
Cooma AI is developing Australian AI estates for health and life-sciences workloads where sensitive data, accelerated compute, human accountability and regulatory boundaries must be designed together.
The decision / Why now
A useful health AI briefing separates research, administrative and clinical intended uses before discussing models or GPUs. That determines the relevant privacy, evidence, safety and procurement work.
OAIC guidance says the Privacy Act applies where AI handles personal information and warns against placing sensitive information into publicly available generative-AI tools.
OAIC · AI privacy guidance ↗The TGA regulates software and AI when their intended purpose meets the medical-device definition, including certain diagnostic, prediction, monitoring and treatment functions.
TGA · AI software regulation ↗The Department of Health's review examined whether current health legislation and regulation provide sufficient clarity and safeguards for AI.
Health department · Final report ↗Workload selector
Select a health or research workload to inspect the likely compute pattern, data sensitivity and operating controls.
01 / Imaging
Create a controlled environment for approved image data, annotation, GPU training and reproducible evaluation before any clinical deployment decision.
The health AI ledger
The ledger is a discovery tool. It does not determine legal status or substitute for privacy, clinical, ethics or regulatory review.
Document legal basis, consent/waiver where relevant, de-identification, cohort provenance, access, retention and deletion.
Track intended use, training data, model version, performance by cohort and known limitations.
Assign clinical/research, privacy, data-custodian and technical accountability with least-privilege access.
Separate research from production; capture runs, outputs, reviewers and release decisions.
Do not make care dependent on an unvalidated service; define recovery and safe clinical fallback.
Proposed deployment pattern
A controlled research and evaluation estate can make data, model and evidence handling more explicit before an organisation decides whether and how a workload should enter production.
Purpose-limited datasets, controlled access and authoritative source systems.
Privacy basis, regulatory triage, evaluation plan, demand profile and decision ownership.
Research, training and evaluation capacity subject to customer and supply agreements.
Keep the workload in research, stop it, or progress through the required assurance and production process.
Australian hosting does not make an AI system clinically safe, legally permitted or TGA compliant. Those conclusions depend on intended use, evidence, organisational controls and the applicable law.
Commercial path
The fastest credible path often starts with research, evaluation, administrative work or another clearly bounded use—not an uncontrolled clinical deployment.
Name the user, decision, data and whether the output could influence diagnosis, treatment or care.
Workload classificationBring clinical/research, privacy, data, regulatory and technology owners into one scope.
Approved discovery basisProduce the data boundary, evidence plan, architecture, GPU envelope and decision gates.
Binding paid SOW · 4–6 weeksBuild and test in a separated environment with pre-agreed measures and oversight.
Paid evaluation milestonesProgress only when legal, clinical, privacy, security and procurement owners approve the path.
Capacity agreement if justifiedHealth & life-sciences briefing
Do not submit health, genetic, personal, clinical-trial or otherwise sensitive information. The first discussion needs only the intended outcome, user and workload type.
Evidence & status
Research updated 23 July 2026
Primary context: OAIC AI privacy guidance · OAIC Guide to Health Privacy · TGA AI software regulation · Health legislation review. Requirements depend on the intended use, entity, data and jurisdiction. Cooma's estate and service descriptions are proposed.