Healthcare admin is full of sensitive personal information, which raises the bar for any AI tool you bring near it. Both Claude and Gemini can help with non-clinical paperwork, but only inside strict privacy controls designed around the Privacy Act and the obligations the OAIC enforces.
Google made a wave of announcements at I/O 2026, and the dust has settled enough to judge them honestly. Plenty of Australian practice managers and clinic owners are now asking what, if anything, they should change. This guide keeps it practical for Australian teams, focused on the trade-offs that actually affect the decision rather than the marketing claims.
What changed at Google I/O 2026
Gemini picked up stronger long-document handling, a cheaper token price, and a more aggressive agentic mode that can take multi-step actions on your behalf. Those are real improvements, and for some workloads they matter. For healthcare admin, though, headline benchmarks are close to irrelevant. The question is not which model scores higher on a coding test. It is which model, wrapped in which controls, keeps identifiable health data safe while still saving your team time.
That reframing is the whole point. A faster, cheaper model that encourages staff to paste a patient record into a chat window is a liability, not a saving. A slightly slower setup with proper data minimisation is the responsible starting point.
Safe use cases for AI in healthcare admin
The wins are in admin, not diagnosis. Booking letters, summaries of non-clinical paperwork, and routine correspondence are fair game when you handle the data carefully and keep a person in the loop.
Drafting appointment, recall, and reminder letters from templates
Summarising non-clinical paperwork and internal process notes
Answering routine practice queries and drafting standard replies
Reformatting and proofreading documents that contain no patient identifiers
Hard boundaries you should not cross
Clinical decisions and identifiable health data need the strictest handling, and often should not reach a general model at all. These boundaries hold regardless of which vendor you choose.
Keep clinical decisions with clinicians, not with a model
Minimise identifiable data sent to any model, and prefer de-identified inputs
Apply the strictest retention and deletion rules to anything that touches health data
Never let an agentic mode email or file patient information without human approval
Privacy by design under the Privacy Act
Health information is a sensitive information category under the Privacy Act, which means a higher standard of consent, handling, and security. Build the controls first, then automate only the safe parts on top of them.
Classify your data before any automation touches it
Mask or remove identifiers wherever the task allows it
Keep an audit record of what was processed, by whom, and why
Check vendor data residency and retention terms before you commit
Where Claude and Gemini actually differ for this work
For non-clinical admin, both models are capable writers. The differences that matter sit around the model, not inside it: how easy it is to constrain what data goes in, how the vendor treats your inputs, and how much human oversight the default workflow builds in.
Our house position at Automata AI is Claude-first, and the reason is practical rather than tribal. We find Claude easier to keep on a tight leash for regulated work: clear system prompts, strong instruction-following on what not to do, and a design posture that does not push staff toward pasting in more data than the task needs. Gemini can be used responsibly too, but its cheaper token price and broader agentic defaults make it easier to over-share by accident, so the guardrails you wrap around it have to work harder.
Instruction discipline: how reliably the model refuses out-of-scope tasks
Data handling: what the vendor does with inputs, and where they are stored
Default oversight: whether the standard workflow keeps a human on approvals
What a breach really costs
A health data breach can cost a practice well over $200,000 once notification under the Notifiable Data Breaches scheme, remediation, and lost trust are counted, and that figure climbs quickly for larger providers. The careful model and a tight privacy design are the only responsible starting point, because the downside is not a slower week. It is a reportable incident, an anxious patient base, and a regulator asking what controls you had in place.
Limit automation to safe admin tasks with no identifiers
Mask identifiers before any processing where you cannot avoid them
Document your controls so you can answer the OAIC with evidence
How to get this right in practice
The pattern across every Australian industry is the same. Automate the routine, keep humans on anything that commits money, law, or client trust, and check accuracy before anything goes out the door. The practices that do well are the ones that start small and stay disciplined.
Start with one high-frequency, low-risk task such as recall letters
Keep a person on anything client-facing or binding
Verify figures and facts before sending
Expand only once a use case has proven itself
Common mistakes to avoid
Across Australian healthcare teams the failure pattern repeats. Owners automate the wrong thing first, let a model touch identifiable data unchecked, or trust output without verifying it. A careful start prevents the costly version of each.
Automating a high-risk task before a safe one
Letting a model commit money or legal positions on its own
Skipping the human check on client-facing work
Assuming a vendor's local handling without verifying it
Scaling before a single use case has proven out
Forgetting to tell staff what is and is not allowed
Key takeaways
Safe use cases live in non-clinical admin, never in diagnosis
Hard boundaries protect identifiable health data and clinical judgement
Privacy by design comes first, automation comes second
Match the tool to the task, keep a human on high-stakes work, and review the choice as models change
Talk to a Claude specialist
Automata AI helps Australian teams design, build, and govern AI workflows with Claude as the core. Book a brainstorm and we will pressure-test your plan against the trade-offs covered above. Get in touch.
