As AI scribes become routine in Australian healthcare, the real challenge is no longer transcription accuracy - it's whether these systems understand clinical context well enough to earn clinicians' trust.
“What is a good night of drinking for you?”
A patient in a rural GP clinic is asked a simple question. To the doctor, the answer offers a starting point for a broader clinical evaluation. To the patient, it’s a standard question. But to the AI scribe, it’s just friendly banter.
The missed datapoint is more than a typo in a transcript. It can substantively impact what is and isn’t picked up in a routine GP visit.
This example stayed with me because it illustrates how easy it is to miss one of the most important elements of healthcare: context. It wasn’t enough for AI to capture what the doctor said. It needed to recognise the clinical intent behind it.
When context is captured accurately, care pathways change. A flagged alcohol history shapes a prescribing decision. A noted social circumstance changes a referral. These aren’t edge cases, they’re the routine clinical decisions that aggregate into patient outcomes.
We caught it because clinicians flagged it. They weren’t passively accepting incorrect output, they were correcting it and pushing for the underlying problem to be fixed. That feedback loop is what made the difference. The question is whether those loops exist at all in most deployments.
Lyrebird started with clinical notes. It became clear how connected notes are to everything else: referrals, care plans, billing and how much that varies by specialty, setting and region.The rural GP clinic, the busy surgical outpatient in a metropolitan hospital, the paediatric ED at 3am.
These are not minor variations. They are completely different clinical environments with different ways of working, different terminology, different patient populations and different clinical risks. No tool can claim to serve all of them well without evidence that it actually does.
This is the real risk facing the industry right now. Not that AI is being deployed, but that it’s being deployed faster than the feedback mechanisms that make it safe to do so. Scale without those loops doesn’t just slow improvement, it obscures where improvement is needed.
Responsible deployment means actively looking for failure, not waiting for it to surface. Our deployment at Gold Coast Hospital, the largest deployment of ambient AI in an Australian hospital setting, with over 1500 active clinicians was subject to an independent, peer-reviewed evaluation, which we agreed to publish regardless of what it found.
The findings were instructive. Strong efficiency gains. Patients reporting more direct time with their doctors. Lyrebird notes scored higher on average on a validated quality instrument. But not 100% of the time.
Across 16 weeks of real-world evaluation across multiple specialties, a number of clinicians encountered something in the note that didn’t look right. In medicine, an average doesn’t absolve you of an outlier. Each anomaly has to be understood and addressed, which is precisely what independent evaluation is designed to surface.
Responsible scaling isn’t about moving slower. It’s about moving differently. A registrar doesn’t specialise on day one. They accumulate enough breadth across settings and patient populations to develop genuine clinical judgment before going deep. The same logic applies to clinical AI. Breadth across contexts builds the understanding needed to go deep where it matters, but only if that breadth is intentional and evaluated.
This year in Australia, at least 42 million consultations will involve AI. That number is only going in one direction. If clinical AI is going to shape what gets trusted, captured and acted on at that scale, generic is not good enough. The standard being set right now will define what patients experience for years.
The tools being built today need to keep earning that trust. Through deliberate deployment, genuine feedback loops, and independent evidence that surfaces what internal teams are too close to see. That standard doesn’t get declared. It gets demonstrated, repeatedly.
This piece is adapted from a keynote delivered at Australia’s Digital Health Festival 2026, Building responsibly as the capability of AI expands
Kai van Lieshout is the founder and CEO of Lyrebird Health. Dr Ray Boyapati is a gastroenterologist and Lyrebird’s chief clinical officer.