What We Learned Deploying Clinical AI Inside a Public Health Record

What We Learned Deploying Clinical AI Inside a Public Health Record Deploying an AI scribe inside a live public Digital Health Record is a different problem to building one — the technology is the easy part. Over the past period running MedTalk AI inside Canberra Health Services' Digital Health Record, with over 200 registered clinicians now using it across 20+ specialties and 15+ regions, the lessons that mattered most weren't about transcription accuracy. They were about trust, integration depth, and governance.

Why governance-first matters

Before MedTalk AI, I led the Digital Delivery of COVIDSafe at the Australian Digital Transformation Agency. That project taught a lesson that shaped everything since: a government-adjacent digital health tool lives or dies on trust, not features. A clinician or a health system will not adopt a tool they can't audit, explain, or hold accountable — no matter how accurate it is. That principle is the reason MedTalk AI was built around Epic and Best Practice Software certification from day one, rather than as a bolt-on integration added later.

Lesson one: integration depth beats raw accuracy

Most AI scribes on the market today produce reasonably accurate transcriptions. What separates a tool clinicians actually keep using from one they abandon after a week is whether the output lands directly in the EHR they already use, in the format they already expect, without a manual copy-paste step. Every percentage point of "accuracy" matters less than removing that one piece of friction.

Lesson two: adoption is a curve, not a switch

Two hundred clinicians did not start using MedTalk AI on day one. Adoption inside a real health system follows a curve — early adopters who are comfortable trying new tools, then a much larger group who need to see it work reliably for their colleagues first, then the holdouts who need the workflow built into onboarding before they'll touch it. Planning the rollout around that curve, specialty by specialty and region by region, mattered more than any single feature decision.

Lesson three: the audit trail is a deployment requirement, not an afterthought

In a public health system, "can you tell us who accessed this note and when" is not a hypothetical compliance question — it's asked early, and often. Building the audit trail and access logging into the architecture from the start, rather than retrofitting it once a health system asked, is the difference between a pilot that gets approved to scale and one that stalls in procurement.

Lesson four: documentation-time savings translate differently across specialties

The headline number — up to 70% reduction in documentation time — isn't uniform. It's highest in specialties with long-form, narrative documentation and lower in fast, templated consults that were already efficient. Reporting an honest range by specialty, rather than a single average, built more trust with clinical leads than a single impressive headline figure would have.

Where this is headed next

The next phase of this work is less about the AI model and more about the governance frameworks health systems and their boards will need to evaluate clinical AI responsibly as it moves from pilot to standard practice — a topic we've written about separately. If you're a health system or board thinking through what to ask before adopting clinical AI, that's worth a read.