Keeping health data onshore — without giving up on AI.

Artificial intelligence has become genuinely useful in exactly the places health services are stretched thinnest. It can sit quietly in the background of a consultation and write the clinical note, so the clinician can look at the patient instead of the keyboard. It can transcribe a phone call and file it against the right client, so a recall or a result is never lost in someone's memory. It can turn a long, messy conversation into a short, searchable summary. For an organisation running clinics across large distances with limited staff, that is not a gimmick — it is time given back to care.

But almost every one of these tools shares the same catch. To work, it sends your audio and your text to a server somewhere overseas, processes it there, and sends the result back. For a suburban retailer that is a footnote nobody reads. For an Aboriginal Community Controlled Health Organisation, it is the whole question — and usually the end of the conversation.

Why sovereignty is not negotiable here

"Data sovereignty" sounds like an IT term, but it is really about two simple things: where your data physically lives, and who gets to control it. For Aboriginal health, both of those carry weight that goes well beyond ordinary privacy.

There is a formal principle behind it. Indigenous Data Sovereignty holds that data about Aboriginal and Torres Strait Islander people and communities should be governed by them, in their interests — not held and used by whoever happens to run the software. In Australia the principle was set out by the Maiam nayri Wingara collective and endorsed at the 2018 Indigenous Data Sovereignty Summit, and it is commonly expressed through the CARE principles: Collective benefit, Authority to control, Responsibility, and Ethics. The thread through all of them is self-determination — an Aboriginal community deciding what happens to information about its own people.

Sitting on top of that is the ordinary regulatory layer every health service must respect: the Privacy Act and the Australian Privacy Principles, health-records and consent laws that vary by state, and the handling rules around My Health Record and protected health information. And underneath both is the thing that actually makes an ACCHO work — community trust. People share sensitive things with a service they believe is theirs. Quietly shipping those conversations to an offshore AI vendor is not a technicality; it breaks the very relationship the service is built on.

Put together, this means an ACCHO does not get to treat "process it onshore, under our control" as a nice-to-have. It is a requirement. Which sets up an awkward problem, because the best-known AI tools are built to run in exactly the opposite way.

What you would actually use it for

Before the test, it is worth being concrete about why a health service would want speech-to-text in the first place. It is the quiet engine underneath a surprising number of everyday jobs:

• Clinical documentation. Transcribe and summarise a consultation and draft the note straight into the clinical system, so the clinician writes less and watches the patient more.
• Phone-call records. Capture inbound and outbound calls in full and file them against the client — so care coordination, recalls and results are never lost to a half-remembered conversation.
• Telehealth. The same capture across remote and outreach sites, where a clinician may be the only one in the room.
• Language support. Transcription as the first step toward bridging English and community languages.
• Triage and recall summaries, and after-hours message capture, so nothing falls through overnight.
• Governance. Board and meeting minutes, and records for social and emotional wellbeing programs.
• Search. Making the whole history findable — every mention of a condition, a medication or a follow-up, across thousands of calls.

Every one of these rests on transcription, and every one of them touches sensitive patient or community information. That is precisely why each one has to run onshore.

The question that actually mattered

Here is the tension. The most capable, best-known transcription engines are offshore cloud services — the ones already embedded, by default, in most phone systems. The quiet assumption in the market is that offshore equals state-of-the-art, and that anything you run locally is a compromise. If that assumption were true, an ACCHO would face a genuinely bad trade-off: sovereignty or quality, pick one.

We did not want to assume. We wanted to measure. So we ran real calls through two engines side by side — the offshore cloud default, and an engine we run entirely on our own infrastructure in Australia — and simply compared what each one wrote down.

How we measured it

We captured 991 real business calls and transcribed every one of them twice from the same audio: once offshore, once onshore. It matters where these calls came from: they were our own everyday business calls — managed-IT support and sales conversations — and deliberately not patient or health-related calls. Running the comparison on our own operational calls let us benchmark the two engines rigorously without any sensitive clinical or community information ever being involved. Then we asked two plain questions of the transcripts. First, does it get the important words right — the names of people, places and organisations that a record is actually searched on? Second, does it capture everything that was said, or does it quietly drop content, especially as calls get longer? For that second question we used a simple stand-in — how many characters of text each engine produced per minute of audio — as a measure of how much of the conversation actually made it onto the page.

The names it gets wrong

On getting a difficult but important name right, the onshore engine was correct on 90.9% of calls; the offshore engine managed 57.8% (Figure 1). In a clinic, this is not cosmetic. The words most likely to be mangled are the local ones — a community's name, a place, a medication — and those are exactly the words a health record is indexed and recalled on. An offshore engine that mishears one name in three is corrupting the field that matters most.

The parts it quietly leaves out

The more worrying finding is not about wrong words, but missing ones. Both engines capture fewer characters per minute as calls get longer — long calls have more pauses and back-and-forth — but the offshore engine falls away far faster (Figure 2). Through short and medium calls the two are close. On calls of twenty minutes or more, the offshore engine captured little more than half as much as the onshore one.

How much actually goes missing

Seen as a percentage — offshore content measured against the onshore transcript of the same call — completeness holds at 93–95% through the one-to-ten-minute range, and then collapses to 64% on the longest calls (Figure 3). Across the whole sample, one call in eight had the offshore engine capturing barely half of what was said. A fifteen-minute care conversation reduced to a couple of sentences, filed as "the record", is worse than no record at all — because everyone downstream trusts it.

What this means for an ACCHO

The result we were hoping for was that the onshore engine would be good enough — close enough that sovereignty would only cost a little quality. What we actually found was that it was better, on both measures, and most decisively on the long, high-value calls. That turns the supposed trade-off on its head. You do not have to choose between keeping data onshore and getting a usable transcript; onshore was the more accurate and more complete option here.

Which means every one of those applications — the ambient clinical note, the complete call record, the searchable history — is available to an ACCHO without compromising Indigenous Data Sovereignty, community trust or the Privacy Act. The technology that had been the reason to say "we can't" is, on this evidence, the reason you now can.

This is one study, and we should be plain about its edges. Character count is a proxy for how much was captured, not a formal word-for-word error rate. The audio is from a single organisation, Australian-accented, so results may differ elsewhere. And the onshore engine is deliberately tuned for local vocabulary — which is part of the point, but does mean an untuned local model would not perform identically. None of that changes the headline: staying onshore did not cost accuracy here; it added it.

For Aboriginal health, the choice was never really AI versus sovereignty. It was a false choice, built on an assumption nobody had tested. With models that run in Australia, under community control, an ACCHO can have the time-saving and the completeness that AI transcription promises and keep faith with the principles that make it a community-controlled service. That is the foundation of Sovereign Call Intelligence — and we are happy to run this same comparison on an organisation's own calls, so the evidence is theirs, not ours.