Nine of Twenty AI Scribes Proposed Care the Doctor Never Ordered

The most consequential number in Canadian clinical artificial intelligence is not 12,019. It is 9.

Twelve thousand and nineteen is the number everyone quotes: the primary care clinicians who enrolled in the national AI Scribe Program between June 2025 and February 2026, roughly a quarter of the country’s family physicians. It looks like adoption, and it is quoted as though it were.

Nine is the number of the twenty AI documentation systems in a provincial procurement that Ontario’s Auditor General found fabricated content — inventing findings, or proposing referrals and blood tests the physician never ordered. The rest of that audit is worse in the detail. Twelve of the twenty transcribed a different drug than the one prescribed. Seventeen missed key details about patients’ mental health. Some vendors had not submitted third-party audit reports, certifications or threat risk assessments at all, and four systems were approved regardless.

Now set that against the document published five weeks earlier. An independent evaluation of the national programme, run by the Centre for Digital Health Evaluation at Women’s College Hospital, reported clinicians saving 1.11 hours a week inside working hours and nearly another hour after them; same-day note completion rising from 55.9 per cent to almost 75; delayed documentation falling from 9.8 per cent to under 3; documentation completeness improving on retrospective chart review; and more than nine clinicians in ten saying the tool added value.

Both documents are true. Nearly every reading of Canadian clinical AI resolves the apparent contradiction by choosing a side — either the scribe is a triumph of the clinician’s day or it is a hallucination engine. Both readings miss the point, and the reason they do is the thing that decides who makes money here over the next three years.

They are measuring different objects

One measures whether the working life improved. The other measures whether the artefact is correct. Nothing in a time-saving self-report speaks to accuracy, and nothing in an accuracy audit speaks to workload.

A tool can hand a physician back four hours a month and still write down the wrong drug. That sentence is the whole argument, and it is why the next phase of this market is not an adoption curve. It is an assurance curve, and the incremental spend goes to whoever sells verification rather than convenience. The Canadian market sizing reflects that split: the add-on layer is CAD 24.0 million in 2025 and CAD 97.0 million by 2033, growing 19.1 per cent inside a market growing 6.0.

Not every error is the same size

The audit collapses three failure modes into one headline, and they are not equally dangerous.

A hallucination inside a narrative note is a documentation defect. It propagates to the chart, to the next clinician and to the billing record, and a physician who reads what they sign can catch it. A hallucination that proposes a referral or orders a blood test crosses from text into action. The physician’s review stops being a proofread and becomes a veto of something already generated — a different cognitive task, performed at a different point in the workflow, with a different failure mode when it is skipped in a busy clinic.

That gradient matters because of where the roadmap is pointed. Asked what the next generation should do, the clinicians in the evaluation named referral drafting, inbox management and clinical decision support. All three sit further along that gradient than transcription. The market is walking deliberately toward the failure mode that just scored 45 per cent — and the procurement response, an enforced attestation that a human reviewed the note, is aimed at exactly that.

The ceiling is context, not model quality

It is tempting to read the audit as a verdict on model quality and assume a better model fixes it. The engineering constraint is elsewhere.

In the 2024 national survey, 73 per cent of physicians named poor system integration or multiple unconnected systems as a major barrier, and 68 per cent reported spending an hour or more a day hunting for patient information they needed. That is the context a scribe is working from. It hears the room and nothing else.

It cannot know that the patient’s metformin changed last week if that change lives in a hospital report that never reached the record. In a market where 95 per cent of physicians already use an electronic record and 65 per cent run exactly one system, accuracy is bounded less by the model than by how much of the patient’s history the model is allowed to see. Which makes integration a technology problem rather than merely a commercial one. Whoever controls the record controls the context, and context is what separates a note that reads well from a note that is right.

Which is why the product is the proof, not the note

The second structural fact is sitting on a vendor’s own home page. Accuro — 7,784 Ontario community seats, the second-largest base in the country — now ships Accuro AI Scribe, powered by Heidi, integrated inside the EMR. The specialist supplies the engine; the platform supplies distribution, integration, billing relationship and brand. The clinician does not choose a scribe and connect it. They choose an EMR and receive one.

That leaves standalone scribe vendors exposed and the platform owners only slightly better off. Three vendor groups hold 90.6 per cent of Ontario’s 27,101 community seats, and 91.5 per cent of those seats are hosted rather than local — the cloud migration that produced a decade of repricing is finished. An EMR vendor can ship a scribe in a quarter. It cannot conjure a validated quality regime, third-party audit evidence and enforced clinical attestation out of its integration layer, and the audit is direct evidence that at least some vendors were not producing those artefacts even when procurement asked for them.

So the defensible product is not a better note. It is the apparatus that makes a note verifiable: extraction of every clinical assertion, an audit trail that separates what the physician said from what the model generated, and an attestation that a human read it. The Auditor General’s recommendation points straight there — IT controls that enforce a physician attestation that notes were reviewed. That single line does more to shape the next product cycle than any privacy statute, because it converts review from a professional habit into a recordable system event.

The general AI stack supplies none of this, because it is domain-specific, unglamorous and cannot be demonstrated in a product demo. It is also what a buyer facing a published 45 per cent hallucination rate will actually pay for.

And then there is the money

All of the above would be a quality essay if this market funded itself. It does not, and that is what makes the 2026 renewal a decision rather than a forecast.

The AI add-on layer was CAD 24.0 million in 2025. Roughly CAD 18.0 million of it is the national programme’s funded licences — twelve months, baseline functionality, everything beyond at the clinician’s cost. Active use is the only public proxy for what happens when that ends: 7,400 of 12,019 clinicians, 61.6 per cent, were using a scribe ten or more times a month by February 2026.

A spread of CAD 5.6 million to CAD 11.1 million is around a quarter of the entire 2025 layer resting on a decision no clinician has yet had to make with their own money.

The renewal case looks unanswerable on paper. 1.11 hours a week across 46 working weeks is 51.06 hours; at CAD 239.38 an hour — CIHI’s CAD 383,000 average gross clinical payment divided by an assumed 1,600 fee-generating hours — that is CAD 12,222.74 of returned time against a CAD 1,500 licence. Eight times over. It is unanswerable on paper precisely because the party who captures the value is not reliably the party who pays: half the clinicians redirected the saved time into other administrative work and only a fifth into additional appointments. In fee-for-service, paperwork pushed to 6 p.m. is a personal benefit and a thin financial one. In a salaried or capitated practice it is not a revenue event at all.

Which is where the auditor’s finding stops being a risk note and becomes a pricing mechanism. Third-party audit reports, certifications, threat risk assessments, validation datasets, clinical safety cases, audit-trail infrastructure and the insurance that follows a documented failure rate are not optional once a province has been criticised for accepting their absence. They are cost of goods. On a CAD 100-a-month scribe, that compresses margin exactly where the 19.1 per cent growth lives — and the revenue cycle is where the saving has to be monetised if the clinician is not going to fund it personally.

Three things that would prove this wrong

One. A disclosed acquisition of a Canadian scribe vendor by a platform owner. That would confirm the component view and collapse the two-layer split this analysis rests on.

Two. Renewal data well below the 61.6 per cent habitual-use rate. If fewer than half of active users re-subscribe at market price, the CAD 24.0 million base is a programme artefact rather than a market.

Three. A provincial requirement that AI-generated notes be flagged, attested or separately audited. That converts an optional productivity tool into a compliance obligation — raising price per seat and lowering attach, moving two variables at once.

How the number was built, and where it is soft

One construction, not two. Roughly 67,800 community physician seats — an extrapolation, because Ontario’s self-reported vendor census is the only large seat-level count in the country (New Brunswick now publishes one too, at 888 seats against Ontario’s 27,101), and applying Ontario’s 68.1 per cent penetration to CIHI’s 99,555 physicians is an assumption, stated rather than buried. CAD 2,800 per core seat, deliberately below Ontario’s CAD 4,800-a-year Enhanced Use subsidy, because that subsidy bundles connectivity a software-only measure should not claim. CAD 1,500 per attached add-on seat, taken from published list — CA$100 to CA$150 a month for the unlimited scribe tier, with enterprise pricing quoted privately and certain to be lower.

Hence a band rather than a point: 4.6 per cent in the low case (CAD 306.3 million by 2033), 6.0 base (CAD 339.8 million), 7.5 high (CAD 380.2 million). Any second route available would have restated the same two unknowns — how many community physicians run which kind of seat, and what each of them pays — dressed as an independent series. One construction with a stated band is worth more than two that share the same guess. The global EMR platform is the distribution channel for this layer, not its competitor, and it is sized on entirely different assumptions.

The practical reading

For a buyer, the question is no longer whether the note is faster. It is whether the note is defensible, and the audit is the strongest available evidence that the market has not answered that yet. For a seller, distribution belongs to whoever owns the seat, so the sellable thing is the capability the seat owner cannot build — evidence, audit trail and attestation. For an investor, 19.1 per cent is a forecast rather than an observation, and the number that turns it into one or the other is 7,400.

The full two-layer sizing, the vendor census, the renewal arithmetic and the four ranked risks are in the Canada EMR and clinical AI add-on market report.