Scribeberry

Ambient capture, note generation, template matching, form population and the autonomous intake agents are all model output, with no services layer or platform business underneath.

The scribe itself is conventional draft and review, with the clinician checking the note before it is pushed or pasted into the record. One adjacent capability goes further: ADA, the Autonomous Doctor Assistant, performs autonomous history taking with patients, and the platform also runs AI receptionists and autonomous booking and rebooking. No confidence threshold, escalation path or abstention behavior is published for any of those patient facing agents. The documentation product keeps a clear human gate; the wider platform adds autonomy that has not been described.

The disclosure gap sits oddly against an otherwise exemplary compliance posture. Accuracy is claimed at 99.9 percent with no methodology, reference standard, denominator or date, which leaves it unfalsifiable, and no model card, named models or evaluation protocol is published. Template counts are also stated inconsistently, appearing as more than 2,000 on the vendor's own site and as more than 200 in a provincial procurement profile. A vendor this precise about encryption ciphers and data residency could be equally precise about model performance and is not.

The vendor's trust center names its providers directly. Microsoft Azure and Google Cloud are named as the cloud providers; Microsoft, Anthropic and Google are named as the main infrastructure providers; and signed data protection or business associate agreements are stated to be in place with Microsoft Azure, Google Cloud, OpenAI and Anthropic.

That reaches the model layer, and it comes with two data terms: content is stated not to be retained permanently, and never to be used to train any model.

One part of the chain is structurally shorter than the rest. Transcription is described as running on the vendor's own self hosted encrypted service rather than through a third party speech provider, so the audio half does not travel to a further party at all.

The vendor's own wording sets the limit. The list is given as an answer to a question rather than as a subprocessor register, and it closes with a clause extending to any other third party vendor used, so it is a naming rather than an enumeration. No model family, version or update policy is given for either model provider, and which provider serves which function is not stated.

External standing rather than measured outcomes. Scribeberry is preferred vendor to the Canadian Rheumatology Association, a specialty society designation, and appears as a named vendor in a Supply Ontario procurement process for AI scribe solutions, which is provincial public procurement participation. Adoption is claimed at more than 30,000 providers.

A professional body selecting a vendor may be a commercial relationship rather than an evaluation, and no assessment criteria for the society designation are published. No study, controlled comparison or published accuracy benchmark exists.

A complete data handling position, with every major commitment covered rather than some of them.

Retention. No audio recording is created or stored at any point; transcription runs in real time on the vendor's own self-hosted encrypted service. Note text is held temporarily and encrypted for the single purpose of synchronizing across a user's devices, and the vendor states nothing encrypted remains once the user deletes the note.

Training use. Patient data is stated never to be used to train any model, with no qualification limiting the promise to third party models only, and the vendor states it cannot see the content itself.

Protection. Encryption is given by named cipher on the product pages, AES-256 at rest and TLS 1.3 in transit, and regional residency is applied rather than offered as a request.

Two of these points are narrower than they first appear. The statement that identifying characteristics are removed refers to voice characteristics such as accent and language, which disappear as a consequence of audio not being retained; that is not the same as de-identifying the text before it reaches a model, and no text de-identification step is described. And the encryption description is looser on the trust center than on the product pages.

One class of question is not answered at all. Nothing published describes safety engineering: no guardrails, no hallucination mitigation, and nothing on how a safety event is identified, escalated and closed. The commitments cover what happens to the data, not what happens when the model is wrong.

Broad regulatory coverage, claimed across two federal regimes and the health privacy statute of every Canadian province, including Quebec's modernized framework under Law 25, which imposes its own consent and transparency obligations distinct from the federal regime. The vendor states directly that it signs business associate agreements.

The customer facing agreement cannot be read before contracting. The trust center's document index does list business associate agreements, but they are the vendor's own upstream instruments with its cloud and model providers. Those establish the chain behind the service rather than the terms a customer would sign.

What is published is substantial. Terms and conditions and the privacy policy are downloadable in full, alongside a privacy impact assessment, an independent security risk assessment, a privacy compliance statement and a privacy controls document. A patient informed consent form is published for clinicians to use, which addresses the covered entity's own obligations rather than only the vendor's.

One question the claim across several regimes raises and does not answer: a business associate agreement is a United States instrument, and the equivalent for a Canadian customer is a different document under a different statute.

SOC 2 Type 2 is named and typed in the vendor's own material, which is real and uncommon at this size. What is missing beyond naming is the artifact: a certification presented as something a buyer can open without asking. Here the trust center states the position and then gates the report. Access to the SOC 2 is by email request, and the vendor notes that some documents require a non-disclosure agreement before release because of its own agreements with third parties.

Neither half of the clock is given. No audit period, no report date, no renewal date and no scope statement appears for the certification in anything published.

A wording question sits alongside it. The trust center describes the position as SOC 2 Type 2 compliant; the product pages describe it as SOC 2 Type 2 certified. A SOC 2 engagement produces an attestation report against a defined scope for a defined period rather than a certification, so the two words describe different things and neither of them is the report.

What is retrievable without asking is substantial: a privacy impact assessment, an independent security risk assessment, a privacy compliance statement and a privacy controls document, all downloadable, plus a continuously updated third party security monitoring dashboard published openly. The dated items carry January 2024 dates, which is the clock question in another form.

No clearance claimed and none required for the documentation product, and no United States device pathway attaches to a note the clinician reviews and signs. The autonomous intake agent is now part of the shipped feature set, and it makes this a clear patient facing case.

ADA and the wider agent framework do not assist a clinician. They speak to patients directly, by inbound or outbound phone call, text or web chat, gathering history and completing forms before the clinician is involved. The vendor's own example opening line has the agent introduce itself as the doctor's assistant calling to gather medical history, and the agent's questions are described as adapting and responding to the patient at the level of a second year resident. That is a clinical competence claim about an autonomous system talking to a patient with no clinician present.

The reasoning that keeps clinical decision support outside device regulation rests on a professional being able to review the basis of what the software produces. A patient is not a professional. Where the software is the party asking the questions, the patient reviews nothing, and the clinician reviews only the summary the agent chose to generate. The vendor does make the full transcript available for cross checking against the generated history, which is a real mitigation, but review after the fact is a different control from review before the output is relied on.

The clinical assistant also produces patient handouts and plain language explanations, which reach the patient directly.

Markets are Canada and the United States, with regional data residency. No United Kingdom or European Union presence is stated. The software as a medical device treatment in those markets would attach to the intake agent even more clearly than to the scribe.

Language access is engineered for a real jurisdictional requirement rather than marketed as a feature: the product documents in both English and French, which is a practical necessity in bilingual provinces rather than a nice addition, and supports capture across 40 or more languages. Consent is also addressed directly, with published guidance for explaining the tool to patients and a reported refusal rate under 5 percent.

That guidance has a double edge: it supplies clinicians with wording to secure agreement, though the suggested language is at least factually anchored in the vendor's actual residency commitment. No fairness statement, subgroup analysis or performance breakdown by language or accent is published.

Nothing the vendor itself publishes creates a route for the recorded person.

Nothing stands behind the output. No warranty, indemnity or remediation commitment is published, and no route is described for a patient who finds an error in a note as distinct from a practice raising it.

The contrast is between the precision of the data handling account and the looseness of the performance claims. Retention, training use and residency are each stated specifically enough to be contradicted. Accuracy is claimed at 99.9 percent with no methodology, reference standard, denominator or date, which cannot be tested. Template counts appear as more than 2,000 on the vendor's own site and more than 200 in a provincial procurement profile. Encryption is given by named cipher on the product pages and as bank level encryption on the security page.

The statutory position is real. An individual in Canada holds an enforceable right to have inaccurate information about them corrected, with regulators behind both the federal and provincial regimes. That right runs against the custodian of the record, which is the practice, and the vendor has published no undertaking to support it.

Broad coverage that is unusually honest about which parts are deep. Native Smart Pull and Push integrations exist for Accuro, Oscar Pro and Jane App. For Epic, Cerner, athenahealth, MYLE, CHR and any other web based record, a Chrome extension pastes the note into the active field, and a desktop application covers anything else through keyboard shortcuts. Telehealth capture works across Zoom, Teams, Google Meet and Doxy.

The vendor states plainly which systems get true bidirectional integration and which get assisted paste, rather than listing every logo as an integration, and every integration is included in the published price with no per connection fee.

Dual jurisdiction residency is stated plainly, and it is the reason to look at this vendor if the purchase is Canadian. Canadian data is stated to remain in Canada on region-specific cloud infrastructure, and the vendor makes the point itself that United States vendors running on major cloud platforms commonly hold Canadian data in United States regions unless configured otherwise.

The vendor's own surfaces differ on one point. The product pages describe regional residency as enforced rather than configurable. The trust center's own answer is softer on one half of it: a Canadian user's data stays in Canada, while a United States user's data is described as able to be routed to United States servers. Able to be routed states a capability rather than a commitment, and the two sentences are not the same promise. The Canadian half is unambiguous; the United States half is not.

Audio is not created or stored at all, which removes the largest residency surface rather than relocating it, and transcription is described as running on the vendor's own self-hosted service. Note text is held temporarily and encrypted only to synchronize a user's devices.

Available across mobile, desktop and a browser extension.

Published, simple and inclusive. The Pro tier is $99 per provider per month for unlimited use, stated on the site rather than behind a quote, and the vendor is explicit that every integration is included with no per feature or per connection add ons, which is where competitors commonly recover margin. A free tier offers three days unlimited then roughly 20 uses a month, which is an evaluation allowance rather than a permanent free plan and is described as such by third parties. An enterprise tier at around $79 per user carries a five provider minimum, which rules out solo and two clinician practices and is worth knowing before comparing headline rates.

Broad coverage, and specific about it. More than 90 specialties are claimed with named coverage across family medicine, internal medicine, psychiatry, obstetrics and gynecology, surgery, pediatrics, nursing and dentistry. Capture spans 40 or more languages with full English and French documentation. Templates can be created by uploading a sample note for the system to copy, which suits specialties whose formats are idiosyncratic. Most distinctively it pre populates Canadian statutory paperwork from encounter data, including CRA Disability Tax Credit forms, provincial referral letters and insurance documents.