Tali AI

Ambient generation, dictation and clinical search are all model driven and there is no non AI business underneath. Worth watching rather than downgrading: Tali now describes 14 connected capabilities spanning pre visit preparation, documentation, forms, billing and outbound patient calls, which is platform expansion. If the commercial center of gravity shifts to workflow breadth, the moat starts to look like the Canadian forms and billing content rather than the model.

Explicit and well described review gate: the note is generated inside Tali, the clinician reviews and edits it there, and only then chooses to insert it into the patient chart as a deliberate action rather than an automatic write. Separating generation from insertion keeps the chart untouched until the clinician acts. No acceptance rate, edit burden figure or confidence threshold is published.

Technical description is generic, citing natural language processing, medical models and speech recognition without naming any of them, and the performance claims offered are speed rather than accuracy, notably documentation three times faster than typing.

One claim deserves substantiation rather than repetition: Tali states its notes were ranked in the top 1 percent for accuracy and quality by the United States Veterans Affairs, which would be a meaningful independent result but is presented without a date, methodology, comparison set or published source.

The architecture is described and no provider is named. The vendor says its language model is trained on medical resources and fine tuned with more than 20,000 accented recordings, and its terms state that the service uses third party providers which the company may change at any time, disclaiming liability for their failures. So the chain has external parties in it, and they are unnamed and swappable.

The training position is clearer than the chain. The FAQ states that the company does not and never will use protected health information to train its AI, and the terms grant the company a perpetual license to customer data on a de identified basis to train its algorithms, excluding personal and personal health information. That is a real commitment with a real carve out: de identified customer content does train the company's own models.

Location is stated by user country. Application data is stored in Canada for Canadian users and in the United States for others, and ambient audio is deleted once the transcript is created. The terms also permit personnel access for support from locations around the world.

No peer reviewed publication, controlled evaluation or independent clinical study is published. The two external reference points are a procurement vetting rather than clinical evidence, being prequalification under the Canada Health Infoway AI Scribe Program, and the unsubstantiated Veterans Affairs accuracy ranking noted above. Physician testimonials are named and specific but remain testimonials.

Handling is specific and training is answered, with one carve out. Data is encrypted in transit and at rest, audio from the ambient scribe is deleted as soon as the transcript is created, and application data is stored in Canada for Canadian users and in the United States for others.

On training, the FAQ states the company does not and never will use protected health information to train its AI. The terms add the carve out: customers grant a perpetual license to their data on a deidentified basis for training the company's algorithms, with personal and personal health information excluded. So deidentified customer content does train the company's own models, under a contractual boundary rather than a help page.

Retention beyond audio and the method behind deidentification are the gaps. No retention period is stated for transcripts or notes, and nothing describes how content is deidentified before it enters training.

HIPAA compliance is claimed and the instrument that makes it real is not published.

Tali is a Canadian company whose home regime is documented with care. Its security page lists compliance with PHIPA, PIPEDA, PIPA and Alberta's HIA and certification under Quebec's TGV framework, and its terms treat the company as a service provider using personal information solely to provide the service. It also sells into the United States, with integrations to Healthie, Net Health, Practice Fusion and Veradigm and United States data storage for users outside Canada, and it lists HIPAA among its compliance claims. Once a vendor invokes HIPAA, what a United States buyer can rely on is the business associate agreement, and none is published. The terms of use, governed by Ontario law, contain no business associate provisions, and neither the FAQ nor the security page states business associate status.

For a Canadian buyer the privacy posture is substantially better, since the home regime is the more fully documented half of the company's privacy material.

SOC 2 Type II is stated consistently on the vendor's own security page and in its terms, and the company also states certification under Quebec's TGV framework for digital health technologies. What is not published is the artifact. The terms say copies of security reports are provided on written request and subject to confidentiality, and the security page offers the privacy impact assessment and penetration test reports on request. No trust portal that opens without a request is published.

One wording point is unsettled. The terms say Tali provides its services on a SOC 2 Type II certified platform, which could describe the company's own examination or the hosting platform beneath it, while the security page says Tali AI itself is SOC 2 Type 2 certified, and nothing says which entity the report names.

No clearance claimed and none required for ambient documentation. No United States device pathway attaches to a note the clinician reviews and signs, and the vendor states the control point plainly in its own help material: the clinician is responsible for the final patient record and should review and edit before it reaches the chart.

The scope delta here is the product set around the scribe. Alongside dictation and ambient documentation the company ships a medical search function and an assistant that works against the record, and its site navigation now markets a clinical decision support module. Medical search and decision support are clinical reference rather than documentation. The terms scope this in writing: the services are not intended to be the basis for treatment decisions, and the customer must review and approve every output for clinical use. How the decision support module squares with that disclaimer is not stated.

The language support matters for a different reason: it changes the artifact rather than the classification. The product supports French and Spanish alongside English, and documents Persian as a beta input language that produces English clinical notes. Cross language generation places a translation step inside the documentation pipeline. That is not a device question in the United States, but it does change what the clinician is attesting to when they sign, because the signed record is not in the language the encounter took place in.

Markets are Canada and the United States. The company describes no United Kingdom or European Union presence. Ambient documentation is treated as software as a medical device in those markets, so entry there would require reassessing this position.

Tali publishes no fairness statement, no subgroup analysis, no accuracy breakdown by accent, dialect, language or speaker population, and no evaluation methodology.

What it does publish is a claim. The vendor states that the product is designed to recognize a broad variety of accents and that it supports a wide range of voices and accents. That is an assertion about speaker robustness itself, published with no number attached. The shape is a common one: the efficiency gain is quantified, at a stated halving of documentation time and a per encounter time saving, while the robustness claim that would matter most to an underserved patient is asserted and never measured.

The exposure is larger here than for an English only product. The scribe supports French and Spanish, and Persian as a beta input language that produces English clinical notes. Cross language generation inserts a translation step into the record, and no evaluation of it was published. A clinician documenting in a second language has no way to see where the system degrades, and nothing in the workflow surfaces it.

The company holds SOC 2 Type II and national program pre qualification, yet publishes nothing that characterizes model performance across speaker populations. That work is plainly achievable: note quality can be evaluated against validated documentation quality instruments and through independent comparative benchmarking.

Nothing the vendor publishes commits it to anything when a note is wrong.

One published claim is unsubstantiated. The company states its notes were ranked in the top 1 percent for accuracy and quality by a United States federal health system. If accurate that would be an independent external assessment, but it is presented with no date, methodology, comparison set or published source, so it cannot be weighed.

The terms point the other way on recourse. The service is provided as is with every warranty disclaimed, including accuracy and reliability; liability is capped at twelve months of fees; the vendor's indemnity covers intellectual property infringement only; and the customer releases the vendor from all claims arising from use of the service, having agreed to review and approve every output for clinical use. That review requirement is the one mechanism in place.

The statutory position is real. Canadian federal privacy law gives an individual the right to challenge the accuracy of personal information held about them and to have it corrected, and provincial health information legislation carries equivalent correction rights, with regulators behind both. The right runs against the holder of the record rather than against this vendor, which has published no undertaking to support it.

Effectively complete coverage of one national EMR market, which is a different and in some ways harder achievement than partial coverage of several. Integrations span Accuro, TELUS Collaborative Health Record, Med Access, Oscar Pro, PS Suite, Healthquest and Intrahealth Profile, and the integration pattern is launch inside the record with one click rather than a parallel window and paste. Access is also available on desktop, mobile and Chrome for clinicians whose system is not covered. The depth sits within Canada, and United States integration depth is asserted rather than evidenced.

Residency is stated and the processing path is left open.

The FAQ says application data is stored in Canada for Canadian users and in the United States for users from other countries, and that ambient audio is deleted as soon as the transcript is created. The terms confirm storage in Canada or the United States as applicable, and add two things. Tali personnel may access data for support from locations around the world, which the terms say is not to be treated as a transfer outside Canada or the United States. And the service uses third party providers that the company may change at any time, none of them named, so where processing by those providers runs is not documented. Tenant isolation is not described.

For a Canadian buyer with a residency requirement the storage commitment is real, and the support access clause and the unnamed providers are the two things to put into the agreement.

A complete published price ladder. A free Personal tier includes 5 ambient scribe sessions and 60 minutes of dictation a month, Premium is CA$45 per month billed annually, Pro is CA$135 per month billed annually, and Enterprise is custom, with a 14 day Pro trial. One structural detail matters: unlimited ambient scribing, the capability most clinicians are evaluating, is unavailable below Pro. Premium buys unlimited dictation only, so the real entry point for regular ambient use is CA$135 rather than CA$45.

Deep in Canadian primary care, thin on specialty enumeration. Supports SOAP and DAP note formats, the latter indicating behavioral health workflows, with ICD-10 code suggestions. Its coverage strength is jurisdictional rather than clinical: a fillable forms library auto populating Canadian government, disability and insurance forms, and billing support for Ontario FHO+ time based codes, which is tuned to one provincial primary care funding model. Outbound patient call documentation extends coverage to telephone encounters. No specialty count or language coverage published.