Empathia AI

Ambient capture, note generation, coding and patient material generation are the whole product, with no services layer or platform business underneath. The company describes a weekly build cycle driven by clinician feature requests, which is a development posture rather than a moat, but it is consistent with a product company rather than a distributor.

Conventional draft and review, with one structural feature that strengthens it: because verbatim dictation is offered alongside AI assisted modes, a clinician can choose a path where the system does not interpret at all for content they want recorded exactly. Notes are finalized by the clinician before use, and live note preview lets them see output forming during the encounter rather than only afterward. No acceptance rate, edit burden figure or confidence threshold is published.

The vendor's data processing agreement names its generative AI suppliers in Annex III: Microsoft Azure providing generative AI services from Montreal, and OpenAI providing generative AI services from the United States. So the providers behind the note are named.

Version and update discipline is missing. Neither the model family nor the version behind either provider is named, nothing says which provider serves a given feature or instance, and nothing published tells a customer when a model underneath the product has changed. No model card, accuracy figure or evaluation methodology is published, and the headline claim of reducing documentation time by over 80 percent carries no denominator or method.

One technical disclosure stands out. Dictation is offered in separate Verbatim and AI Assisted modes, which tells a clinician plainly when the system is transcribing what was said and when it is interpreting it. That is the difference between a record and a summary, and the product surfaces it as a choice.

The model providers, the hosting arrangement and the subprocessor list are all published, in Annex III of the Data Processing Agreement at a public address, each party with its function and its location.

The list names Amazon Web Services for cloud hosting in Oregon and Quebec, Microsoft Azure providing generative AI services from Montreal, OpenAI providing generative AI services from the United States, Mixpanel for user behavior analytics, Sumo Logic for monitoring and logging, Atlassian for bug tracking, Google as identity provider, and Freshworks for customer relationship management. The model layer is named explicitly and split across two providers in two countries.

The list also includes the operational parties. Monitoring, logging and bug tracking services routinely ingest application logs, and application logs in clinical systems frequently carry identifiers or fragments of clinical text unless deliberately scrubbed, so their presence on the list puts them in view.

The governing terms are published alongside. Subprocessors are bound by the agreement, the vendor remains liable for their failures, changes require thirty days' advance notice with a right to object, and personal data is returned or deleted within thirty days of termination. On training, the vendor states it does not use original customer audio recordings or transcripts for model training, and that synthesized data may be used to train and evaluate its systems.

Two residuals remain. The model family and version behind each generative provider are not named, which is a question for what is under the hood rather than for who holds the data. And the boundary around learning for each clinician is still unstated: the product adapts to each clinician's editing patterns within roughly five encounters, and nothing says whether that adaptation stays inside the clinician's own tenant or informs anything shared.

External validation exists, and it comes in three different kinds that are easy to run together.

The strongest item is institutional vetting. Empathia states it is a preferred vendor across multiple Canadian provincial programs and with the national health infrastructure agency. That is an assessment carried out by a public body with no commercial interest in the outcome, a kind of evidence with no United States equivalent.

Named site evidence sits below that. An Ontario family health team reports charting time down by around two hours daily, and several named physicians appear in published case studies. Named sites and named clinicians are meaningfully better than anonymous testimonials, because a reader could in principle contact them, but these are vendor collected accounts rather than measurements taken by anyone independent.

The third item is not evidence at all. The American Academy of Family Physicians offers this product as a member benefit. That is an explicitly commercial discount relationship, and a professional body's willingness to negotiate a rate for its members says nothing about whether the product performs. It reads as endorsement and is not one.

There is no study, controlled comparison or published accuracy benchmark. What is evidenced is institutional vetting and reported operational benefit, not measured clinical performance.

The controls are real and what they establish sharpens the central question rather than closing it.

Retention is configurable per artifact and stated with precision: encounter recordings from immediate deletion after scribing to a maximum of 365 days, transcripts on the same range, generated notes and summaries from one day to seven years, and user profile data removed within 30 days of a subscription ending. An instant deletion option exists. That is a full schedule, published per artifact.

The default is where the caution sits. A health authority's own deployment guidance for this product tells clinicians to toggle instant deletion on and to select the shortest retention option available, which indicates the minimal setting is a choice the customer must make rather than the one they receive. The vendor separately states notes are stored for one year by default. For a product recording encounters of up to two hours, the default matters more than the option, because most buyers never change a default.

On training the position is partly stated and the important half is not. The vendor states it does not use original customer audio recordings or transcripts for model training or any external purpose, and that synthesized data may be used to train and evaluate its systems. Separately, the product adapts to each clinician over time, learning editing patterns, tone and formatting within roughly five encounters. Adaptation of that kind is use of customer derived content by definition, and the commitment about original recordings and transcripts does not reach it. What is not stated is the boundary: whether that learning is confined to the individual clinician's own tenant, or whether edits and corrections inform anything shared across customers. Those are entirely different propositions for a covered entity.

De-identification for quality assurance is stated to follow the Safe Harbor standard with re-identification prohibited by policy, which is a specific and checkable commitment.

Two deployments exist, and the vendor's guidance states that data remains in Canada only if the Canadian version is used, while its United States material states storage in the United States. Which instance a given account runs on decides which of the two applies.

The United States compliance claim is made repeatedly and is not supported by an instrument.

What is claimed is broad. The vendor describes itself as fully compliant with HIPAA, PIPEDA and PHIPA, states that all operations are conducted in strict compliance with HIPAA, and describes itself as one of the few scribes compliant across both national regimes, with separate integrations for each market.

What is published is a Data Processing Agreement, and it is a substantial document: it names Empathia AI Inc. as Processor and the customer as Controller, binds sub-processors with liability retained under Article 28(4), sets breach notification, audit rights on fifteen days' notice, and return or deletion of personal data within thirty days of termination. It names health-related data as a category of sensitive data transferred.

It is also, expressly, a European instrument. Its stated purpose is to comply with the requirements of EU GDPR, it attaches the European Commission's Standard Contractual Clauses, and it identifies the competent supervisory authority by reference to the EU clauses. The vendor separately states that its services are available exclusively to customers within Canada and the United States. So the published agreement governs a regime the vendor says it does not serve, and the regime it does claim has no published instrument.

The vendor makes no statement that it acts as a business associate, and no business associate agreement, template or scope statement is published. The nearest thing runs the other way: the vendor states that it requires all of its own third-party service providers to sign business associate agreements. That is the downstream half of the chain, and it counts in the vendor's favor, but it is not the vendor telling a covered entity what it will sign with them.

The rest of the vendor's published material is strong, and none of it fills this gap. SOC 2 Type II and ISO 27001 are held, a trust center is published, retention is stated per artifact type, and a named sub-processor list exists. The gap is one specific instrument, not a question of the vendor's seriousness.

SOC 2 Type II and ISO 27001 are both held, and both are independent examinations rather than self descriptions. The information security management system is stated to comply with ISO/IEC 27001:2022, and independent third party risk assessments are described as running at least annually alongside penetration testing and vulnerability management.

A trust center is published at trust.empathia.ai, linked from the compliance page and from the footer of every page, and the vendor directs buyers there for current attestations and documentation.

Also published: hosting on named clouds with the Canadian offering in Montreal, encryption in transit at TLS 1.2 or above and at rest under AES-256 with keys in a managed service, multifactor authentication, role based access with logged administrative actions, and deidentification for quality assurance against the Safe Harbor standard.

The trust center may hold dated, scoped, retrievable reports. Neither certification claim on the public pages carries an audit period, and attestations of this kind expire. And scope is a live question here, because the vendor runs three cloud providers across two national deployments, and nothing public says which systems and which jurisdiction a given report covers.

No clearance claimed and none required for ambient documentation. The clinician reviews and finalizes the note, and the vendor states that timestamps and audit trails accompany every encounter to support documentation integrity.

What governs varies by market and the vendor operates in two. In the United States no device pathway attaches. In Canada the obligations come from professional regulation rather than a device authority, and the vendor references the national physician mutual defense organization's guidance and provincial College documentation standards, alongside a recommendation that clinicians inform patients an AI scribe is in use. Referencing the regime that actually applies is better than silence, though the engagement is a line in a frequently asked questions page rather than published guidance or consent wording.

Two boundaries sit around the documentation core. The product suggests diagnostic and procedural codes, described as helping avoid missed billing opportunities, which is the reimbursement boundary and the phrase is doing work: a system oriented toward opportunities missed is pointed in one direction.

The second is patient facing. The product prepares patient education handouts in real time. Material generated by a model and handed to a patient is a different object from a note a clinician signs, because the reasoning that keeps clinical decision support outside device regulation depends on a professional being able to review the basis, and a patient is not a professional. Nothing published says whether a clinician reviews each handout before it is given, or what sources the educational content draws on.

Markets are Canada and the United States. No United Kingdom or European Union presence is described, and in those markets ambient documentation is treated as software as a medical device.

Engineers for the language access failure mode rather than disclaiming it. Multi speaker recording is built to distinguish clinician, patient and caregiver, and the vendor names the specific contexts it is for, including pediatrics, geriatrics and mental health visits where a family member or INTERPRETER is actively participating, so speech is attributed rather than merged.

Transcription runs across more than 30 languages and the product generates bilingual patient handouts, extending language access past the note to what the patient takes home. No fairness statement, subgroup analysis or performance breakdown by language or accent is published, so the design intent is stated rather than demonstrated.

Dictation is offered in separate verbatim and assisted modes, so the clinician is told plainly when the system is transcribing what was said and when it is interpreting it. That is the difference between a record and a summary, and knowing which one you are reviewing determines how you review it. It tells the person signing where the machine's judgment enters. Configurable retention and an instant deletion option add further customer control.

What is absent is any commitment about correctness. The headline claim of reducing documentation time by more than 80 percent carries no denominator or method, no accuracy figure or evaluation is published, and no warranty, indemnity or remediation commitment is stated. No route is described for a patient who finds an error.

The retention default matters. A health authority's own guidance for this product instructs clinicians to switch instant deletion on and choose the shortest retention available, which indicates the protective setting is a choice the customer must make rather than the default they receive. For a product recording encounters of up to two hours, the default is very likely what most practices will live with.

The product sells into Canada as well as the United States, and Canadian law gives the recorded individual an enforceable correction right against the holder of the record. Which side of that a practice sits on depends on where a given deployment processes data; a United States practice does not inherit the Canadian floor by buying from a vendor that also sells in Canada.

Nothing published states the retention defaults, which regional instance an account runs on, which country's law governs the contract, or what the vendor undertakes when a note is wrong.

Integration is claimed as tailored per market, with separate EMR connections maintained for United States and Canadian systems, which is more deliberate than a single integration list stretched across both.

A named list is published: Accuro, OSCAR, Juno, Plexia, Elation, Epic, athenahealth, eClinicalWorks, NextGen, DrChrono, Practice Fusion, Office Ally and TouchWorks. The Canadian names at the front of that list, Accuro, OSCAR, Juno and Plexia, are the meaningful part.

The list states which systems are named and not what the connection does in any of them. No integration architecture, write back mechanism, or certification program membership is published, so nothing says whether a given name means an embedded application, a note delivered into the chart, or a copy step. Breadth across two national markets is evidenced; depth in any one system is not.

Partial, with one useful property. The product records encounters up to two hours even on low internet connectivity, which matters in rural and remote practice where most cloud dependent scribes simply fail, and is a deployment characteristic rather than a marketing claim.

Against that, no hosting region, residency option or subprocessor detail is published, and residency is a live question for a vendor claiming Canadian provincial compliance, where data location expectations are explicit.

Partial. A free trial is offered, plans are described as flexible and affordable with selectable add ons, and a discounted rate is available through American Academy of Family Physicians membership, so a buyer knows the shape of the offer and that a discount route exists. No rate card, tier structure or per clinician figure is published, so none of it can be budgeted from.

Coverage is enumerated rather than claimed, across three dimensions.

Specialty pages span psychiatry, pediatrics, oncology with staging and multi modal treatment terminology, ENT including procedure documentation, allergy and immunotherapy, emergency with discharge summaries, and chronic disease with multi problem visits. These are described with the vocabulary each specialty actually uses rather than as a list of names, which is the difference between a product configured for a specialty and a product that accepts its patients.

Clinician coverage extends well beyond physicians, to nurse practitioners, physician assistants, physiotherapists, occupational therapists, dietitians and social workers. Allied health is built in rather than treated as an afterthought.

Encounter coverage is equally wide: in person, a built in telehealth capability handling both phone and video with an automatic note, and recording from third party platforms including Zoom and Doximity. Phone is the notable inclusion, since a great deal of real clinical work happens on a call that most ambient products cannot capture. Notes can be organized by problem rather than by visit, which suits longitudinal chronic care rather than episodic documentation, and transcription runs across more than 30 languages.

What is missing is validation rather than breadth. Nothing published shows performance by specialty, and a product documenting oncology staging and psychiatric assessment with one engine is doing two quite different jobs.