Aura AI Scribe

Insight Health describes itself as an AI agentic platform and every product in it is model output: the Aura scribe, the Virtual Care Assistant conducting patient interviews, front desk, phone triage and referrals agents. No services layer or platform business underneath.

The scribe itself is conventional draft and review, generating notes and coding suggestions for the clinician to check before use.

Two adjacent agents on the same platform carry materially more autonomy and are not described in oversight terms: a phone triage agent, which triages patients by phone, and the Virtual Care Assistant, which conducts pre visit and post visit patient interviews to capture medical histories unsupervised. No confidence threshold, escalation path or abstention behavior is published for either.

That autonomy belongs to the wider suite rather than to the documentation product.

Some real mechanical description: automatic speech recognition with speaker diarization is named rather than implied, note templates carry section level instructions that steer generation, and the white label product documents structured JSON output and event listeners for partners. But no accuracy figure, model card, named models or evaluation methodology is published, so nothing about performance can be assessed.

The boundary is drawn well and no party is named. On the boundary the published commitments are detailed: a configurable retention period defaulting to twenty one days, destruction performed to a named federal media sanitization standard rather than described as secure deletion, provider ownership of all patient data, the vendor acting solely as a processor, no sharing with third parties, and no use beyond service delivery.

That last phrase carries the training question by construction, since training on customer content is use beyond service delivery, though a plainer formulation would leave less room. On enumeration there is nothing: no model, model family or provider, no hosting arrangement and no subprocessor list, despite useful mechanical description elsewhere including named speech recognition with speaker separation and documented structured output for white label partners.

The white label arrangement is itself a supply chain fact: this engine appears inside other companies' products, so a practice may be using it without knowing the name. One scope question remains open: patient facing agents conduct pre visit history taking, answer inbound calls and triage symptoms, and nothing states whether the retention default and sanitization standard cover recorded speech from a patient who never attended a visit.

Platform vendor validation rather than clinical study, and the distinction matters. The Epic Toolbox designation in the Ambient Voice Recognition Notes category is a technical verification by Epic that the product follows its recommended connection practices, which is a higher bar than a marketplace listing. The athenahealth Marketplace agreement adds a second platform relationship.

Multiple named clinicians are quoted, including one who reports trialing five platforms before selecting it, and an $11 million Series A closed in April 2026. None of that measures documentation quality or clinical outcome: no study, controlled comparison or accuracy benchmark is published, and the vendor claims of two hours saved daily and four to five additional patients weekly carry no denominator.

The stewardship disclosures are largely present, and the open question is breadth.

The vendor publishes a configurable retention period defaulting to twenty one days, destruction performed to the United States federal media sanitization standard, provider ownership of all patient data, the vendor acting solely as a processor, no sharing with third parties, and no use beyond service delivery.

That last phrase carries the training question and largely answers it. Use beyond service delivery is what training on customer content would be, so the commitment reaches it by construction. It is still a use limitation rather than an explicit statement about model training, and a plainer formulation, saying directly that clinical content is never used to train or fine tune models, would leave less room.

Scope is the open question. The platform is not a scribe alone. Patient facing agents conduct pre visit history taking by voice or text, answer inbound calls and triage symptoms, and contact patients after procedures about recovery, symptoms and medication adherence. Each of those produces recorded patient speech outside a clinical encounter, and none of the published retention material states whether it is governed by the same twenty one day default and the same sanitization standard as the scribe.

A retention policy written for encounter audio may or may not cover a recorded telephone triage call from a patient who never attended a visit at all.

The business associate terms can be read before any agreement conversation, and they cover the free tier as well.

Business associate status is stated in the instrument itself: Insight Health AI, Inc. is named as Business Associate, and the recitals state that by providing services and receiving protected health information it shall become a business associate of the customer.

The agreement is available in full at insighthealth.ai/service-baa, dated 17 January 2024, and is incorporated by reference into the terms of service, which state that in the absence of a separately negotiated agreement the standard one at that address governs. A buyer can read what will apply before any sales contact.

The tier it applies at is stated explicitly. The Trust and Safety page states that the standard business associate agreement is available to all users across all payment tiers, so the free tier is inside the arrangement rather than outside it.

The agreement expressly extends its obligations to subcontractors. Under Disclosure to Third Parties it requires a written agreement with each subcontractor or agent having access to protected health information, binding them to the same types of restrictions, terms and conditions. The vendor separately states that it holds business associate agreements with all of its own vendors.

The instrument is substantive rather than a stub: minimum necessary limits, safeguards under Subpart C of 45 C.F.R. Part 164, breach reporting within ten business days with a defined content list, improper use reporting within fifteen, ten business day turnarounds on individual access, amendment and accounting requests with five business days to forward a direct individual request, and return or destruction on termination extending to subcontractors' holdings.

The agreement attaches through an Order Form, so it depends on one existing for the plan being bought. It also applies only to the extent the customer is a covered entity, a limit that bears on any customer that is itself a business associate.

SOC 2 TYPE II is named and typed. The Epic Toolbox designation adds an independent technical review of connection practices, which is a different kind of assurance but a real one. Neither the audit date nor the report scope is given.

No clearance claimed and none required for the documentation product. No United States device pathway attaches to a note the clinician reviews and signs.

The regulatory question is larger than a single feature. The platform runs patient facing agents across the whole care episode. One conducts pre visit history taking by voice or text, described by the vendor as capturing critical information just as a physician or physician assistant would. One answers inbound calls, triages symptoms and determines whether the issue is urgent. One contacts patients after procedures at scheduled intervals, asks about symptoms and medication adherence, and flags concerning reports for clinical review.

Determining urgency from reported symptoms is a clinical judgment, not an administrative one, and it is being made in a conversation with a patient where no professional is present. The reasoning that keeps clinical decision support outside device regulation depends on a professional being able to review the basis of an output before relying on it. In a triage call the output is acted on immediately, by the patient, on the basis of what the agent told them.

The vendor's oversight position stands out because it is stated rather than implied. Its leadership describes a safety net for escalation and routing and says a clinician is always in the loop. The questions that follow are the useful ones: what triggers escalation, how a false negative would be detected, and what a patient is told when the agent is uncertain.

The scribe also suggests diagnostic and procedural codes, framed as optimizing reimbursement while maintaining compliance, a mild form of coding automation.

No fairness statement, no subgroup analysis, no accent or dialect performance disclosure and no language coverage claim is published.

For this platform the absence is not principally a documentation quality question. The reasoning needs stating in full because it changes the stakes.

In a conventional ambient scribe, speech recognition failure produces a worse note. The clinician was present at the encounter, reads the draft, and can correct what the system misheard. The error is contained by the person who witnessed the conversation.

Here the platform's agents speak to patients directly and alone. They take histories before a visit, answer inbound calls and triage symptoms, and follow up after procedures asking about recovery and medication adherence. No clinician is present. The system's understanding of what the patient said is not checked against anyone else's memory of the conversation, and in the triage case it feeds an urgency determination acted on immediately.

So performance across accents, dialects, and speech affected by illness, distress, age or hearing difficulty stops being a quality issue and becomes a safety one. The vendor's own material observes that patients calling a practice are often anxious or unwell. That is precisely the speech condition least likely to be well represented in training data, and it is the population the agent is designed to serve.

The disclosure that would matter here is therefore narrower and more urgent. Not overall note accuracy, but how reliably the agents understand patients across the range of people who actually call, and what happens when they do not.

A mechanism addressed to output correctness is published.

The vendor publishes what it calls a Supervised Autonomous Framework for Evaluations: a supervisory model that continuously monitors output for medical terminology accuracy, consistency with the source audio, and potential hallucinations. Consistency with audio is source traceability in substance, checking the generated note against what was actually said. Combined with clinician review before signing, mechanisms exist to catch a wrong output before it becomes a signed record.

Nothing stands behind the output. No accuracy or error figure, no false negative rate for the supervisory layer, no published limitation set and no remediation obligation is published. A framework that detects hallucinations without a published detection rate is a control whose effectiveness is unstated. The terms of service contain indemnification and limitation of liability provisions, which the business associate agreement cross references and states will govern performance under it; their content may change the picture in either direction.

What makes the remaining absence weigh more than for a scribe alone is scope. The platform runs patient facing agents that take a pre visit history by voice or text, answer inbound calls and triage symptoms, and contact patients after procedures about recovery, symptoms and medication adherence. A triage interaction is a judgment about urgency made without a clinician present, and a post procedure check about symptoms and adherence is a clinical conversation with a patient at home. The supervisory framework is described for the scribe's documentation output; nothing establishes that it governs the triage agent, which is the surface where a wrong output reaches a patient directly with no review step.

The white label distribution compounds it, because where this engine sits inside another company's product the practice may not know which entity to ask.

The integration position is verified by the platform vendors themselves rather than asserted. The Epic Toolbox designation for Ambient Voice Recognition Notes means Epic has confirmed the product follows its recommended connection practices, and the scribe launches directly from Haiku with the encounter captured in context. It is listed on the athenahealth Marketplace and names eClinicalWorks and AdvancedMD alongside. The Chrome extension is designed to stay inside the same browser window as a web based record rather than forcing application switching, which is a small but real workflow difference.

Aura Anywhere exposes the whole capability to EHR vendors as a developer product, returning structured JSON mapped to the partner's own data model with event hooks and staging infrastructure. That is an integration surface built for embedding rather than for connecting, and it is a different kind of depth from a marketplace listing: the partner's product becomes the delivery vehicle. The vendor markets partnership routes to EHR vendors, transcription and scribe companies and private equity firms, which indicates the white label channel is a primary distribution strategy rather than a side offering.

The Epic designation and the marketplace listing are third party verifiable. The designation confirms connection practice rather than depth of write back: the product connects to named systems in ways those vendors have confirmed, and how deeply each writes is not shown.

Specificity is the strength of these deployment terms.

The vendor publishes a default retention period of twenty one days, describes it as configurable, and names the standard used to destroy data at the end of it: the United States federal guideline for media sanitization. Naming a period a buyer can hold them to, and naming the standard governing how destruction is performed, is materially more useful than a bare statement that data is deleted.

Alongside it the vendor states that the healthcare provider retains ownership of all patient data, that it acts solely as a processor, that there is no sharing with third parties, and no use beyond service delivery.

Three things are missing. No hosting region is named and no residency option is offered. No subprocessor list is published, and nothing establishes whether a third party model service processes the encounter or what it retains, which matters because a retention commitment binds the vendor and not necessarily its suppliers.

The vendor also sells the scribe as an embeddable component that other software companies install into their own products, delivered through a package manager with the host's branding and single sign on. So a clinician may be using this vendor's scribe inside another company's application without that being apparent, and the retention and processing terms that govern it will sit in an agreement they never saw.

A complete four tier ladder is documented: a free forever plan, Pro at $39.99 per clinician per month or 480 per year, Team at 79.99 per month or 960 per year, and a custom enterprise tier, with monthly and annual rates both stated. Coding suggestions sit in the Pro tier rather than the free one, so the free plan does not include them.

The gap is provenance rather than completeness. The full ladder comes from a third party software directory rather than the vendor's own pricing page. A directory can be out of date, and a price a vendor has not published itself is a price it has not committed to.

Outpatient focused, consistent with the Epic Ambient Voice Recognition Notes category it is designated in. Specialty fit is achieved through customizable note formats definable per specialty, per practice or per individual provider, with section level instructions steering generation, which is a more granular customization model than a fixed template library. Named users span physicians and advanced practice providers across several specialties. No specialty count or language coverage is published, and no inpatient or emergency capability is described.