Sunoh.ai

The artifact is wholly model generated: without ambient speech recognition and summarization there is no note, no dialogue flow and no pre filled order. Graded on the same test applied to Abridge. Worth separating from the Oracle Health case in this category, which is graded C: Sunoh is a separately branded product carrying its own per provider price rather than a feature folded into an EHR contract. What Sunoh inherits from its parent is distribution, not the intelligence, and distribution is assessed on the interoperability axis rather than here.

The human gate is real and clearly described: output stays a draft until the clinician edits and imports it into the Progress Note. The grade is C anyway because of reach. Sunoh claims to be the first ambient scribe to facilitate order entry and pre fills medications, laboratory tests, imaging and referrals from overheard conversation, while publishing no accuracy figure at all, let alone one specific to orders.

Drafting a medication order is a clinical action rather than a text suggestion, so the disclosure burden is higher, not lower. Same call and same reasoning as Oracle Health Clinical AI Agent, which is the useful comparison here.

No accuracy figure, no model card, no named models and no evaluation methodology published. The gap is conspicuous because competitors in the same price band do publish a transcription accuracy rate. A second candour flag: the vendor's own pages carry inconsistent adoption claims, variously citing more than 50,000, more than 80,000 and more than 100,000 physicians, with no denominator or date attached to any of them. A buyer cannot tell from published material which figure is current or what it counts.

Nothing published identifies any party in the chain. No foundation model provider, model class or version is named, no hosting or cloud arrangement is disclosed, no sub processor list was located, and no evaluation or architecture description exists to reason from. The gap matters more than usual because of where this product sits.

The engine is distributed largely through record system partners and appears inside other vendors' workflows, so it is itself a link in somebody else's chain, and a health system buying an ambient capability from its record vendor may not realise a separate company processes the encounter at all. A buyer in that position should establish the identity of every party in the path rather than assuming the record vendor is the only one.

No explicit statement was located on whether customer audio or transcripts are used to train models, which is the commitment that would at least bound the chain even without naming it. Ask for the sub processor list, the training position in writing, and a data flow diagram showing every entity that touches encounter audio.

Evidence is customer testimonial and vendor case study, for example a named clinic reported saving over two hours daily on documentation. No peer reviewed publication, no controlled comparison, no independent evaluation and no published accuracy benchmark located. Graded C on the standing precedent that scale of use and satisfaction anecdotes do not substitute for evidence of benefit. This sits in stark contrast to the same category's randomized trial evidence for Nabla and Microsoft Dragon Copilot.

The vendor states it signs business associate agreements and applies HIPAA administrative, physical and technical safeguards with industry standard encryption, and third party reviews report a short audio retention window of roughly seven days. That retention figure was reported by comparison sites rather than confirmed on Sunoh's own materials in this pass, so treat it as unverified and ask for it in writing. No explicit statement was located on whether customer audio or transcripts are used to train models, which is the decisive commitment on this axis.

States plainly that it executes business associate agreements and follows HIPAA safeguards, and it is sold self serve with a free trial, which implies a standard rather than individually negotiated agreement. The specific terms are not published for inspection before contracting.

A second pass again failed to reach this vendor's own security documentation, and the reason is worth recording because it shapes how the claim should be read.

A SOC 2 attestation is asserted for this product in several places. All of them are third party comparison pages, and several are published by direct competitors. Those same pages also state that the vendor publishes a dedicated security overview covering encryption, audio retention, data residency, use of data for training, and data sharing. If that page exists it would answer much of this axis and several others.

This index grades from evidence retrieved from the vendor, and neither the attestation nor the security overview was confirmed on the vendor's own materials in either pass. So the grade records what a counterparty can verify directly, which remains thin.

One distinction is worth drawing rather than treating all third party assertions alike. An aggregator repeating a vendor's marketing carries little weight, because such sites copy one another. A competitor's comparison page conceding that a rival holds a certification runs against that publisher's interest, which makes it a different and somewhat better class of evidence. It is still not the vendor evidencing its own posture, and a buyer should not accept a rival's characterisation of a supplier's controls in either direction.

This remains a refresh candidate with a specific next step rather than a general one: locate the vendor's own security overview, confirm the report type and period, and check whether the disclosures on audio retention and training use are stated by the vendor rather than only described by others.

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 provider is described as editing the draft before importing it.

Two features sit beyond narrative documentation. The product is described as drafting a note together with suggested labs and orders, and as pre filling orders to speed sign off. Order suggestion is a step past recording an encounter, because an order is an action taken on a patient rather than a description of one, and pre filling is a further step again since it places the proposed action in front of the clinician already formed. Establish what the clinician confirms for an order specifically, as distinct from what they confirm for the note.

The second is cross language generation, and here it is the headline feature rather than an edge case. The vendor's positioning is that clinician and patient converse in the patient's preferred language while the product produces an English transcript and note. That is a translation step embedded in the documentation pipeline, and it changes what the signature attests to: the signed record is not in the language the encounter took place in, and where the clinician does not speak that language they cannot check the translation they are signing.

One peer in this category publishes an explicit caveat on exactly this, stating that its product cannot act as an interpreter and that a native speaker must corroborate the note. No comparable caveat was located here, and for a vendor making multilingual capability its central claim the absence is more conspicuous than it would be for a vendor offering it as a minor feature.

Markets appear to be United States only.

The absence here sits directly underneath the vendor's own headline claim, which is what makes it the sharpest instance of this pattern in the category. No fairness statement, no subgroup analysis, no accuracy breakdown by language, accent or dialect, and no evaluation methodology were located.

Multilingual capability is not a secondary feature on this record. It is the central differentiator, marketed explicitly at the multicultural United States patient population, with the product positioned so that clinician and patient can speak in the patient's preferred language while an English note is produced.

That claim carries an evidentiary burden the vendor has not met, and the second pass makes the gap wider rather than narrower in two ways.

First, the language list itself is not published. A third party review notes that language coverage is not disclosed and advises prospective buyers to ask before purchasing. A vendor whose principal claim is breadth of language support, and which does not say which languages, has left the buyer unable to check the one thing the product is being sold on.

Second, translation quality is the actual risk and it is unaddressed. Speech recognition degrades with accent and dialect, and a cross language pipeline adds a translation step on top of that, so errors can compound rather than simply appear. The population most exposed is the one the product is explicitly marketed to serve, and nothing in the workflow surfaces a degraded result to the clinician signing the note.

The methods exist. Peers have submitted note quality to published evaluation against validated documentation instruments, and one publishes an explicit limitation on cross language use.

No accuracy or error figure, no published limitations, no warranty, indemnity or remediation commitment was located in two passes, in a category where several competitors do publish a transcription accuracy rate. What replaces measurement here is a candour problem that bears directly on reliance.

The vendor's own pages carry inconsistent adoption claims, variously citing more than 50,000, more than 80,000 and more than 100,000 physicians, with no date and no denominator attached to any of them. A buyer cannot tell which figure is current or what it counts, and published numbers that contradict each other are a poor foundation for any representation a health system might rely on. The distribution model compounds the position.

This engine reaches many buyers embedded inside a partner's record system, so the contract, the support path and the commercial relationship frequently run to that partner rather than to this vendor, which means the party a customer can hold to anything is not the party that built the model.

A retention window of roughly seven days for audio is reported by comparison sites rather than confirmed on the vendor's own materials, so it is not relied on here and is worth obtaining in writing alongside the rest. Ask which entity carries the performance representations in your particular purchase, and ask for an accuracy figure with a definition attached.

Deep and native inside eClinicalWorks: desktop EHR, eClinicalTouch for iPad and eClinicalMobile, with structured import directly into the Progress Note and order pre fill on mobile rather than copy and paste. Outside eCW the picture is weaker than the marketing implies. Sunoh brands itself EHR agnostic and names Epic, athenahealth and Oracle Health, but third party reviews record that depth varies by site and no integration architecture is documented. Buy it for eClinicalWorks and treat the cross EHR claim as requiring site specific proof.

No hosting region, residency option or subprocessor list was confirmed from the vendor's own materials, and nothing establishes where inference runs or whether a third party model service is involved.

Two structural facts are now clearer and both belong on the record.

The first is corporate. This product is not an independent scribe that happens to integrate well with one electronic record system. It is developed by a company within that record system vendor's own group. That changes how the integration depth should be read, since the deepest connection is to a sibling product rather than to a negotiated partner, and it changes the data question too: a buyer should establish whether encounter content stays within the scribe's own boundary or moves within the wider group, and what the terms permit.

The second is the integration surface, which is split. Native integrations exist with a small number of named record systems, with the strongest being to the sibling platform and one connection made through a modern interoperability interface. Everything outside that list is reached through a browser extension that pushes notes into any browser based record system. That is the fourth vendor in this lane to reach the chart through the browser rather than a purpose built interface, and the same questions apply: whose credentials the extension operates under, what page access it holds, how its writes appear in the audit log, and what breaks when the host system changes.

So a buyer's deployment answer differs sharply by which record system they run. Establish which path applies before assuming the integration described in marketing is the one available to you.

Publishes a price, which alone puts it ahead of most of this category, and offers a free trial with no contract and no credit card. The listed rate is 149 US dollars per provider per month, also expressed as roughly 1.25 US dollars per visit. Held at B rather than A for two stated qualifiers that make the published number incomplete: the rate is labelled promotional or limited time, and eClinicalWorks states that additional monthly charges may apply without specifying them. A published price that is provisional and admits unquantified add ons is better than silence but is not a rate card a buyer can budget from.

The broadest specialty reach of any independent scribe assessed in this index, and it is enumerated rather than implied. Beyond primary care, dermatology, rheumatology and behavioral health with DAP notes, it produces dental periodontal charting and dental treatment plans and covers vision.

Dental and optometry documentation is a genuinely different note structure that almost no competing ambient scribe attempts, and it is the clearest reason a multi specialty or dental group would shortlist Sunoh over a physician focused tool. Multilingual capture is claimed across the range, though without published performance by language.