iScribeHealth

Generative ambient documentation is real but it was added to an existing business rather than being the business. iScribeHealth was founded in 2015 around a mobile documentation app, speech to text dictation, transcription and human virtual scribe services, all of which it still sells. Strip out the generative layer and a functioning company remains, which is the moat is not the model pattern this index applies to Conveyor AI and Chartnote in this same category. The continuing human scribe offering also puts it in the same structural family as Speke, where AI and human documentation are sold side by side.

Conventional draft and review, with the review step described explicitly as taking seconds before the provider leaves with their last patient. The dictation path is inherently clinician driven, since the provider speaks into a named note section rather than the system deciding placement. Held at B because no acceptance rate, edit burden figure or confidence threshold is published for the ambient path, where the system does choose what goes where.

No accuracy figure, model card, named models or evaluation methodology located. Capability is described in outcome terms, generative AI and speech to text producing documentation in the moment, with the published numbers measuring time and cost saved rather than correctness.

Nothing identifies any party in the chain: no model or model family, no foundation model provider, no hosting arrangement and no sub processor list was located in two passes. The more important omission here is human rather than technical, and this record is one of several in the lane that makes the point.

Named individuals at the vendor routinely read notes describing real patient encounters, because expert reviewers drawn from the company's original scribe workforce check generated notes as part of the standard process. So the parties receiving encounter content include people, and the enumeration a buyer needs covers who they are, where they sit and under what controls rather than model providers alone.

None of it is published: not whether reviewers are employees or contracted, not their location, not whether their access is logged and time bound, and not whether they see identifiable content or de identified text. A pattern is now visible across this lane and it is worth stating plainly. The vendors that market human review most prominently disclose least about how it is governed. Review is presented as a quality feature, which it is, and treated as though it raises no access question, which it plainly does. Ask who reviews, where, on what content, and with what logging.

Named customer testimony including an operations executive at a large orthopaedic group, and specific cost figures of 15,000 to 25,000 US dollars saved per provider per year through reduced transcription spend. That figure is more concrete than most claims here because it measures a displaced line item rather than a subjective benefit, but it is vendor reported, dated, and describes cost substitution rather than documentation quality or clinical benefit. No study, controlled evaluation or third party assessment located.

No statement on audio or transcript retention, de identification or training use was located. The extra dimension the earlier assessment identified is now confirmed rather than hypothetical, and it should be stated plainly.

The company began as a human scribe business operating around a hundred and twenty scribes. On concluding that AI could produce accurate notes, it did not remove people from the process. Its chief executive has described retaining the most expert of those scribes to review AI generated notes, both for accuracy against what was recorded and to reproduce each physician's style.

So human review is part of the model rather than a legacy service, and that changes what this axis is asking. The question is not only what systems retain, but who reads. Named individuals at the vendor routinely read notes describing real patient encounters. That is a legitimate design which plainly improves output, and it is an access surface a buyer must be able to assess.

The questions are specific. Do reviewers see the audio or only the generated note. Are they employees or contracted. Where are they located. Is their access logged and time bound. Is a patient's identifiable information visible to them, or is review performed on de identified text. Does every note pass through review or only a sample, and can a practice opt out.

This index has now recorded several vendors with people in the loop and a clear pattern: the ones that market human review describe it prominently and disclose least about how it is governed.

Ask who reviews, where, and under what controls.

No published business associate agreement posture was located for this product.

An agreement exists, since the vendor handles recorded clinical encounters and its reviewers read the resulting notes. What is absent is any published statement of scope, and this record has two features that make scope the substantive question rather than a formality.

The first is the review workforce. Where people employed or contracted by the vendor read notes describing patient encounters, the agreement needs to address workforce access explicitly: who qualifies as workforce, what training and confidentiality obligations apply, whether any reviewers are contractors or offshore, and how the vendor evidences compliance for people rather than systems. An agreement drafted for a software product may not reach any of that.

The second is distribution. The company markets alongside a major record system vendor, which raises the question this index has now recorded several times. A practice reaching the product through that route should establish which entity is party to their agreement, whether the record system vendor is a party or a conduit, and who is accountable in a breach originating with either. Do not assume an existing record system agreement extends to a partner application.

One further question follows from the company's history. Where a vendor previously delivered a human scribe service to the same customers, older agreements may still be in force describing that service rather than an AI product, and the two have different data flows.

Ask for the agreement, its workforce provisions, and which entity is party to it.

No named or dated attestation, no report of either type and no trust centre were located.

The human review layer is what makes this absence more consequential than the tier norm, and the argument is the same one this index applied to a peer operating certified coders over every note. A vendor whose model depends on staff reading customer content carries a heavier disclosure burden than one running a fully automated pipeline, not a lighter one. The controls that matter are precisely the ones an independent examination describes: how reviewer access is provisioned and revoked, whether it is logged, whether it is scoped to assigned work or open across the estate, how personnel are screened, and what happens on departure.

None of that is published, and none of it is inferable from a compliance claim.

The company's history sharpens it further. This was a services business before it was a software business, and services businesses accumulate operational access patterns that are appropriate for delivering a service and unexamined as a security posture. The transition from a hundred and twenty scribes to a smaller expert review team is a reduction in surface, and the surface that remains is the one an assessment would look hardest at.

The partnership context raises the expectation too. Distribution alongside a major record system vendor normally involves that vendor's own security review, so an assessment of this company probably exists in someone's hands.

Ask what external testing has been performed, what the reviewer access model is, and whether the record system partner's diligence findings can be shared.

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

The product is narrow in scope. No coding engine, risk adjustment scoring, decision support layer, order generation or patient facing component was located. It produces documentation.

One feature of the design is worth recording on this axis because it bears on the assumption the whole regulatory position rests on. The vendor operates an expert human review step between generation and the clinician, checking notes for accuracy against what was recorded and for fidelity to the physician's style. That is an additional quality control layer that most products in this category do not have, and on its face it strengthens rather than weakens the position: a second party has examined the note before the clinician sees it.

It also introduces a question worth asking. Where a note has already been reviewed by an expert before reaching the clinician, the clinician may reasonably review it less carefully than an unchecked draft. That is a rational response to a real quality improvement, and it means the clinician's own attestation is being made against a document they may scrutinise less. The regulatory position depends on the signing clinician's review, not the vendor's.

So the useful question is whether the workflow makes clear what has been checked and what has not. A note presented as reviewed should indicate what the review covered, so the clinician knows which parts remain theirs to verify.

Ask what reviewers check, and how that is surfaced to the signing clinician.

No fairness statement, subgroup analysis, accent or dialect performance disclosure or language coverage claim was located.

The absence is more striking here than for most small vendors, because this company has the one thing almost nobody in this category possesses: expert human reviewers reading AI generated notes across many physicians and many encounters, continuously.

That workforce is a detection mechanism. People who read machine generated notes all day, against the recordings they came from, are positioned to notice patterns no automated metric surfaces: that transcription degrades for particular speakers, that certain accents produce more corrections, that particular clinical vocabulary is mishandled. They are effectively running a continuous qualitative evaluation of exactly the thing this axis measures.

So the question is not whether the vendor could produce evidence but whether anyone is collecting what its reviewers already see. Establish whether correction patterns are logged and analysed, whether reviewers can flag systematic issues rather than only fix individual notes, and whether any of that has ever been examined by speaker characteristics.

That is a low cost ask. It requires no new evaluation infrastructure, only attention to data the review process already generates, and it would place this vendor ahead of far better funded competitors on the axis this index has found least answered across the entire category.

The converse is the risk. A review layer that silently corrects systematic errors makes them invisible, because the output looks fine while the underlying model continues to fail for the same population.

Ask what the reviewers see, in aggregate, and whether anyone is looking.

Expert human review sits in the pipeline and it is a genuine correction mechanism, which is what places this in the middle band rather than at the floor. The company began as a human scribe business and, on concluding that models could produce accurate notes, retained its most expert scribes to review generated notes against what was recorded and to match each physician's style.

A person checking machine output against the source before it reaches the clinician is the same class of control that earns credit elsewhere in this lane, and it operates before the note is signed, which is the point at which a documentation error becomes a medicolegal fact. What is not established is its scope or its governance.

Nothing states whether every note passes through review or only a sample, whether a practice can opt out, whether reviewers see the audio or only the generated text, or how their access is logged and bounded. A correction step of unknown coverage cannot be relied on as one. Nothing else supports the grade.

No accuracy figure, error rate or evaluation methodology was located, the published numbers measure time and cost saved rather than correctness, and no warranty, indemnity or remediation commitment was found. Ask what proportion of notes are reviewed, by whom, and what the vendor commits to for the ones that are not.

Genuine integration rather than transfer, and unusually granular for a mobile product. Documentation returns directly into the EHR with no copy and paste, and dictated content can be placed into specific note sections including chief complaint, history of present illness, review of systems, physical exam and assessment and plan, which is section level rather than dumping a block of text.

Bidirectional flow between app and record is described as a core design property. athenahealth is a named partner and Veradigm is listed. Held at B because the supported EHR list is not enumerated, depth outside athenahealth was not verified, and the product is iOS only.

No hosting region, residency option or subprocessor detail was located. Delivery is a cloud service with a mobile client, and the vendor names deep record system integration as one of its three stated differentiators, marketing alongside a major ambulatory record platform.

The residency question has a dimension here that most records in this lane do not carry, and it follows from the review workforce rather than from infrastructure. Where people read notes describing patient encounters, the location of those people is a data location question in its own right. Servers in one jurisdiction and reviewers in another is a common arrangement and a lawful one when properly papered, and a buyer cannot assess it without being told it exists.

So this axis has two answers to obtain rather than one: where systems process and store, and where the humans who read the output sit. The second is the part no infrastructure diagram will show.

The ordinary questions remain unanswered alongside it. Nothing establishes which model service generates the notes or what it retains, which matters for a vendor whose differentiator is reproducing each physician's individual style, since style adaptation implies something is retained and applied per clinician.

One practical note on the integration claim. Deep integration is asserted rather than described, with no standard, mechanism or scope named. A buyer should establish what it consists of for their own record system rather than reading the partnership as a specification.

Ask for the processing region, the reviewer location, the model provider, and the integration mechanism.

Thin but not silent. A 14 day free trial with installation support is published, and the vendor commits to deployment within a day, both of which tell a buyer something about cost of adoption. No rate card, tier structure or per provider figure was located, and third party directories describe pricing simply as tailored. Compare the vendors graded A on this axis, which publish the mechanism and the number.

Ambulatory and mobile by design, aimed at clinicians moving through a clinic rather than working at a fixed workstation, with telehealth capability included and orthopaedics among named users. Two real bounds: the product is iOS only, and no specialty count, specialty tuning or language coverage is published, so breadth cannot be compared against the vendors in this category that enumerate both.