Chartnote

The ambient scribe is real model work, but it sits on top of a product that predates it and would survive without it. Chartnote's original and still distinguishing asset is a library of thousands of medical snippets and dot phrases plus a custom template system, which is curated content rather than intelligence. Remove the AI and a useful documentation tool remains, which is not true of the pure plays. The moat is the content library, not the model.

Conventional draft and review with the same structural gate as other transfer based products here: the note is generated in Chartnote and the clinician moves it into the record, so nothing files itself and a human is necessarily in the loop at the point of entry. No acceptance rate, edit burden figure or confidence threshold is published.

No accuracy figure, model card, named models or evaluation methodology is published. Capability is described functionally, generating SOAP notes from patient conversations with high accuracy medical speech to text, without any measurement attached to either claim.

Nothing identifies any party in the chain. No model or model family, no foundation model provider, no hosting arrangement and no subprocessor list is published. Health privacy compliance is asserted alongside robust security protocols, which describes posture rather than naming anyone who processes an encounter.

The self serve distribution makes this gap harder to close through the usual route. A health system can compel a subprocessor list through procurement. An individual clinician adopting the tool has no leverage and no counterparty other than a signup page, so published disclosure is the only available source, and there is none on this question.

The company publishes a privacy and security page and a trust portal. Either may name a model provider or list subprocessors. A trust portal is the conventional place for a subprocessor list, which makes reading it the obvious next step rather than a speculative one.

App store reviews and short clinician quotes only. No study, controlled evaluation, accuracy benchmark, third party rating or named health system deployment is publicly available. Much of the comparative material circulating about this vendor is published by competitors selling against it, and by Chartnote's own blog comparing itself to others, so the surrounding commentary is self interested.

Very little is published, and that is itself the finding.

No retention schedule, deletion commitment, deidentification practice, encryption detail or training use statement is published for this product. The vendor's own material describes health privacy compliance with robust security protocols, which is an assertion about posture rather than an account of what happens to an encounter.

The distribution model is what makes thin documentation consequential here. This is a self serve tool that an individual clinician can adopt without any institutional review. Nobody with expertise in retention, training use or subprocessors sits anywhere in that adoption path. Where a hospital buys a scribe, procurement extracts these answers privately and the vendor's website matters less. Where a clinician signs up alone, published material is the only mechanism, and there is very little of it.

Four answers are missing. Nothing published says whether audio is retained or for how long, whether transcripts are retained, whether any of it is used to train or improve models, or which third parties process the encounter. Getting those answers on record falls to the clinician as a professional responsibility.

None of this suggests the answers are unfavorable. It records that they are not obtainable without asking, from a vendor whose buyers are least equipped to ask.

Health privacy compliance is asserted in the company's own material, and the company publishes a dedicated privacy and security page at chartnote.com/hipaa covering the United States, Canadian and European regimes, plus a trust portal and a published account of how it reached compliance, which describes continuous control monitoring through Vanta and quarterly security reviews.

The agreement itself is not published. The dedicated page and a help desk article on the same subject may state it. So a compliance surface exists and is substantial; whether it includes a business associate agreement is the open question.

The distinction between the two things is the whole question. Compliance is a posture a company asserts about its own practices. The agreement is a contract that makes the company legally accountable to the covered entity, defines permitted uses and creates breach notification obligations. A clinician who uses a tool on protected health information without an executed agreement has a problem regardless of how well built the product is, and the compliance claim does not cure it.

For a self serve product this is the single most important thing to establish and the easiest for a company to answer. Nothing published says whether an agreement comes with an account at signup.

This is not a finding that no agreement exists, only that none can be confirmed before contracting.

No named or dated attestation, no report of either type, no penetration testing statement and no trust center is published.

The company appears small, and the gap is in what a counterparty can verify, not a finding that controls are absent. The usual proportionality argument applies with one qualification worth keeping.

That argument runs: a small vendor can reasonably defer the cost of an attestation because its buyers are sophisticated and will extract the answers through procurement instead. It does not hold here. This tool is adopted directly by individual clinicians. There is no procurement, so there is no alternative mechanism, and publishing becomes the only route by which a buyer learns anything.

That does not mean a small vendor must hold a report. It means the substitutes matter more: a plainly written security page describing encryption, access control, retention and subprocessors would do most of the work an attestation does for a buyer of this size, at a fraction of the cost.

The vendor publishes a comparison of scribes on its own blog which includes an accurate limitation about its own integration depth. A company willing to name its own weakness in its own marketing is likely to answer a direct security question straight.

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, and that counts as a positive rather than an absence. It combines ambient capture, dictation and smart templates to produce a note. No coding engine, risk adjustment scoring, decision support layer, order generation, patient facing component or revenue cycle module is described. The vendor's own description of itself is confined to documentation.

The distribution model has one consequence, and it is not a criticism. Because the tool is adopted directly by individual clinicians rather than procured by institutions, the ordinary checkpoint at which an organization would assess a clinical AI product before it touches patients does not occur. Nothing is being circumvented, since no pathway attaches and no rule requires that review. But the professional obligation, the consent obligation and the vetting obligation all rest on the individual who signs the note, without the institutional support a hospital deployment provides.

That holds for any self serve product, not only this one. The clinicians most likely to adopt this way are those least likely to have read anything about the obligations they are assuming.

Local consent requirements for recording vary by state and are the clinician's responsibility rather than the vendor's.

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

Where a vendor makes no multilingual claim, its silence on cross language performance does not contradict a marketing promise, and that is a materially different position from a vendor advertising dozens of languages with no evidence behind any of them. Chartnote sits in the first group.

What remains is the accent and dialect question, which applies to every English language product regardless of what it claims, because speech recognition performance varies across speakers whether or not a vendor mentions it.

The distribution model shapes how that question can be pursued. Adoption is by individual clinicians rather than institutions, so there is no organization positioned to run an evaluation, aggregate error patterns across a practice, or notice that documentation quality differs systematically for some patients. Each clinician sees only their own encounters, which is exactly the vantage point from which distributional differences are invisible.

So the realistic ask is narrower than for an enterprise vendor. Not a published subgroup study, which a company this size will not have, but a straight answer to what the underlying speech model is and what its published performance characteristics are. If the vendor licenses a commercial speech service, that provider may have published evaluation data the vendor could simply point to.

No accuracy or error figure, no published limitations and no warranty, indemnity or remediation commitment is published. Capability is described functionally, generating structured notes from patient conversations with high accuracy speech to text, with no measurement attached to either claim.

The distribution model determines what that absence means in practice. This is a self serve product an individual clinician can adopt without any institutional review, starting on a free tier. Where a hospital buys a scribe, procurement extracts these answers privately and published material matters less. Where a clinician signs up alone, published material is the only mechanism, and there is very little of it. Nobody with expertise in accuracy, retention or remediation sits anywhere in the adoption path.

The company does publish a privacy and security page and a trust portal, and describes continuous control monitoring and quarterly security reviews. Those speak to how the system is operated rather than to what happens when its output is wrong. Either page may address accuracy or remedies.

The distribution model leaves three answers missing. Nothing published says whether audio and transcripts are retained or for how long, whether either trains models, or what the company commits to when a note is wrong. Having those answers on record is the clinician's professional responsibility.

None of this suggests the answers would be unfavorable. It records that they are not obtainable without asking, from a company whose buyers are least equipped to ask.

Transfer based rather than integrated: finished notes move into any web based EMR, which makes the product usable anywhere and connected to nothing. There is no structured field population, no chart context flowing in and no write back, so the clinician carries the note across the gap. For a solo or small practice that means zero implementation, and for anyone wanting the note to arrive as discrete data it means this is not the right tool. The phrasing is at least honest about what it does.

No hosting region, residency detail or subprocessor information is published. Delivery is cloud with web and mobile access.

One fact the vendor publishes about itself stands out. On its own blog it maintains a comparison of ambient scribes which includes itself, and the entry states plainly that integration with certain record systems may be limited. That is a real limitation named by the vendor in its own marketing.

It is also the disclosure that matters most for deployment. Where integration is limited, the note leaves the product by hand, which means the vendor's environment holds the encounter and its output for whatever interval elapses before a clinician moves it, and no institutional system has visibility of what was captured. Nothing published says which record systems are supported or by what mechanism, and the answer decides whether a clinician's workflow is a connected one or a copy and paste one.

The rest is unanswered. Nothing establishes where processing runs, whether an external model service handles the encounter, or what that service retains. For a product built on generative AI, the model provider question is the one most likely to have a simple answer the vendor could publish today.

The pricing structure is public and unusually sensible, tiered by monthly AI note volume rather than per seat: a free Basic tier with 15 AI scribe notes a month plus templates and snippets, a Pro tier capped around 100 notes with unlimited Copilot tools, and an unlimited Max tier, with subscriptions purchasable directly through the app store. Volume based tiering means a part time or low volume clinician pays proportionately, and it prices the product to how much a clinician documents rather than to the fact of employing them.

The amounts are not all published by the company. Only the top tier figure, around $100 a month, is public, and it comes from a third party comparison rather than from the vendor, so the full ladder is not confirmed by the company's own material. Publishing all four tier prices on the site would be a small change: the structure is already the hard part and it is already public.

One note on how to read the free tier. Fifteen AI scribe notes a month is usable rather than a trial, but nothing published says whether the business associate agreement extends to the free tier.

Ambulatory and urgent care focused, with family medicine, emergency medicine, internal medicine and urgent care named, and support for SOAP, DAP and custom note structures. The snippet and template library is the practical mechanism for specialty fit, meaning coverage depends on what a clinician builds rather than on models tuned per specialty. No specialty count, no language coverage and no inpatient capability documented.