CarePilot

CarePilot is the intelligence layer rather than a platform carrying one. A listed software group contracted it specifically to supply AI capability across an EHR it already owned, which is the clearest possible evidence that the model work is the product. This is the counterpart to the C graded on EverHealth Scribe: same intelligence, but here it belongs to the vendor rather than being licensed in.

The oversight model is now described explicitly, and it is the clearest statement of its kind in this category.

The distribution partner's published transparency note states that the system does not autonomously enter documentation into the medical record, that all generated output is presented for clinician review and editing before explicit submission to the chart, and that providers retain full authority over three named things: acceptance or rejection of a generated note, editing of structured clinical content, and final submission to the record.

Naming the three separately is what lifts this above the usual assurance. Most vendors in this lane say a clinician reviews the note, which leaves open whether structured data such as codes travels on a different path from narrative text. This states that structured clinical content is editable and that submission is a distinct explicit act. That is the right decomposition and almost nobody makes it.

The attribution caveat from the earlier assessment stands and is narrower than it was. The statement is the distributor's, describing the embedded deployment, and this company publishes nothing under its own name. But a gate that prevents autonomous writing to the chart is a property of the system rather than of a deployment, so it most likely describes the product itself.

What is unstated is the quantitative side: no acceptance rate, no edit burden figure, no confidence threshold and no description of what happens when the system is uncertain. The partner separately notes that clinician corrections refine the model over time, which means the review step is also a training signal and deserves its own question.

Ask what the review gate looks like when contracted directly rather than consumed inside the record system.

No accuracy figure, model card, named models, architecture description or evaluation methodology was located for CarePilot under its own name. Everything publicly documented about its performance arrives through its distribution partner's announcements rather than from the company itself.

Nothing published under this company's own name identifies any party in its chain: no model or model family, no foundation model provider, no hosting arrangement and no sub processor list was located in two passes. Everything publicly documented about this product's behaviour arrives through a distribution partner's announcements rather than from the company that builds it, which is the second instance of that asymmetry recorded in this lane and it repeats a structure worth naming.

The builder of a resold engine typically publishes less than the reseller, because the reseller is the one with a marketing motive and a customer to reassure, while the builder has neither and is not the party a health system is talking to. The consequence is that the most authoritative source about a model is the party least able to speak for it. The embedded route makes the practical problem concrete.

A practice using the record system branded product is sending patient audio to a third party, and the branding does not make that obvious, so the first thing many buyers would need to establish is that this company exists in their data path at all. Ask the reseller to name the engine, then ask this company directly for the model, hosting arrangement and sub processor list, and establish whether the reseller's published commitments bind it.

Two operating figures exist and their provenance is unusual enough to describe precisely. Practices using the embedded product are reported to have eliminated an average of eight minutes of documentation time per visit, with a 32 percent increase in same day claim submissions, and both are explicitly attributed to CarePilot data in a listed company's press release.

A public company putting its name to a partner's numbers is a modest step above an unattributed vendor claim, since it carries securities disclosure exposure. It remains vendor derived with no denominator, cohort, time period or methodology, and no study or independent assessment was located.

This remains the highest value gap on the record, though the second pass finds a partial answer in the distribution partner's transparency note.

That note states health privacy compliant handling with encryption and role based access, that there is no secondary use of identifiable patient data unrelated to care, and that de identified data may be used to improve system performance.

The training position is therefore stated in outline, and it is the middle of the three formulations this index now sees. Weakest is silence. Strongest is a flat commitment never to train on clinical content, or a permission gate placing the decision with the customer. This sits between: identifiable data is fenced, de identified data is available for improvement. That is a legitimate and common arrangement, and it turns entirely on a word the note does not define. De identification has a specific meaning under the privacy rule with two recognised methods, and the note says which outcome is claimed rather than which method is used.

A related statement from the partner's product material sharpens it. Clinician corrections are described as refining the model over time. So the review step is also a training signal, which means a clinician editing a note is contributing to model improvement, probably without understanding that.

What is still absent under either name is a retention schedule. Nothing states how long audio or transcripts persist, or whether audio is retained at all.

The embedded route adds the point the earlier assessment made and it stands: a practice using the record system branded product needs to know that patient audio reaches a third party, which the branding does not make obvious.

Ask for the retention schedule, the de identification method, and whether corrections train a shared model.

No published business associate agreement posture was located under this company's own name.

The two routes to market create two different chains and a buyer needs to know which they are in.

Contracting directly, the practice is the covered entity and this company is the business associate, so an agreement between them is required before any protected health information is handled. Nothing published establishes that one is offered or on what terms.

Using the record system branded product, the practice contracts with the record system vendor, and this company sits behind it as a subcontractor. The rule provides for that: a business associate must have written agreements with the subcontractors it uses. What it does not provide is visibility. A practice signing with the record system vendor may not realise a separate company processes their encounters at all, and the agreement they executed will name the vendor they recognise.

That is the substance of this axis for this record. Ask the record system vendor to confirm the subcontractor relationship in writing, to confirm the agreement covers it, and to state what happens to your data if the partnership ends. A supplier relationship between two companies can be terminated on terms neither of them negotiated with you.

One detail worth noting for anyone tracing the corporate chain. The mobile application carrying the record system brand is published by a differently named legal entity again, so at least three names attach to one product.

Ask which entity is party to your agreement, and whether the direct and embedded routes carry the same terms.

No named or dated attestation, no report of either type and no trust centre were located under this company's own name.

The distribution partner describes health privacy compliant infrastructure with encryption, access control and secure storage. That is a description of design rather than an examination of it, and it describes the partner's platform rather than this supplier's environment.

That distinction is the important one here, and it is a version of a pattern this index has recorded elsewhere. A practice adopting the record system branded product will encounter the record system vendor's security material, which is substantial because that company is a listed business with an established platform. None of it examines the supplier that actually processes the encounter audio. The assurance a buyer reads and the entity handling their patients' voices are not the same organisation.

So the question to put to the record system vendor is narrow and answerable: what security assessment did you perform on this supplier before embedding it, does your own attestation scope cover the supplier's processing, and will you share the outcome. A vendor embedding a third party into a clinical workflow has almost certainly done that diligence, and its result is the artefact a practice actually needs.

For a practice contracting directly, none of that applies and there is nothing published at all. Ask for whatever external testing exists.

The transparency note the partner publishes is a meaningful governance artefact and it is not a security attestation; it describes intended behaviour rather than examined controls.

No clearance claimed and none required for ambient documentation, and this axis reads better than the category norm because an explicit statement of intended purpose exists.

The distribution partner's transparency note states that the product is not intended to diagnose, treat, or recommend clinical care decisions, and describes its scope as outpatient documentation support for licensed clinicians in real time or post encounter workflows.

That matters more than it may appear. Intended purpose is the test that determines whether software falls inside the device framework, and it is set by what the manufacturer claims rather than by what the technology could do. A published statement of intended use, phrased in the terms the regulatory question actually turns on, is the single most useful thing a vendor in this category can provide, and it is almost entirely absent from this lane. Most vendors leave a reader to infer intended purpose from a feature list, and several use marketing language that pulls in the opposite direction from their regulatory position.

One capability sits at the edge of that statement and should be checked against it. The product recommends diagnosis codes. Code recommendation is ordinarily treated as administrative rather than clinical, and it is a recommendation about how an encounter is characterised for payment. That is compatible with the stated intended use and it is the boundary a buyer should confirm has been assessed.

The attribution caveat applies as elsewhere: the statement is the distributor's and this company publishes nothing under its own name.

Ask whether the same intended use statement governs the direct product, and whether the coding capability was included in that assessment.

This is the first commitment in this category that directly addresses the question this axis asks, and it is worth setting out carefully including the attribution problem attached to it.

The distribution partner publishes a structured AI transparency note for the embedded product. It states that evaluation includes accent based assessment alongside provider feedback, note completeness and usage derived measures, that monitoring is continuous, and that on detection of disparities the response includes adjustments to prompts, pipeline, tooling and model fine tuning. It also publishes a limitations section naming audio quality, overlapping speech, accents and jargon as factors affecting output.

No other vendor across this entire category publishes a commitment to accent based evaluation with a stated remediation path. Most publish nothing; several assert accent handling without evidence; one invites clinicians to test it themselves. This is the only record where a named evaluation activity and a response to detected disparity both appear.

The attribution caveat matters and does not defeat it. The note is published by the distributor, not by this company, and this index does not ordinarily transfer a partner's claim onto a supplier's record. But accent evaluation and model fine tuning are activities only whoever controls the model can perform, and the note names this company as the builder. So the commitment is most likely describing this vendor's work, reported by its partner.

What is still missing is the evidence itself: no results, no populations, no thresholds, no methodology.

A practice contracting directly should confirm the same evaluation applies outside the embedded route, and ask what the assessments have found.

Two passes located no accuracy or error figure under this company's own name, no published limitations and no warranty, indemnity or remediation commitment. This is the builder side of an embedded arrangement and it produces the same split this index has recorded on the other resold engine in this lane: the company controlling model behaviour has no relationship with the health system, and the company holding the contract does not control the model.

Neither publishes a document covering the combined product, so a question about what happens when a note is wrong has no single addressee. Two features specific to this arrangement sharpen it. The partner material states that clinician corrections refine the model over time, so the review performed inside the reseller's product feeds a model this company controls, which means quality improvement flows to the builder while accountability stays ambiguous.

And the branding conceals the arrangement, so a practice may not know a separate company processes its encounters, which makes it unlikely to ask any of these questions at all. Establish which entity carries the performance representations, whether any supplier warranty passes through to the practice, and whether the reseller's published commitments are contractually binding on this company.

Narrow and properly documented, which is better than broad and asserted. CarePilot holds its own listing in the DrChrono Marketplace and the authorisation mechanism is described concretely: clinicians sign in with their DrChrono credentials to authorise the integration, and each provider at a practice authorises individually. Per provider authorisation is a meaningful detail because it means access is granted and revoked at the clinician level rather than the practice level. No integration with any EHR other than DrChrono was located.

No hosting region, residency option or subprocessor detail was located under this company's own name. Delivery is cloud, reaching clinicians either through a browser based application contracted directly or embedded inside a record system's web and mobile applications.

The embedded route is the one that needs describing, because it makes the data path longer than a clinician would assume from the interface. A provider taps a control inside their record system, speaks to a patient, and a structured note appears in the chart. What has actually happened is that audio has left the record system vendor's environment for a supplier's, been processed there, and returned. Three organisations have touched the encounter and the clinician has interacted with one brand.

So the residency question has to be asked of the supplier rather than the record system vendor, and the record system vendor may not be able to answer it. Where does audio go, is it processed in the same jurisdiction, which model provider handles it, and what does that provider retain.

One fact makes the chain concrete. The mobile application carrying the record system's brand is published by a third differently named legal entity, so the corporate structure behind the product is not visible from the interface.

The partner's transparency note addresses oversight, evaluation and data use, and does not address hosting location or subprocessors at all.

Ask for the processing location, the subprocessor list, and the model provider, and put the questions to the party that actually processes rather than the party that sold it to you.

No published rate card for the direct product and no statement on how the embedded route is priced. A practice on DrChrono may be able to reach the same intelligence two ways, through the marketplace listing or through EverHealth Scribe, and neither price is public, so the comparison cannot currently be made.

No specialty count, note format list or language support was located under this company's own name.

The distribution partner's transparency note supplies the setting, and it is stated with useful precision: outpatient documentation support for licensed clinicians, in real time or post encounter workflows. Naming both timings matters, because a product usable after the encounter serves a different workflow from one requiring live capture, and many vendors leave that ambiguous.

The note structure is also partially described, covering history of present illness, assessment and plan, alongside diagnosis code recommendation.

What remains unstated is breadth. No specialty coverage is claimed under either name, no language support is mentioned, and no indication is given of which practice types the product has been tuned for. That is a gap worth pressing rather than assuming, because the partner's transparency note names accents and jargon among the factors affecting output, which implies specialty vocabulary is a known variable.

The practical bound noted in the earlier assessment stands: reach through the embedded route is shaped by the record system's own base, which serves independent ambulatory practices, so the product has been exercised most in that setting whatever its nominal scope.

A specialist practice should therefore treat specialty fit as unevidenced rather than assumed. Ask which specialties the product has been used in at volume, request sample notes in your own specialty before contracting, and establish whether templates exist for your note formats or whether you will be adapting a general structure.