Apricot Health

Generative AI drafting the documentation is the product. There is no electronic medical record, no services business and no prior software underneath it, and the company describes itself as AI native. What it automates is the drafting of OASIS and visit documentation from captured content, which is the model doing the work rather than a workflow layer presenting someone else's output.

The oversight advantage here is architectural as well as procedural. Because documentation is drafted from a guided interview the clinician completes after the visit, plus structured capture during it, the clinician supplies the content rather than a model inferring it from overheard conversation. That materially narrows the fabrication surface compared with ambient generation, and the company argues the point explicitly, positioning against approaches it characterizes as riskier.

An in home visit guide reduces omission during capture, and the company's own representatives state plainly that the product drafts OASIS forms from patient documents but does not replace clinical decisions, with agency review keeping final decisions with the team.

None of it is quantified: no accuracy or error rate, no confidence signal, no abstention behavior when the interview leaves a required field unsupported, and no statement of what the agency review layer actually catches.

The company names transparent reasoning as a product property, meaning a clinician can see why a recommendation or draft entry was produced rather than receiving it unexplained. That is a real feature claim. Nothing else is disclosed: no model or model family named, no accuracy figure for OASIS field completion or narrative generation, no evaluation methodology, no error taxonomy and no model card. For a product whose output is a regulated assessment instrument, per field accuracy is the disclosure that matters and none exists publicly.

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, and no training position is stated in either direction. The founder's published checklist tells agencies to require business associate agreements with language model providers, with training data restrictions included; the vendor does not say which providers it uses or what its own agreements with them contain.

The content types are unusual and unlikely to share one retention treatment. Inside the home the product captures signatures, vital signs, photographs and voice notes, and uploaded documents feed the same pipeline. Photographs need their own answer: an image is identifiable in ways text is not, may capture the patient's body, their living conditions and other people present, and persists as an image that cannot be redacted the way a transcript can. Captured consents form a separate category, since a signed consent is both evidence and personal data. Nothing published says what happens to captured content when a visit is abandoned or a referral does not proceed, when the material exists but no episode of care was created to govern it.

Strong commercial signals, and customer evidence that is named but not independently measured. A Series A led by Insight Partners in October 2025 is meaningful external scrutiny, and a co branded integration page on the Netsmart marketplace is a documented partnership rather than an assertion.

Customers are named in the vendor's own press releases, including Reliant at Home, Royal Care and Advanced Home Health and Hospice, and a vendor case study reports that Choice Health at Home cut documentation time by 85 percent. No controlled comparison or independent evaluation is published, though the case study itself may set out its method.

The headline efficiency figure needs reading carefully. Start of Care documentation falling from 120 minutes to 15 to 30 minutes has been footnoted as measured against leading EMR time targets rather than against observed time at a customer. A target is not a baseline, and comparing measured performance to somebody's target overstates the delta by whatever gap already existed between target and reality.

The vendor publishes no retention period, training use statement or deidentification posture.

The architecture changes the shape of the exposure rather than removing it. This is not primarily ambient capture of a patient conversation. The core flow is a guided interview completed with the clinician after the visit, whose answers, together with uploaded documents, photographs and other in home data, are transformed into structured documentation. So the principal generative input is the nurse's own account rather than a recording of the patient speaking.

That is a lower exposure design than continuous encounter recording, and it counts in the product's favor. A clinician answering structured questions produces far less incidental content than an open microphone in a patient's living room.

What remains is still substantial and specific. The product captures signatures, vital signs, photographs and voice notes inside the home. Photographs are the item that most needs its own answer, for the reason that applies across home based products: a wound photograph is identifiable in ways text is not, may capture the patient's body, their living conditions and other people present, and persists as an image that cannot be redacted the way a transcript can. Consents captured in the home are a second category that needs its own answer.

The gaps are narrower than an ambient product's but no less pointed. Nothing published gives the retention period for photographs specifically or for voice notes, says whether either is used to improve models, or describes what happens to captured content if a visit is abandoned or a referral does not proceed.

A privacy notice is published, and its scope is what matters here. It states that it describes the processing of personal information provided, collected or disclosed for the websites and online platforms that link to it, and the categories it enumerates are contact, commercial, employment, usage and communication information. There is no patient data, no protected health information, no clinical content and no reference to the documentation service anywhere in it. It discloses Google Analytics and Meta pixels and discusses interest based advertising. It is a marketing site notice, and an unusually complete one.

The position falls short in two ways at once. Compliance is claimed without the underlying document: a HIPAA badge sits in the footer of every page alongside a SOC badge, with no business associate status stated, no agreement published, no scope and no execution path. And the published privacy notice covers the website rather than the service that handles patients.

The distinction matters for a buyer because the two read identically from the outside. A vendor with no privacy documentation and a vendor with thorough privacy documentation about the wrong thing both present a footer badge and a policy link. Here the documentation is real and its subject is website visitors and job applicants.

The gap matters more here because of what the product does. Apricot documents start of care assessments in patients' homes, with clinician controlled audio capture, and the assessment drives episode payment. Nothing published states business associate status, offers an agreement, or describes what happens to captured audio.

SOC 2 Type I is stated first party: the vendor announced in September 2025 that an examination confirmed its controls are suitably designed, and a SOC badge appears in its site footer.

The report type is the substance here. A Type I report assesses whether controls are suitably designed at a single point in time. A Type II report tests whether those controls actually operated effectively across a defined period, usually several months. They are different assurances, and the gap between them is where most real control failures live. The vendor states the type accurately rather than saying only SOC 2.

Type I is normally the first step toward Type II, a year after the announcement no Type II is published, and no date for one is stated. The current report is evidence about design rather than operation.

Also unresolved: no trust center carrying the report, its date or a scope statement, and no ISO 27001 or HITRUST alongside it. Scope matters for a platform capturing photographs, signatures, vital signs and voice notes inside patients' homes.

No FDA clearance or device authorization is claimed and none is expected, since the product completes a documentation and assessment instrument rather than making a diagnosis. The regulatory exposure that does matter sits in the CMS conditions of participation for home health and in the accuracy of the OASIS assessment itself, which drives both quality reporting and payment.

The counterweights are substantial. The company states that clinician informed recommendations keep coding and documentation aligned while transparent reasoning and agency review leave every final decision with the team, its representatives state directly that it does not replace clinical decisions, the founder ran a home health agency for seven years, and the design is described as informed by extensive field research with nurses.

The founder has also published a checklist for agencies vetting AI vendors that asks about anti upcoding policies and compliance sign off. Against that, OASIS is not a neutral form: it drives case mix under the Patient Driven Groupings Model, so a product that shapes OASIS responses is operating on the instrument that determines what a thirty day episode pays. No fairness, subgroup or demographic performance disclosure is published.

There is one mechanism in its favor and one absence against it. The vendor names transparent reasoning as a product property, so a clinician can see why a recommendation or a draft entry was produced rather than receiving it unexplained. That is a real correction mechanism at the point of review, because a clinician who can see the basis for an entry can reject it on that basis rather than on instinct.

The design also has a lower exposure architecture than continuous recording, since the principal generative input is a guided interview completed with the clinician after the visit, and a clinician answering structured questions produces far less incidental content than an open microphone in someone's living room. The absence is the one that matters most for what this product outputs.

Its deliverable is a regulated assessment instrument whose fields drive payment grouping and quality reporting, and no per field accuracy figure, error taxonomy or evaluation methodology exists publicly. For a product of this kind, per field accuracy is the disclosure a buyer needs, because errors are not evenly distributed across an instrument and the consequential fields are a small subset. No warranty, indemnity or remediation commitment is published.

One integration is properly evidenced rather than merely claimed, which is worth more than a long unnamed list. Apricot integrates with Netsmart myUnity, the post acute care record, and appears on the Netsmart marketplace with a co branded page authored jointly with that vendor.

Being carried in a post acute EMR's own catalog is a form of external validation, since the platform vendor very likely applies its own review before listing, and it indicates the integration is supported rather than improvised. The product is separately demonstrated for Axxess users. Netsmart is the only integration described in any depth, the other record systems that dominate post acute care are not addressed, and no integration mechanism or standard is described.

No hosting model, cloud provider, region or residency commitment is published, and nothing published lists subprocessors or addresses third party model services.

The specific gap for this product is capture. It happens inside patients' homes, where connectivity is frequently poor and often absent, and the product collects signatures, vital signs, photographs and voice notes there before any documentation is generated. So content exists on a device, in a house, for some period before it reaches anything the vendor controls.

Nothing published states what that period is or how it is protected: whether capture works offline, what is held on the device until it syncs, whether it is encrypted at rest there, what happens if a device is lost between the visit and the sync, or whether an agency can remotely revoke or wipe. Those are operational and security questions at once, and for a workforce driving between homes all day they are not edge cases.

The generation step raises the ordinary questions on top. The guided interview and the documents, photographs and notes accompanying it are processed somewhere, and nothing establishes where, by whom, or what a model provider retains.

The product reaches customers partly through a post acute record vendor's marketplace and integrates with that vendor's platform, and nothing says which organization's infrastructure holds what at each stage.

No price, tier or pricing mechanism is published.

The Netsmart marketplace listing may carry pricing or packaging detail the vendor's own site omits, as partner catalogs occasionally do.

Narrow by design and deep within the narrow part. The product addresses home health and post acute care, and within that it began with the Start of Care visit, the most complex and time consuming documentation event in home health, covering it end to end across medication reconciliation, wound documentation, calendar plotting, interventions and goals and narratives. That is instrument level work rather than generic summarization, since OASIS is a specific regulated assessment.

Support is described as extending to other OASIS, therapy and routine visits. The breadth beyond Start of Care is recent or announced rather than demonstrated, and coverage does not extend past this one care setting.