Roger Healthcare

Referral summarization, chart pre fill, ambient capture, the real time prompting about missed items, and completion of the OASIS and routine notes are all model output. No human scribe tier, no coding services layer and no underlying EHR that the AI is a feature of. The company positions itself as a documentation platform sold to agencies, and the RPA sync is plumbing around the model rather than the product itself.

The oversight question here is not whether a human reviews but when, and the ordering is the weak point.

The described sequence is that the system fills out the entire chart in the agency record system within minutes and the clinician then reviews the result without changing existing workflows. That is write first, review second. The safer ordering is the reverse, where generated content populates the record only after the clinician has approved it, because a completed document invites confirmation rather than scrutiny and because anything corrected afterward leaves the original entry behind.

Two real mitigations sit alongside that.

The assessment is not submitted to the payer until the clinician signs it, so a draft sitting in the record system is not yet a regulatory filing. The signature is still a working gate even though the write has already happened.

And real time prompting during the visit is an oversight strength. Prompting a clinician about items not yet addressed while they are still in the patient's home catches omissions at the only moment they can still be resolved by asking. Compliance feedback at the point of care is materially better than a quality query raised days later, when the answer has to be reconstructed.

What is missing is everything about the system's own behavior. No confidence threshold, no abstention behavior when the audio or the inference is poor, and no escalation path is published, so nothing states what happens when the system is unsure rather than wrong.

The approach is named. The vendor describes a multistage artificial intelligence pipeline trained on millions of rows of home health visit data, and separately names its record system sync as a proprietary robotic process automation system. Both are real architectural statements rather than a claim of proprietary technology, and both say what the product is and what to test.

The domain training claim is also given a concrete worked example rather than asserted, which is useful: when a patient says they walked to the mailbox yesterday, the system reads that as implying particular functional scores for ambulation. That is exactly the inference a general model would miss and exactly the inference that carries risk, since a functional score drives payment and quality reporting. Showing the mechanism opens it to interrogation.

The version and update discipline behind those names is absent. No model card, named underlying model, accuracy figure or evaluation methodology is published, and nothing states how or whether customers are told when the pipeline changes. On a product that infers regulated scores from conversational cues, a silent model change alters what the scores are inferred from.

The headline 80 percent time saving is described as proven without a study, denominator or comparator behind it, which is a claim about evidence rather than a measurement.

The corpus and the architecture are described and no party is named, which places this mid band. The vendor states a multi stage pipeline trained on millions of rows of home health visit data, and separately names its record system synchronization as a proprietary robotic process automation layer.

Both are real architectural statements, and the corpus description implicitly answers a question competitors dodge: this model was trained on real assessment data, which is why it outperforms a general scribe on these forms. That makes provenance the question rather than an accusation.

Nothing published says whose data that was, on what basis it was used, or whether a customer agency's own content joins it, and a specialist model trained on real records is only as defensible as the permissions behind the records. On enumeration there is nothing: no model or model family, no foundation model provider, no hosting arrangement and no sub processor list. One content type needs its own answer.

The product captures photographs of medication bottles to extract prescription data, and a medication label carries the patient's name, frequently their address, the prescriber and the pharmacy, so the image contains identifiers well beyond the drug.

Two named agencies and no study behind either.

New Day Healthcare, reported at 17,000 patients across six states, has rolled the product out company wide. Central Coast Home Health is published as a second named customer, with a stated timeline of pilot to full deployment in one month and integration completed in under a week. Named agencies adopting across an organization rather than piloting, with a deployment timeline attached, shows the product surviving contact with a real operation.

What none of it establishes is benefit. The headline 80 percent time saving is asserted as proven, site wide, with no cohort, baseline, comparator or method. Describing an unpublished result as proven is a stronger claim than an unsupported percentage, because it asserts that verification has happened. The Central Coast material describes adoption speed and clinician reaction rather than measured outcomes, and the remaining evidence is unattributed clinician quotes. No measurement of documentation accuracy, assessment completeness or scoring agreement appears anywhere, and scoring agreement is the one that matters most here, since the product infers regulated functional scores that drive payment.

One conflict cannot be settled from either vendor's materials. A competitor's published comparison lists New Day Healthcare as its own customer while this vendor publishes a company wide rollout there. That may mean parallel evaluation, a switch in either direction, or different business units within a large organization, but both products cannot be the sole documentation system across it.

No retention schedule, audio deletion commitment, encryption detail, de identification practice or training use statement is published.

Three content questions are specific to this product.

The training corpus first. The model is described as trained on millions of rows of home health visit data. Nothing says whose data that was, on what basis it was used, or whether a customer agency's content joins it. A specialist model trained on real assessment data is very likely why this product handles these forms better than a general scribe, and the provenance still has to be answerable.

The access surface second. The integration is a robotic process automation layer holding credentials to the agency record system, and it does not deposit a note: it populates an entire regulated assessment, field by field. So the vendor holds a durable credential to a system containing every patient the agency serves, not only the ones its product documented. The access surface is the whole record rather than the encounters in scope.

The third is new and easy to overlook. The product captures photographs of medication bottles to extract prescription data. A medication label carries the patient's name, frequently their address, the prescriber and the pharmacy, so the image holds identifiers well beyond the drug, and an image is harder to redact selectively than text. Nothing published states how long those photographs are held, whether they are retained after extraction, or whether they are used for improvement.

The home setting adds a fourth consideration: recording happens in a residence where family members and caregivers may be present and are not the patient.

No health privacy statement, business associate agreement reference or compliance page appears in the vendor's published materials.

The fair reading is that agreements plainly exist. The platform records patient encounters and writes into record systems at named enterprise agencies, and no agency of that size onboards a vendor handling protected health information without one. The absence is far more likely a website gap than a compliance gap, and for a company of this size it is a simple thing to fix.

No terms are visible before a sales conversation, and a vendor whose entire product involves recording patients in their own homes and holding durable credentials to an agency's record system has not published a single line about the contractual basis for either.

The scope of the agreement matters, because the platform now spans more than visit documentation. The intake product processes inbound referral packets, creates patient records automatically in the record system, and runs eligibility checks against a large number of payers. Referral processing means the vendor holds protected health information about people who have not yet been admitted and may never be, and who have no relationship with the agency at all if the referral is declined. Eligibility checking sends identifiers onward to payers. An agreement drafted for ambient documentation may not anticipate either.

The home setting adds a further question. Recording occurs in a patient's residence, where family members and others may be present and are not the patient.

No named or dated certification, audit report, penetration testing statement or trust center exists.

What makes that matter more here than for a typical scribe is the architecture. The integration is a robotic process automation layer holding credentials to the agency record system, which it drives to fill in every field of the chart.

That access model is what an examination would need to cover, and the questions it raises are not the ones a standard scribe assessment answers. Where are the credentials stored and how are they protected at rest. Whose account does the automation operate under, and does it appear in the record system's audit log as that person. Can an agency distinguish an automated write from a human one after the fact. What happens on staff departure, when a credential should be revoked but an automation depends on it. And what is the blast radius if the vendor's environment is compromised, given the credential reaches the whole record system rather than the visits the vendor documented.

Those are answerable questions and none of the answers is published. They are also the questions an agency's own auditors will eventually ask, since the agency remains accountable for what appears in its records.

The customer base raises the expectation rather than lowering it. The vendor states enterprise agencies use the platform, and enterprise procurement normally requires a report before a vendor touches the record system. The company is small, which is a fair mitigating consideration on timing but not on the access model, because the access model is the same at any headcount.

No device claim is made and the product is scoped as documentation; no pathway is stated and no clearance is claimed.

The regulatory weight here sits with the payment and survey regime rather than device oversight, and it is heavier than it looks. The assessment instrument this product completes, including the functional series, drives both quality reporting and case mix payment, so a score inferred from conversation lands in survey findings and in reimbursement rather than only in a note.

The vendor states its position on review plainly. It markets being fully artificial intelligence driven as a differentiator, contrasting itself with services that use people to quality check notes and framing that review as a delay of a day or more. That is a coherent product argument about speed. It also means the only review between an inferred functional score and a submitted assessment is the clinician's own.

Scope is the second point. The product generates a structured plan of care from the visit's findings, which is a step past recording an assessment, and the sibling intake product generates diagnosis codes described as maximizing reimbursement. Neither is device territory; both are payment integrity territory.

Mixed, with a real positive and a real gap.

The positive is that the compliance mechanism runs forward rather than backward. The system prompts the clinician about undocumented items while they are still in the patient's home, which prevents an omission rather than querying it afterward. In a segment built around quality assurance staff returning charts to clinicians days later, moving that check to the moment the patient is still present is a structural improvement, not a feature: the clinician can simply ask.

The revenue framing also stays off the coding gradient. The stated mechanism is throughput, serving more patients without adding staff, rather than coding intensity or capture. Speed is not a coding metric, and nothing in the documentation product is sold on lifting reimbursement. The sibling intake product is a different matter and is assessed separately.

The gap is disclosure, and it is total. No fairness statement, subgroup analysis, or accent and dialect performance breakdown is published, despite multi-language capture being sold as a feature with no language count given.

That gap is especially sharp in this setting. The home health workforce is among the most linguistically diverse in United States healthcare, and the patients are elderly, frequently with hearing, speech or cognitive impairment, recorded in uncontrolled home acoustics rather than a consulting room. Every factor that degrades speech recognition is present at once, and the system is inferring regulated functional scores from that audio.

Nothing published describes how the model is evaluated, how often it is retested, or what happens when a release changes its behavior.

A review step is published, and the design weakens it.

The clinician reviews the completed chart and signs the assessment before it is submitted, and the signature is the gate that matters: until it happens, a draft sitting in the record system is not yet a regulatory filing. Real time prompting during the visit adds a second mechanism, surfacing items not yet addressed while the clinician is still with the patient and can still ask.

Two things weaken it, and they are specific to this architecture.

The review follows the write rather than preceding it. The system fills the chart first and the clinician reviews afterward, so what is being checked is a completed document rather than a proposal, and the burden of noticing an error falls on someone reading text that already looks finished.

And the forensic mechanism is compromised by the integration itself. The automation layer holds stored credentials to the agency record system and acts under a named person's identity, so the audit log stops being a reliable account of who entered what. For a regulated assessment that drives payment and quality reporting, establishing whether a field was entered by a clinician or generated by software is exactly the evidence a later dispute or audit would need, and it is the evidence this design removes.

The credential model has a further consequence. The vendor holds durable access to a system containing every patient the agency serves, not only those its product documented, so the access surface is the whole record rather than the encounters in scope.

Nothing stands behind the output. No accuracy or error figure, no published limitation and no warranty, indemnity or remediation commitment is published, and the headline 80 percent time saving is described as proven with no study, denominator or comparator behind it. Describing an unpublished result as proven is a stronger claim than the evidence supports.

Nothing says whose credentials are used, whether the automation's entries are separately identifiable in the audit trail, or what the vendor commits to when a field is wrong.

No record system is named anywhere. For a home health agency the systems that matter are a short and well known list, and which of them this product supports today is not published.

What is described is functionally deep and architecturally brittle, and the vendor says so plainly. The sync is a proprietary robotic process automation layer operating in near real time, working in both directions: pulling attachments, schedules and referral packets out of the record system before the visit and filling in the chart afterward. Reading referral documents out and writing a full structured assessment back is a wide scope for an automation layer, since each direction has to survive a different set of screens.

The costs of that architecture are the standard ones. Robotic process automation drives the user interface rather than a sanctioned data channel, so it breaks when the vendor of the record system changes its screens, and it carries no contract with that vendor to prevent such changes. The audit trail shows an automated session rather than an integration, which matters here because the thing being written is a regulated assessment.

Naming the mechanism rather than calling it integration is candid, and it shows exactly what to test.

No hosting region, residency option, subprocessor detail or offline behavior is published. Delivery is a cloud service with a mobile capture application and a robotic process automation connector into the agency record system.

Two questions are specific to this product rather than generic to the category.

Offline behavior is the first, and it is unusual for this to be unaddressed. Capture happens in patients' homes, frequently in rural areas and inside buildings with poor reception, and the workflow explicitly continues after the visit with additional dictation. Nothing published states what the application does when the signal drops mid-visit, what is held on the handset until it syncs, whether it is encrypted at rest there, or what happens if a device is lost with unsynced visits on it. For a product whose entire capture surface is mobile and off-site, that is the first operational question an agency would ask.

The second follows from the connector. Because integration works by driving the record system's interface under stored credentials rather than through a permissioned interface, the vendor's environment holds a durable path into the agency's system. So the residency question is not only where audio and transcripts sit. It is also where those credentials sit, and what network path the automation runs over to reach a record system that may itself be hosted elsewhere.

Which record systems the connector supports is not published either, so a buyer cannot establish coverage without asking.

No price, unit of pricing, tier structure or trial is published anywhere, and every route into the product ends at a booked demo. That position was checked against the vendor's own estate rather than assumed: the site, the customer stories and the blog all terminate in the same call to action and none carries a figure.

The one price signal available is a customer testimonial stating that competing products were more expensive and did less. That is a positioning claim from an unnamed source rather than a disclosure, and it is not credited.

The absence is more consequential here than the size of the vendor suggests, because of who the buyer is. Home health agencies run on thin, regulated margins with reimbursement set by a payment model rather than by negotiation, so the cost per clinician per month is the number that determines whether the product is viable at all for a given agency. A buyer cannot begin that calculation without entering a sales process.

The unit question matters as much as the figure. Pricing per clinician, per visit, per assessment completed or per agency produce very different economics for an organization with a large part time and per visit workforce, which is normal in home health.

Home health only, and deep within it.

The product covers the comprehensive assessment rather than a summary of the conversation, and it names the instruments rather than the specialty: the OASIS including the M1800 series and the GG functional scores, the Plan of Care, wound measurements and medication reconciliation, alongside routine visit notes. Naming the instrument is the difference between domain depth and a template library, because an instrument has defined items a product either completes or does not.

It also targets the specific failure modes a general scribe hits in this setting rather than describing the setting in the abstract: reading a list of fifteen medications from bottles, handling regulated forms, and capturing wound measurements. Those are the tasks that make home health documentation hard, and addressing them individually is evidence of real work rather than repositioning.

The workflow suits the setting too. Referral summarization before the visit, prompting during it and dictation afterward match a clinician who arrives cold at a house with a hundred page discharge packet behind them.

Capture in multiple languages is supported, with no language count published.

Coverage is one care setting in one country, with no hospice, no facility based post acute care and nothing beyond home health. That narrowness is a deliberate and coherent choice rather than a gap, and it is also the whole footprint: an agency running both home health and hospice lines would be covered for one of them.