Andy AI

Capture, chart drafting, the OASIS itself, wound and medication documentation, and the QA pass over the full history and physical are all model generated, and coding is described as AI powered. The human contribution is described as review and confirmation rather than production: clinicians confirm the documentation, and coding is fully reviewed by home health coders. What the vendor does not state is how much those coders change. If reviewers are substantively rewriting coding output rather than verifying it, that part of the product is a service with an AI front end.

A review gate is described at more than one point. Documentation is drafted by AI, confirmed by the clinician, then synced to the record; a QA layer cross checks the visit against the history and physical and prior visits; and coding is described as fully reviewed by home health coders.

The companion app states that it does not generate an OASIS on the device and that the agency must connect its record system for the completed chart to appear, so nothing reaches the record through an unsanctioned path. Still, no reviewer credential or QA coverage is stated, no threshold or abstention behavior is published, and it is not stated whether the QA layer reviews every chart or a sample.

The vendor publishes no model card, named model, accuracy methodology or evaluation design. The performance claims are stated as outcomes rather than as measurements: 45 percent more detailed charts, 24 percent more patient data captured, 45 percent more conditions identified, double the clinician productivity. Each of those needs a denominator and a comparator to mean anything, and none is given. A third party review site describes the accuracy as superhuman, which is that site's language rather than the vendor's, and is not evidence.

Nothing the vendor publishes identifies any party in the chain: no model or model family, no foundation model provider and no subprocessor list. The app store listing says captured data uploads to a HIPAA compliant cloud, which names a property rather than a provider. Providers may be named in the privacy policy, which sits behind a third party viewer.

The architecture makes the omission concrete. The vendor states the assessment is not generated on the device, and that an agency must connect its record system to the vendor's cloud to receive completed output, so capture, upload and generation are separate stages, each involving parties that are not named. Coding is described as reviewed by home health coders, a human party in the chain whose employment and location are not stated.

The content type raises the stakes. This application captures images as well as audio, which in home based care means wound photographs taken inside a patient's home. A photograph may capture the patient's body, their living conditions and other people present, and it persists as an image rather than as text that could be redacted. Nothing published states how long photographs are held, whether they are stored separately, or whether they are excluded from any model improvement.

Andy publishes no study, controlled evaluation, third party assessment or named customer agency, and no deployment count of any kind. Every figure is vendor reported and uncited: one hour saved per start of care chart, 90 minutes reduced to a 15 minute review, export in under five days, a 7 percent case mix increase, double productivity.

The funding and founder credentials are verifiable through Y Combinator, Crunchbase and Dealroom and are real, but investor validation is not clinical evidence. The company also declined to submit products for a CB Insights analyst briefing, which is neutral in itself but means no third party analyst view exists either.

The only published commitment is that captured data uploads to a cloud described as meeting health privacy requirements. There is no retention schedule, no audio deletion commitment, no deidentification practice and no statement on whether customer content is used to train or improve models.

Two things make that gap larger here than for a clinic based scribe.

The application captures images as well as audio. In home health those are wound photographs, and they are taken inside a patient's home. A photograph of a wound is identifiable in ways a transcript is not, it may capture the patient's body, their living conditions and other people present, and it persists as an image rather than as text that could be redacted. Nothing published states how long photographs are held, whether they are stored separately from the note, or whether they are excluded from any model improvement.

Second, the mobile application's store disclosures indicate that contact information and user content are collected and linked to user identity. Those labels are vendor declared summaries rather than a policy, and they describe the app rather than the platform, so they settle little on their own. What is linked to whom, and for what purpose, is not stated anywhere beyond the store listing.

The processing architecture makes the questions sharper. The vendor states the application does not generate the assessment on the device, and that an agency must connect its record system to the vendor's cloud to receive completed output. So capture, upload and generation are separated, and each stage has its own retention question.

HIPAA compliance is claimed on the vendor's homepage and in the app store listing, which describes uploads to a HIPAA compliant cloud. No business associate status is stated, and no agreement terms, availability or scope are published. The privacy policy sits in a third party policy viewer and may say more.

The agency rather than the individual clinician is the contracting party, which means an agreement conversation normally precedes use.

The stakes are specific to the setting. Clinicians capture audio and wound photographs inside patients' homes on phones, and the agency remains the covered entity for everything captured.

Controls are asserted and nothing independent stands behind them. The site describes the product as HIPAA compliant and the app as uploading to a HIPAA compliant cloud, which is a regulatory claim rather than an attestation. No certification, audit report, penetration testing statement or trust center is published, and the site footer links only resources, a blog, the privacy policy, support and a data rights page.

The exposure profile is what makes the absence matter. This is a mobile first product used outside any controlled environment. Clinicians carry it into patients' homes, capture audio and photographs there, and upload from whatever connection is available, on personal or agency issued phones rather than managed workstations.

An independent examination is the mechanism that tells a buyer how that surface is governed: how devices are enrolled and revoked, what is retained on the handset before upload and for how long, what happens when a phone is lost with unsynced visits on it, how access is provisioned for agency staff, and whether it is logged in a way the agency can audit.

The company is young and venture backed, and a full attestation program is a material cost at that size. The gap is in what a counterparty can verify before contracting, not a sign that controls are absent.

The Medicare payment and quality framework the product operates in is real, but it is not a regulatory pathway for the software. No clearance is claimed and none is required for documentation a clinician reviews and confirms, and the vendor states that documentation is confirmed by clinicians before it reaches the record.

The framework that actually governs the output is the OASIS, which drives both the quality reporting behind agency star ratings and the case mix weighting behind episode payment, so failures land in survey findings and payment integrity rather than in device safety.

One issue is subtler than the usual coding concern. The vendor has said its system scores OASIS functional items on what is safe for the patient rather than on what the patient currently does. That is the correct convention, and clinicians commonly score habitual activity instead, so applying the rule is a real quality improvement. It also runs one way: scoring on safe ability produces lower functional scores, lower functional scores raise case mix weighting, and the vendor reports a 7 percent higher case mix weight while marketing capture that justifies higher functional impairment. The interpretation the product applies is simultaneously the more accurate one and the more remunerative one, and nothing published separates corrected under scoring from drift.

The central issue is one number: agencies using Andy are reported to see a 7 percent higher case mix weight, which the vendor markets as capturing every clinical detail to justify higher functional impairment and comorbidity adjustments. Case mix weighting determines what a 30 day home health episode pays, so this is the home health instance of the question every documentation tool tied to payment raises: whether more complete capture is also more remunerative capture, and who checks the difference.

There are real counterweights. Andy frames its notes as accurate and defensible, requires clinician confirmation before documentation reaches the record, has coders review coding, and the underlying claim is that more complete capture of conditions actually present produces a more accurate case mix, which under a system that pays for documented clinical complexity is a legitimate argument rather than only a commercial one. There is no contingent pricing and no human removed.

Separately, no fairness statement, subgroup analysis or accent and dialect disclosure is published, which matters in a segment where the workforce and the patient population are both linguistically diverse and the recording happens in the patient's home.

The vendor describes review before commit: documentation is drafted by AI and confirmed by clinicians before it syncs to the record, a QA layer cross checks the visit against the history and physical and prior visits, and coding is described as fully reviewed by home health coders. That is a real mechanism for challenging an output.

Nothing stands behind the output. There is no accuracy or error figure, no published limitation, and no warranty, indemnity or remediation commitment. The published results are outcomes rather than measurements: a 7 percent higher case mix weight, visit to export in under five days, and twice as many patient visits, measured across agencies using the product with a caveat that results vary.

The case mix figure deserves separate attention because the vendor presents it as capturing every clinical detail to justify higher functional impairment and comorbidity adjustments. That may mean the system captures what a rushed clinician would have missed, which is the intended reading, or that it documents impairment and conditions the clinician did not assess. In home health the distinction is not academic, because documented functional status and comorbidities drive episode payment, and nothing published shows which is happening.

The architecture is write back rather than transfer. The completed chart including the OASIS is generated into the agency's record system, the agency connects that system to Andy's cloud as a deployment step, and the companion app deliberately does not produce an OASIS on the device, so the record of truth stays in the agency's system. That is the right shape.

The gap is that no record system is named anywhere. The homepage says only that Andy works with major home health EMRs. For an agency running WellSky, MatrixCare, HCHB, Axxess, Devero or Netsmart, whether its system is supported today is the first question, and the answer is not published.

No hosting region, residency option or subprocessor list is published, and nothing says whether a third party model service processes the encounter or what it retains.

The architecture is split in a way that matters. The vendor states plainly that the application does not generate the assessment on the device. Capture happens on the phone, data uploads to the vendor's cloud, generation happens there, and completed output reaches the clinician only once the agency has connected its record system to that cloud. Processing is server side by design and the mobile application is a capture and upload client.

That makes offline behavior a first order question rather than a detail. Home health visits happen in patients' homes, frequently in rural areas and inside buildings with poor reception. Nothing published states what the application does when the signal drops mid visit: whether capture continues locally, what is held on the handset until a connection returns, how long it stays there, whether it is encrypted at rest on the device, and what happens if the visit is never synced. That is both an operational question and a security one.

No price published, and access runs through a demo. One commercial disclosure is useful: the vendor states that scribing, QA and coding together cost less than the cost of outsourced QA or coding, which anchors the buyer to a cost line they already carry rather than to a per clinician software rate. That is a benchmarkable claim even without a number, and it tells a buyer what is being displaced. The unit of charge and the drivers of the quote are not published.

Narrow by design and deep within it. Home health only, spanning start of care through discharge, covering both nursing and therapy disciplines with therapy evaluations and exercise documentation captured during treatment. The depth shows in the detail: it names the instruments rather than counting specialties, working to the OASIS and PDGM and organizing the product around start of care as the pivotal chart.

Wound documentation via photograph and medication list reconciliation are the two field specific capabilities. The limits are a single care setting in a single country with no language breadth described, and home health only rather than the wider post acute picture.