Lime Health AI

Ambient capture and structured assessment generation are the product, and the company's stated trajectory runs outward from the scribe into a full AI native EMR for home health, on the argument that every ambient visit already produces the structured data an EMR stores. The certified coder review layer is verification of model output rather than a services business the AI was added to.

The oversight architecture here is complete, and the word that carries it is default.

A certified home health coder verifies every assessment item, diagnosis code and visit detail before anything syncs to the record system, and the company states explicitly that this review is standard with every engagement rather than an upsell. That distinction does the work. Human verification sold as a tier is a control the buyer may decline, may decline for cost reasons, and may decline without understanding what they are declining; verification included by default is a property of the product rather than of the purchase order. A control with known universal coverage is worth considerably more than one whose coverage depends on what was bought.

Three further elements complete it. The reviewer is a credentialed domain specialist rather than a general scribe, which matters when the artifact is a federal assessment instrument that drives payment grouping rather than a narrative note. Real time quality checks flag missed assessment sections and weak narratives while the clinician is still in the patient's home, so gaps are closed at source rather than by a return visit or a later query. And the review is audit logged, producing a record of what was checked and when.

Coding is positioned explicitly as suggesting codes for review rather than replacing judgment, which is consistent with the rest of the design rather than an isolated assurance.

One consequence comes with it: universal review means a human workforce reads every chart by design. That is the cost of this architecture and the two should be evaluated as one decision rather than separately.

ICD-10 suggestions are delivered with clinical evidence mapping, a useful architectural feature that shows which parts of the documentation support each code. That is the same traceability principle as Abridge's Linked Evidence, and it is rarer in coding than in note generation.

Against that, no accuracy figure, model card, named models or evaluation methodology is published, and the vendor's own time saving claims are inconsistent, appearing as a 75 percent reduction in documentation time in one place and 50 percent in another.

One tier of the chain is established explicitly and the other is not described at all.

What is established, and stated as a design property rather than disclosed reluctantly, is that a certified home health coder reads every chart. Because verification is universal rather than optional, a human workforce sees one hundred percent of encounters this product touches, not a subset the buyer selected. So the answer to who else sits between the record and the answer is known in kind: people do, on every chart, by design. An audit logged review trail records that the access occurred.

Three things are known: a human tier exists, its coverage is total, and the reviewers are credentialed in the relevant discipline.

What is not published is who those people are. Not the employing entity, not whether coders are employees or contracted, not their location or whether review happens onshore, not how they are vetted, and not what they see beyond the coded fields. An audit trail records that access occurred without telling a buyer who holds it.

The technical tier is unaddressed. No model or model family, no foundation model provider, no hosting arrangement and no subprocessor list is published.

The human tier is at its maximum here, and that is the price of the product's oversight strength. The two are one decision. An organization that wants every chart checked by a credentialed coder is also accepting that every chart is read by someone outside it, and should know who.

Third party review presence on G2 including a named medical coder and OASIS reviewer, which is a venue the vendor does not control. Beyond that, claims of two to three hours saved daily and elimination of after hours charting carry no denominator, cohort or method, and no study, controlled evaluation, accuracy benchmark or named agency customer is published.

The audit logged review trail is a real control, and a signed business associate agreement is stated plainly rather than implied.

One tension is sharp here, and it is the direct cost of the oversight architecture. Because coder verification is universal rather than optional, a human workforce reads one hundred percent of charts rather than a selected subset. Every encounter this product touches is read by a person outside the organization as a matter of design, not as an exception or a quality sample.

That is not a criticism of the design. It is the mechanism that makes the oversight worth what it is worth, and choosing this product means choosing both halves. The point is that the two should be evaluated as one decision rather than credited separately, and that patients are not told about the second half by anyone in the chain.

What is not published is the part needed to assess it. Nothing published says where that coder workforce sits, whether reviewers are employees or contracted, or what they can see beyond the coded fields. Nothing addresses audio or transcript retention, deidentification practice, or whether customer content is used to train models. For a product whose review step generates a continuous stream of human corrections to machine output, the training question is pointed: corrections are exactly the material a model developer would want.

States HIPAA compliance and a signed business associate agreement directly, alongside audit logging. Specific terms are not published for inspection, and nothing published says whether the agreement covers the certified coder review workforce or only the software.

The vendor publishes no SOC 2 report of either type, no HITRUST certification, no ISO 27001, no penetration testing statement and no trust center. Enterprise grade security remains a marketing phrase rather than a standard.

The service model, rather than the software, makes that absence matter more here. Every note this product generates is reviewed by a credentialed home health coder before it reaches the customer's record system. The vendor presents that as standard with every engagement rather than as a paid upgrade, and as a quality control it is a strength.

It also means the vendor employs or contracts people whose job is to read patients' clinical documentation at scale, routinely and by design. That is a materially larger internal access surface than a pure software vendor carries, and it is precisely the territory an independent examination covers: how staff are vetted, how access is provisioned and revoked, what a reviewer can see beyond the note in front of them, whether that access is logged, and what happens when a reviewer leaves.

None of it is answerable from published material. The coding credential held by reviewers speaks to their competence at coding and says nothing about the controls governing their access to patient information.

No clearance claimed and no United States device pathway attaches. A real and demanding regulatory framework governs this product. It simply is not device regulation.

The instruments involved are federally mandated. OASIS-E2 is the home health patient assessment, HOPE is its hospice counterpart, and the patient driven groupings model is the methodology home health is reimbursed under. Accuracy failures surface as survey findings, claim denials and audit exposure rather than as device incidents, and CMS assessment and payment policy is the governing regime.

That produces a sharp incentive structure. A scribe suggesting billing codes shapes reimbursement for care already delivered; here the same answers serve as the clinical assessment, as the payment case mix, and through publicly reported quality measures as the agency's own quality score. One artifact drives three outcomes, and software populating its items determines all three at once. Nothing says which direction the system resolves ambiguity where an assessment item could defensibly be scored two ways.

HOPE deserves separate attention because it is recent. It replaced the previous hospice assessment instrument in 2025, so any tool completing it works against a young instrument carrying less accumulated interpretive guidance than OASIS has. Nothing says how the vendor keeps pace as that guidance develops, or what happens to documentation generated against a superseded version of either instrument.

The coding leans toward reimbursement, but with strong counterweights.

The reimbursement language is present: ICD-10 codes are described as validated for PDGM optimization, and PDGM grouping determines what an episode pays, so optimization is not a neutral word.

What sits against it is unusual. Codes are mapped to the clinical evidence supporting them, so a reviewer can see why each was suggested. Coding is stated to be a co pilot rather than a replacement for human judgment. And a credentialed coder verifies every one before submission, which puts an independent professional between the model's suggestion and the claim.

The vendor also publishes a decomposed accuracy claim. Structured assessment items are said to approach 99 percent because the model has explicit field mappings, while narrative sections approach 95 percent on first generation. Stating a mechanism for why two parts of the same output differ is a real disclosure habit.

It still falls short in two ways. The decomposition is by task type rather than by population, so it shows how performance varies across output formats and says nothing about how it varies across speakers and settings. And the figures are vendor generated, with no methodology, denominator or independent validation published, so they are claims rather than measurements.

No fairness statement and no performance breakdown by speaker characteristic is published, and language coverage is English and Spanish only. For a workforce visiting patients in their own homes, frequently elderly and frequently with speech affected by illness, that is the breakdown which would matter most.

Two mechanisms here work together. The first is traceability applied to coding: diagnosis code suggestions are delivered with clinical evidence mapping showing which parts of the documentation support each code. That converts a coder's review from a trust exercise into a checkable one.

It matters more in coding, because a wrong code shapes what is billed and creates false claims exposure for the organization submitting it, and a coder who can see the supporting documentation can refuse a suggestion on the evidence rather than on instinct. The second is that human coder verification is universal rather than optional, so the correction step covers every chart. A correction mechanism with known coverage is worth considerably more than one with unknown coverage.

An audit logged review trail and a signed business associate agreement complete the position. Nothing stands behind the output. No accuracy figure or evaluation methodology is published, the vendor's own time saving claims are inconsistent at 75 percent in one place and 50 percent in another, and no warranty, indemnity or remediation commitment is stated.

Effectively complete coverage of one segment's EMR market, with the depth that matters. As of mid 2026 Lime integrates bidirectionally with WellSky, MatrixCare, HCHB, Axxess, DSL and Netsmart, which between them account for the overwhelming majority of home health and hospice agencies.

More importantly it syncs structured OASIS fields and care plans rather than pushing narrative text, which is the difference between a scribe that saves typing and one that completes the regulatory instrument. Verified notes sync ready for billing with, in the vendor's phrasing, zero rework.

For this product, residency is a different question from the usual one.

The conventional gaps remain. No hosting region is stated, no residency option is offered, and no subprocessor list is published. Nothing establishes whether encounter audio or transcript is processed by a third party model service, which provider that would be, or what it retains. Delivery is cloud based, with capture through a mobile application used inside patients' homes and native integration into the major post acute record systems.

The material fact here is not architectural. Every note passes through a credentialed human reviewer at the vendor before it reaches the customer's record system. Data residency is normally a question about which jurisdiction's servers hold the data and therefore whose authorities can reach it. Here there is a second and more immediate answer: the documentation is read by identifiable people working for the vendor, and where those people sit is a residency fact in its own right. A note reviewed by a person in another country has crossed a border regardless of where the database lives.

That is not a criticism of the model, which is a real quality control and is offered as standard rather than as a paid extra. It is a scoping point that no architecture diagram shows, and an easy one to overlook.

Nothing published says whether the hosting region can be pinned by contract, which model provider processes the encounter, or where the review workforce is located. The last of those may matter more than the first two.

No rate card, tier structure or pricing model is published. The mobile application is free to download but the platform is sold to agencies through a demo led process. The price necessarily includes human coder review of every chart, so it does not compare directly with software only per clinician rates elsewhere in the market.

Deep in one setting, and specialized at the level of the regulatory instrument rather than the note template. Coverage spans home health and hospice for nurses, therapists and field clinicians, generating OASIS-E2 assessments, HOPE hospice evaluations, daily visit notes, care plans and ICD-10 coding, with PDGM and homebound status documentation handled as first class concerns.

The vendor's own argument is the right one: an ambient tool that produces only SOAP notes does not serve this setting, because the document that matters is a structured CMS assessment. English and Spanish. Mobile and iPad first, which suits clinicians documenting in patients' homes rather than at a workstation.