EliseAI

Graded on the indexed HealthAI product line, where conversational AI across voice, text, email, and chat is the product itself. The company's property management business is out of scope for this record.

Escalation posture disclosed: agents hand off to staff when a more personal touch is needed, and the vendor reports 95 percent of inquiries handled without human intervention. Escalation criteria are described at a marketing level rather than in a formal certification or governance framework.

The least technical disclosure in this category. The platform is described as conversational and agentic AI without any account of what sits underneath: no model architecture, no base model, no description of the training corpus, no retrieval or grounding approach, and no benchmark results. That absence is more noticeable here than elsewhere because competitors in this category publish model family details, parameter counts, evaluation frameworks and accuracy figures.

One claim repeated across the healthcare pages deserves scrutiny rather than credit: patient data is described as 100 percent safe and accurate. No system is either, and an unfalsifiable absolute is weaker evidence than a measured figure with a stated method would be. The operational claim that the platform resolves around 95 percent of patient inquiries is more useful but is a containment statistic rather than a statement about the model. Worth asking what the system is built on, how responses are grounded, and what the measured accuracy is.

Compliance is asserted consistently, with two recognised standards named across the healthcare pages, and no specific technical control over protected health information is described anywhere. That gap is what holds the grade and it is visible only in context, which is why the comparison belongs on the record.

Competitors in this same segment publish concrete mechanisms: redaction of identifiers from conversations before they reach a model, identity verification before any protected detail is disclosed, an architectural boundary the data does not cross, or a named sub processor list. Here the claim rests on the certification alone, and a certification establishes that controls exist and were tested while saying nothing about what the company does with the data those controls protect.

Nothing states how long call recordings and transcripts are retained, whether patient conversations inform model development, what deletion is available on termination, or which sub processors sit in the voice path. That last one matters because a call typically passes through telephony, speech recognition and synthesis providers before reaching the model, so a health system that has assessed this vendor has assessed one party in a chain of four, and it cannot name the other three. Ask for the sub processor list, retention on audio as distinct from transcripts, the training position and deletion terms, and put all four in contract.

Operational claims are vendor and partner published: 95 percent inquiry containment, and a partner reported figure that only 15 percent of callers recognize the agent as AI. No independent validation retrieved.

Compliance is asserted consistently, with HIPAA and SOC 2 Type II named across the healthcare pages, but no specific technical control over protected health information is described anywhere. That gap is what holds the grade, and it is visible only in context: competitors in this category publish concrete mechanisms, whether redaction of identifiers from conversations, identity verification before any protected detail is disclosed, or an architectural boundary the data does not cross.

Here the claim rests on the certification alone. Nothing states how long call recordings and transcripts are retained, whether patient conversations inform model development, what deletion is available on termination, or which sub processors sit in the voice path, which matters because a call typically passes through telephony, speech recognition and synthesis providers before reaching the model. Worth putting retention, training use and the sub processor list into contract rather than relying on the compliance statement.

HIPAA compliance stated on product pages. BAA availability and execution terms are not published; a buyer must confirm BAA posture in the sales process.

SOC 2 Type II and HIPAA compliance stated on the HealthAI product pages, a cleaner certification claim than many peers. No dedicated public trust center with control level detail was retrieved.

Administrative automation rather than a clinical product, so no FDA clearance applies and none is claimed. The company describes its healthcare work as automating non clinical tasks, which is a sensible scoping statement, though it appears in press coverage and general positioning rather than as a published regulatory posture. Beyond that, nothing maps the platform to any framework a health system is measured against.

There is no reference to HIPAA Security Rule safeguards as a structure, to health sector cybersecurity performance goals, or to recognised practice standards, and no statement of what the system is prohibited from doing. Given the platform answers patient questions about coverage and payments and performs intake, the boundary between administrative response and advice is worth defining explicitly rather than leaving to the non clinical label. Worth asking for the written scope of what the agent will not answer and how it escalates.

No AI governance material was located: no responsible AI resource, no framework or certification, no model documentation, no human oversight model beyond handoff to staff, and no published accuracy, containment or escalation failure rates. The language question is the sharpest gap.

Multilingual capability is marketed prominently, with the healthcare pages citing support for more than fifty languages, while third party accounts describe voice in a far smaller number with written support across roughly fifty. Those are materially different claims, and no performance data is published for any language. A voice system that mishears an accent fails differently from one that mistranslates text, and neither is measured here.

This mirrors a pattern the index has found elsewhere: broad language counts marketed as a feature with no evidence of comparable performance across them. Worth asking which languages are supported in voice specifically, and for accuracy by language.

This is the least technical disclosure in the category, and the absence is more noticeable here than it would be elsewhere because the comparison is immediate: competitors in this same segment publish model family details, architecture compositions, evaluation frameworks and accuracy figures, so the silence is a choice rather than a category norm.

No architecture, base model, training corpus description, retrieval or grounding approach, or benchmark result was located, and no warranty, indemnity or remediation commitment. One claim repeated across the healthcare pages deserves scrutiny rather than credit. Patient data is described as one hundred per cent safe and accurate.

No system is either, and the construction is the one this index treats consistently: an unfalsifiable absolute is weaker evidence than a measured figure with a stated method, because it cannot be checked, cannot be missed and would be falsified by a single counterexample the vendor would then have to explain. It also does the specific damage this index has recorded elsewhere, since a buyer who accepts it has no reason to build the review step that would catch the exception.

The operational claim that the platform resolves around ninety five per cent of patient inquiries is more useful and is a containment statistic rather than a statement about correctness: it counts calls that ended without a human, not calls that ended well. Ask what the system is built on, how responses are grounded, and the measured accuracy on clinical intents.

Books, cancels, and reschedules directly in the EHR with real time updates, and integrates with practice management and revenue cycle systems. Third party coverage references FHIR based integration. Named EHR marketplace listings not verified.

Delivered as a vendor hosted cloud service that integrates with existing electronic health record, practice management and revenue cycle systems rather than replacing them. Nothing further is published. No hosting provider is named, no cloud region or residency commitment is offered, no tenant isolation model is described, and there is no customer hosted option.

The omission carries particular weight for a voice product, because the call path crosses telephony and speech providers before reaching the model and each is a place where audio and its transcript exist. It carries additional weight here because the same underlying platform serves a large housing business alongside healthcare, which makes the separation between those environments a question a health system should ask rather than assume. Worth establishing where calls are processed and stored, how long recordings persist, and how healthcare data is isolated from the company's other lines of business.

No pricing, pricing metric or contract structure is published, which is the norm in this category rather than an outlier. What distinguishes this record is where the quantified evidence sits. The company publishes specific, attributable outcome figures for its property management business, including large increases in lease conversion and resident engagement and a halving of delinquencies, and it offers a core housing product free of charge.

For healthcare the published evidence thins to a single headline claim, that roughly 95 percent of patient inquiries are resolved without waiting, with no methodology, deployment size or measurement period attached and no attributed customer results of the kind competitors publish. A buyer evaluating the healthcare business is therefore reading commercial proof drawn substantially from a different industry. Worth requesting healthcare specific reference customers with named outcomes, and establishing the pricing metric early since per call and per provider models diverge sharply with volume.

Real specialty depth exists where it has been built, and it is more than a landing page. Women's health coverage describes postpartum workflows, fertility consultation promotion and hormone therapy pathways, and orthopaedics is addressed separately, with third party accounts also citing surgery centres and rehabilitation facilities.

Channel coverage is genuinely broad, spanning voice, text, email and chat from one platform, and integration extends past scheduling into revenue cycle systems, which brings billing and payment conversations into scope rather than only appointments.

Held below the top of the band because the published specialty set is narrower than competitors who enumerate twenty or more, and because no coverage is described for the institutional settings that shape this market, with nothing addressing federally qualified health centres, children's hospitals or payers. Healthcare is also one of two verticals the company serves rather than its whole business. Worth asking which specialties have purpose built workflows today versus generic handling.