SmarterDx

The model is the labour, notwithstanding that a human validates every output. Second level review of every chart at more than 30,000 data points each is work no coding department performs today, and it is not performed at all without the model. Remove it and there is no workflow layer, network or platform left that a customer would pay for.

The company's framing invites the opposite reading and should not be taken at face value on this axis. Its stated position is empowering documentation and coding teams rather than replacing people, which is a claim about who decides rather than about who does the work. The engine still reads the entire record, forms the clinical judgement about what is missing, and produces the finding; the human confirms or rejects it. That is model centrality with human authority over the output, and the two are different questions.

The autonomy boundary is the clearest in the index because there effectively is not one. Nothing is submitted by the vendor. Findings are surfaced for the customer's clinical documentation and coding teams to validate before final billing, and the company states this as a design principle rather than a limitation. A buyer evaluating this alongside Fathom is choosing between opposite philosophies rather than different automation rates.

Oversight is correspondingly complete. Every finding is stated to be fully attributable and auditable back to evidence in the chart, which is what makes validation possible rather than nominal, and the company positions defensibility as the product rather than throughput.

What is missing is the number that matters for a tool of this shape. No precision rate, false positive rate or acceptance rate for surfaced findings is published anywhere. For an autonomous coder the question is what share of charts it handles correctly; for a finding surfacer the question is what share of its findings a validator accepts. A low acceptance rate wastes scarce documentation specialist time and a systematically permissive one creates compliance exposure, and neither is disclosed. Ask what percentage of surfaced findings customers accept on review.

The training data disclosure is the most specific in this part of the index. Models are stated to be trained on more than 21 million patient encounters of real record system data, and the company draws the contrast explicitly against competitors it characterises as trained on limited, non clinical or publicly available data. Scale and type of training corpus are exactly what this axis asks for and almost nobody states.

Processing is described concretely: the full clinical record at more than 30,000 data points per chart, with a stated commitment to review every chart rather than prioritising a subset. That last point is a real design disclosure, because prioritisation determines which patients get scrutiny and the company forgoes it.

The claim to understand clinical reasoning, paired with complete auditability and fully attributable findings, is the architectural position. What sits behind it is unstated: no foundation model, model class, version or architecture is named, and no precision, recall or accuracy figure of any kind is published. Implementation is stated at under eight weeks.

The training corpus is disclosed and its origin is not, which is an unusual shape for this axis and lands at the top of the C band.

More than 21 million real patient encounters is a substantial and specific provenance statement. It also raises the question it does not answer: whose encounters, obtained under what agreements, de identified by what standard, and whether customer charts processed under service contracts contribute to the shared model that other customers benefit from. A vendor whose differentiator is the scale of its clinical training corpus has a stronger obligation to explain how that corpus was assembled than one training on public data, and nothing published does.

On parties there is little. A public cloud provider appears as a badge, which indicates a hosting relationship without disclosing its scope. No foundation model provider, model class or version is named, and no sub processor list was located. The acquisition of a note generation company in September 2025 folded another technology stack into the platform whose components are not separately described.

Ask how the 21 million encounter corpus was sourced and consented, whether customer data contributes to the shared model, and for a sub processor list.

The best reference set in this part of the index, and a central claim that resists verification.

Eleven health systems are named as current clients, including large multi state operators and academic medical centres. Three case studies are attributed to named executives with titles at named institutions, which is materially stronger than the anonymised role and bed count references used by CombineHealth and RapidClaims. A chief financial officer at McLaren Health is quoted reporting more than 11 million dollars in annualised net new revenue alongside review of 100 percent of clinical data across 100 percent of charts. Third party signal includes a KLAS client satisfaction score of 98 and an innovation award listing.

The caveat is structural rather than a matter of missing paperwork. Net new revenue found is a counterfactual: it measures money the organisation asserts it would not otherwise have captured, which cannot be observed directly and depends on assumptions about what the prior process would have caught. A 5 to 1 return and 2 million dollars per 10,000 discharges rest on the same construction. No independent audit, controlled comparison or peer reviewed evaluation of that counterfactual exists, and a satisfaction score measures how clients feel rather than what the tool found.

Held at B because the reference set is exceptional and the headline metric is unfalsifiable as published. Ask how baseline capture was established at a reference site.

The credentials imply a control programme and the stewardship questions are untouched.

A controls report and a state government cloud authorisation both require assessed security controls, and the latter is granted only after review of a defined cloud service, so a programme exists and has been examined by outside parties. That is more external scrutiny of the hosting environment than most vendors in this part of the index can show.

None of it answers what this axis asks. There is no retention schedule, no data ownership or deletion statement, no access control detail, no encryption specification, no data minimisation commitment, and no incident or breach disclosure.

The unanswered question that matters most is specific to this vendor's differentiator. Models are stated to be trained on more than 21 million real patient encounters, and the product ingests the complete record for every discharge at a large number of health systems. Whether customer documentation processed under service contracts feeds that shared corpus, and under what de identification standard, is the single most consequential stewardship question on this record, and nothing published addresses it in either direction.

Health privacy appears as a badge with nothing behind it. No control enumeration, no de identification statement, no access control description, and nothing comparable to the itemised programme CombineHealth publishes on a dedicated page.

What sits underneath is real but general rather than health specific. A controls report and a state cloud authorisation both cover domains that overlap substantially with the health security rule, and the state programme in question accepts health framework assessments as evidence, which situates the two frameworks relative to each other without the company holding the health one. No certification against a health specific framework was located, unlike two competitors in the adjacent coding lane.

No business associate agreement posture, template, negotiation stance or execution requirement was located, which is a notable gap for a vendor contracting with academic medical centres and multi state systems and processing complete inpatient records for every discharge.

Graded C because externally assessed controls exist and the health specific layer is asserted rather than evidenced.

A credential most healthcare vendors do not hold, alongside one that is stated too loosely, and no trust center.

The state government cloud authorisation is the notable one. It is issued by a state authority rather than self declared, granted against a defined cloud service, time bound, and required for vendors serving that state's public institutions, which makes it a procurement gate rather than a marketing badge. No competitor built in this sweep holds a government issued authorisation of any kind, and it is a meaningful differentiator for public university systems and state facilities.

The controls report is presented as a certification without a report type. That framework produces an attestation rather than a certification, and the distinction between the two report types is the whole question, since one tests design at a point in time and the other tests operating effectiveness over a period. XpertDox and CombineHealth both state their type; this vendor does not.

No trust center exists as a standing page. The badges sit on the home page with no supporting detail, no report availability or request process, no audit period, auditor or date, no penetration testing disclosure and no vulnerability disclosure policy. Graded B on the strength of the government authorisation, held below A by the untyped report and the absent trust center. Ask for the report type and period, and the authorisation level and expiry.

No device pathway applies and none is claimed. Identifying documentation gaps and revenue opportunities from a clinical record is an administrative determination rather than a clinical one, so the absence of a clearance is correct and is not a gap.

The regulatory weight sits elsewhere and is heavier here than for a straightforward coding engine. The product surfaces missing diagnoses on inpatient charts, and added diagnoses are precisely what shifts payment weight in the inpatient classification system. Documentation improvement that moves payment groupings is among the most actively examined areas in federal programme integrity work, where the question is whether an added diagnosis reflects care actually delivered and documented or a grouping optimised after the fact.

The company's own design is its best answer: findings are attributable to chart evidence and validated by the provider's clinical staff before billing, which puts a licensed human between the model and the claim. That is a genuinely strong posture and it is a posture rather than a control framework. Nothing published describes the documentation standard applied to a surfaced diagnosis, audit sampling of accepted findings, or what happens when a payer challenges one. Graded C because the position is correctly represented and undocumented.

Two design commitments do real governance work and are stated plainly.

Every finding is attributable to evidence in the chart and fully auditable, which means an accepted finding arrives with its own justification attached rather than as an assertion a reviewer must reconstruct. And the company commits to reviewing every chart at every data point rather than prioritising a subset, explicitly contrasting itself with vendors that prioritise. Prioritisation is a quiet source of bias in this category, because it decides which patients receive scrutiny, and forgoing it removes that failure mode entirely. Combined with human validation of every output, the case level governance here is stronger than anything else in this part of the index.

The population view is missing and the commercial direction makes it matter. The product exists to increase captured revenue, and its output is a recommendation to add diagnoses. No distribution of accepted findings against an expected benchmark, no breakdown by specialty, payer, physician or documentation style, no acceptance or reversal rate, no bias or fairness testing, and no external audit of finding validity is published.

Held at B because the instruments that exist are genuine and unusually well suited to the risk, and no result from them is disclosed.

No performance figure of any kind is published, which is the condition that puts a record here regardless of how good the product is.

There is no accuracy rate, no precision or false positive rate for surfaced findings, no acceptance rate on validation, and no confidence threshold. Without one, there is no stated level against which a shortfall could be measured, and therefore nothing to hold the vendor to. No service level agreement, warranty, indemnity or remediation commitment was located.

The architecture genuinely limits how much recourse a buyer should expect, and it also relocates the risk rather than removing it. Because a clinician or coder validates every finding before billing, the provider owns every submitted claim outright and the vendor never touches a payer. That is defensible design and it means the false claims exposure created by an accepted but unsupportable diagnosis sits entirely with the health system. The defensible framing in the company's marketing refers to the auditability of the evidence trail, not to the vendor standing behind the finding.

One pre emptive note for future passes: further return on investment or revenue figures cannot move this grade. A 5 to 1 return and a risk free framing are commercial outcome claims, not accuracy or recourse measures, and the underlying revenue figure is itself a counterfactual. Only a published finding precision measurement, an acceptance rate, or a contractual commitment on either will change it.

The weakest axis on this record relative to the company's size, and the gap is disclosure rather than capability.

No record system is named anywhere in the published material. No integration standard is described, no interface mechanism is specified, no vendor marketplace or programme listing was located, and no statement addresses whether the connection is bidirectional or read only. Competitors a fraction of this company's size enumerate ten or twelve platforms by name and specify their standards.

What can be established is indirect and is not treated as evidence here. The named clients run enterprise record systems, the product ingests more than 30,000 data points per chart across every discharge, and implementation is stated at under eight weeks, all of which imply mature and deep integration in practice. Funding materials reference adding new clinical and financial data integrations. Inference from customer lists is not disclosure, and this record does not grade it as such.

Graded C because a buyer cannot determine from published material how the product would connect to their environment, what the integration burden is, or whether findings write back. Ask which systems are supported, through what standards, and whether any integration is certified by the record vendor.

One genuine signal, and the specifics still absent.

A state government cloud authorisation is the most useful thing here, because that programme certifies a defined cloud service rather than a company, is granted for a fixed term, and requires the provider to describe its deployment model and security measures to the assessing authority. Its existence therefore establishes that a documented deployment description was submitted to and reviewed by an outside body. A major public cloud provider also appears as a badge.

What is not published is what a buyer needs. No processing or storage region, no tenancy model, no residency commitment, no customer controlled or single tenant option, and no statement on whether the authorisation covers the whole platform or one service. The badge indicating a cloud relationship is a logo rather than a hosting disclosure and is not treated as one here.

The volume makes the question live rather than formal. Complete inpatient records for every discharge across eleven named health systems move for processing continuously. Ask for the regions, the tenancy model, which services the state authorisation covers and at which level, and whether that certification is current.

No price is published and the economics are unusually modellable, because the company supplies both an instrument and a unit driver.

A public return calculator is keyed to annual discharges grouped by payment classification, which is the correct denominator for an inpatient product and lets a buyer estimate against a figure they already know precisely. Around it sit a stated 5 to 1 return from day one, an average of 2 million dollars in net new annual revenue per 10,000 patient discharges, an average of 3.5 million dollars realised annually, implementation in under eight weeks, and 100 percent client retention.

One thing the disclosure implies without confirming deserves a direct question. A vendor publishing a fixed return ratio as proven, and describing its return as risk free in funding materials, is signalling that price is tied to value delivered rather than set as a flat licence. If the commercial model is a share of identified or realised revenue, that has very different budget and audit consequences from a subscription, and it is not stated anywhere.

Held at B rather than A because the unit of charge is absent, and the four other product lines beyond pre bill review carry no economic disclosure at all. Ask whether pricing is contingent on realised recovery, and what the risk free framing means contractually.

Depth in inpatient acute care, breadth across revenue cycle stages, and effectively nothing outside the hospital.

The setting is unambiguous and the evidence corroborates it. The return calculator is keyed to discharges grouped by payment classification, the named clients are health systems and academic medical centres rather than physician groups, and the product logic centres on diagnoses that change inpatient payment weight. Within that setting the coverage runs across the whole cycle: authorisation and medical necessity before care, documentation and charge capture around the encounter, and denials and underpayment recovery after the claim.

What is absent bounds the record clearly. No ambulatory or professional coding coverage is described, no specialty enumeration exists of the kind RapidClaims publishes, and nothing addresses coding regimes outside the United States. This is a hospital product and it does not pretend otherwise.

Held at B rather than A because coverage is described by revenue cycle stage rather than by clinical specialty, so a buyer cannot tell whether performance in, say, complex surgical or oncology documentation differs from general medicine, and the finding quality question on this record makes that variation material.