Arintra

The model does the work a coder previously did, from chart capture through to claim submission. Documentation arrives from the record system, large language models paired with clinical knowledge graphs interpret it in clinical context, and billing codes return with no human in the path on the majority of charts. There is no workflow layer, network or data pipeline underneath that would function if the model were removed, because the model is the labour.

One qualification belongs on the record for symmetry with the rest of this lane. Any autonomous coder must route the charts it declines, so a human coding capability sits behind the automation, and the company has extended the product into documentation improvement and denials work where the model advises rather than acts. Neither dilutes the core: the sold unit is autonomous code assignment.

The boundary of the automation is published as a number, which is the first thing this axis asks for. A third party marketplace listing states more than 82 percent of charts coded autonomously, with coders freed to handle complex cases, and the company describes coding running from chart capture to claim submission inside the record system with no human intervention on the automated share.

What separates this from the A grade in this lane is the absence of two things the category leader publishes. There is no stated accuracy figure anywhere in the company's materials, only the adjective high, so a buyer knows what share of charts the engine takes but not how often it is right on them. And there is no contractual service level agreement on automation rate, accuracy or turnaround, so the published automation rate remains a marketing figure rather than an enforceable obligation with a counterparty.

The oversight mechanism that is published is the explainable audit trail attached to every coding decision, which gives a reviewer a way to interrogate an individual assignment after the fact. That is real, and it is post hoc review rather than a confidence threshold governing what the engine hands off in the first place. Ask for the autonomy threshold and the accuracy figure behind it.

The mechanism is named more specifically than most of this category manages. Large language models paired with clinical knowledge graphs, applied to unstructured charts, interpreting clinical intent and patient progression across a visit rather than matching keywords. That is a genuine architectural statement and it distinguishes the approach both from the pure deep learning engines and from the computational linguistics route Nym Health takes.

Output scope is enumerated precisely: evaluation and management levels, CPT, ICD-10, HCC and HCPCS codes, with modifiers and units. Scope enumeration tells a buyer what still needs a coder and is rarer than it sounds.

Absent is everything below the mechanism. No foundation model provider, model class or version is named, no training data provenance or scale is described, and no accuracy figure is published at all, by coding element or in aggregate. The analytics dashboard is stated to report coding accuracy to the customer, which means the number exists and is measured; it is simply not disclosed publicly.

The company describes itself as generative AI native and states that large language models sit at the centre of the product, so the existence of a model supply chain is acknowledged rather than obscured. That is the floor, and it is where the disclosure stops.

On enumeration of parties there is nothing. No foundation model provider is named, no model class or version is given, no hosting arrangement is described, no sub processor list was located, and no position was found on whether customer documentation contributes to model development. The clinical knowledge graph raises a second unanswered question, since a curated ontology is either built in house or licensed and neither is stated.

The data crossing that unnamed chain is the complete clinical note for every encounter coded rather than an extract, and the company reports monthly coding volume increasing more than fivefold in a year, so both sensitivity and throughput are high. Ask for the base model, the sub processor list, the knowledge graph provenance and the training position.

Independent assessment exists and is named, which puts this above most early stage records in the lane. KLAS published an Emerging Company Spotlight on Arintra in 2026 examining customer experience with the autonomous coding solution, and the company reports a customer performance score of 93 out of 100 against a stated 2026 Best in KLAS software average of 81.1. Named customers include Mercyhealth, which described coder capacity outpaced by expansion past 200 care locations, and Med First. Endeavor Health Ventures invested as a customer, which is a weaker but real signal.

Two limits hold this at B. The KLAS report sits behind a subscriber login, so what is verifiable from outside is that the assessment happened and the score the vendor quotes from it, not the findings themselves. And no peer reviewed publication, prospective study or independently audited accuracy measurement exists.

A specific caution for anyone quoting the outcome figures: they do not agree across the company's own materials. Revenue uplift appears as five percent or more in most places and seven percent on a marketplace listing, and accounts receivable improvement appears as a twelve percent reduction in one release and a sixty four percent reduction in pre accounts receivable days in another. Those measure different things and are presented interchangeably.

The certification that exists was framed by the company and by the certifying body around cybersecurity and information protection rather than around data stewardship, and that framing matches its scope. Entry tier HITRUST e1 plus asserted SOC 2 and privacy posture tells a buyer that baseline security controls were assessed. It does not answer what this axis asks.

None of the stewardship specifics are published. There is no retention schedule, no statement on whether customer documentation contributes to model development or improvement, no de identification position, and no description of how the clinical knowledge graph is maintained or whether customer encounters inform it. That last question is particular to this architecture: a knowledge graph curated from customer charts is a materially different data proposition from one built out of published coding guidance, and the company does not say which it is.

The exposure is not small. The product ingests the complete clinical note for every encounter, and monthly coding volume grew more than fivefold in a year. Held at C rather than higher because the certification is genuine and entry tier, and rather than lower because the company names it with a date and a scope.

Compliance is stated by the company and, more usefully, is backed by an external certification rather than resting on the assertion alone. HITRUST e1 was awarded in April 2026, and the framework maps a certifiable control set onto health privacy requirements, so a buyer has something assessed by a third party to point at.

The tier matters and is stated plainly here rather than treated as a single badge. The e1 assessment is the entry level of the three tiers, covering a foundational control set, below the i1 that Fathom holds and well below the risk based r2. It is a real certification and the lightest one available.

No business associate agreement posture, template, negotiation stance or execution requirement was located publicly. Given the company contracts with health systems and processes complete charts, agreements clearly exist in detail; none of it is public.

A named certification carrying its tier and date, which is the disclosure this index consistently rewards. HITRUST e1, awarded April 2026, announced with the certifying body quoted and the scope described as an assessment of implemented security controls and operational maturity.

SOC 2 and privacy compliance are also claimed, and they are treated differently here. Both appear only inside a phrase describing an existing compliance posture, with no report type, no scope statement, no audit period, no auditor and no date. Under the credential test a certification counts when it carries a verb and a scope boundary, and these carry neither, so they are recorded as asserted rather than counted as evidence.

Held at B rather than A because no trust center, report availability statement, penetration testing disclosure or vulnerability disclosure policy was located in a dedicated pass. A prospective buyer should ask for the report type and audit period, which will move this grade in either direction.

No device pathway applies and none is claimed. Assigning billing codes from documentation is an administrative determination rather than a clinical one, so the absence of a clearance is correct and is not a gap in the record.

The regulatory exposure sits in claims submission rather than device law. Codes submitted to a payer are representations, and where they are wrong the framework is federal false claims enforcement, which lands on the billing provider rather than on the software vendor. The company positions the product around compliant reimbursement and states the audit trail supports compliance and clinician education, which is the right posture, and a posture is not an allocation of liability.

Graded C rather than higher because nothing published addresses how responsibility is apportioned when an automated code is later found incorrect, and rather than lower because the regulatory position itself is correctly and honestly represented.

Stronger than the category leader on the one governance question specific to coding, which is systematic drift. A model assigning evaluation and management levels slightly high across a population raises revenue and compliance exposure together, one assigning slightly low does the reverse quietly, and neither error announces itself in a single encounter.

Arintra publishes that its analytics dashboard reports evaluation and management distribution alongside upcode and downcode trends. That is precisely the aggregate view the drift question requires, and it is surfaced to the customer as a standing report rather than produced on request during an audit. Set beside the explainable audit trail at the individual decision level, the customer has both the population view and the case view.

Held at B rather than A because what is published is the existence of the instrument, not any result from it. No distribution of assigned levels against an expected benchmark is disclosed, no breakdown by specialty, payer or physician documentation style is offered, and no bias testing, fairness evaluation or model validation methodology is described. The company can see drift; the index cannot.

The weakest axis on this record, and the gap is structural rather than incidental. No accuracy figure is published at all, in aggregate or by coding element, so there is no stated performance level against which a shortfall could even be measured. No confidence threshold governing autonomous handling is disclosed. No service level agreement, warranty, indemnity or remediation commitment was located, and no denial or reversal rate for codes the engine assigned is published.

That matters more than the grade alone conveys, because codes submitted on a claim are representations to a payer and the provider carries the false claims exposure. A vendor coding more than 82 percent of charts without a human touch, publishing no accuracy figure and accepting no contractual commitment on one, has taken on the work without taking on any of the risk.

One pre emptive note for future passes: further marketing outcome figures cannot move this grade. Revenue uplift, denial reduction and accounts receivable improvement measure commercial benefit rather than accuracy or recourse, and the company's own numbers for those already conflict across sources. Only a published accuracy measurement, a stated autonomy threshold, or a contractual commitment on either will change it.

The strongest interoperability evidence in this lane, and it rests on a named program rather than an adjective. Arintra has been available through the Epic Toolbox since December 2024, which is a formal distribution listing with a review process and a scope boundary behind it, not a marketing claim about integration quality. That satisfies the credential test in a way that a third party calling an integration smooth does not.

Integration is bidirectional across three named record systems: Epic, Cerner and Athenahealth. The product reads clinical notes from the encounter and writes direct to billing charges back with no manual retyping, so it operates inside the record system rather than beside it, and a customer reference specifically cites the platform running within Athena from day one. Addressing both the documentation side and the claim submission side is the part most vendors in this category leave undescribed.

Graded A rather than B because the distribution credential is named and externally verifiable, three record systems are supported bidirectionally rather than one, and both ends of the pipeline are covered. What would strengthen it further is an interface standard statement, since no position on FHIR HL7 or other standards is published.

The company describes where the product sits in the workflow, inside the record system, and says nothing about where it sits in infrastructure. No hosting model, cloud provider, region, residency commitment, single tenant option or customer controlled deployment was located.

Volume and geography together make this a live question rather than a formality. Complete clinical notes for every encounter across a majority of ambulatory specialties move for processing, at a volume the company states grew more than fivefold in a year, while engineering operations sit in Bengaluru and the customers are United States health systems. Whether processing, storage or support access crosses a border is unanswered, and for a health system carrying residency requirements that is the first question rather than a detail.

Ask for the hosting arrangement, the processing and storage regions, and whether the offshore engineering team holds access to production data.

No price is published, and the economics are only directionally described. The company states it cut coding costs by nearly one third for customers, which is a percentage against a figure the buyer already knows precisely, namely current coding spend, and that framing is what this index credits when absolute pricing is withheld.

It stays at C rather than the B the lane leader earns because the claim is thinner in three ways. It appears in a growth and revenue press release rather than as a product economics disclosure, it is a single directional figure rather than a range with a stated average, and no pricing mechanism is described anywhere. Whether the model is per encounter, per chart, per coder replaced or a volume band changes the risk profile substantially, particularly for an organization whose volume is growing.

Retention is claimed as industry leading with no number attached, and thirteen enterprise deals in one hundred days is a sales figure rather than a pricing one. Ask for the pricing mechanism, the minimum commitment, and whether any portion of the fee is at risk against the automation rate.

Broad within one function and unusually explicit about where the coverage reaches. The company states autonomous coding across a majority of ambulatory care specialties including primary care, cardiology and pediatrics, plus urgent care, the emergency department, inpatient rounding, surgery, and the diagnostic specialties radiology and pathology. Coverage of the coding elements is complete rather than partial.

The buyer is a health system or physician group. Expansion out of ambulatory into inpatient rounding and the diagnostic specialties happened across 2025 and is recent, so depth in the newer settings is likely thinner than in the ambulatory core. A buyer whose volume concentrates in the emergency department or inpatient should ask for references in that specific setting rather than relying on the list.

Nothing addresses coding regimes outside the United States, which is a real boundary given that coding systems are national. The engineering presence in India is an operations fact, not a coverage signal.