Charta Health

The economic claim depends entirely on the model. The company is explicit that its solutions are built on large language models that read any clinical documentation, extract meaning, and run custom analysis modules per note type, drawing an explicit contrast with rules based engines relying on NLP and user generated rules. The product only exists because reviewing 100 percent of charts pre bill was previously too costly to be realistic.

The oversight model is configurable and the company says so plainly, which is the right answer for autonomous coding. It states the platform can autonomously code every note, autocorrect coding mistakes directly in the EHR, or queue problem charts for human in the loop intervention, customized to the practice's preferred level of oversight.

Two supporting mechanisms matter: the AI provides direct citations to the underlying documentation so a reviewer can validate a suggestion rather than accept it blind, and coding opportunities are surfaced for simple approval or rejection.

More specific than most on approach without disclosing the stack. The company names large language models as the substrate, custom built with healthcare billing knowledge, and describes per client implementations built to replicate that client's own review criteria with an onboarding period of feedback until the AI follows their standards consistently. No model provider, evaluation methodology, or accuracy benchmark is published, and the customization model means performance is client specific rather than a single published number.

The approach is named and no party in it is. The company identifies large language models as the substrate, describes them as custom built with healthcare billing knowledge, and explains that implementations are per client and tuned to replicate that client's own review criteria over an onboarding period. That tells a buyer the shape of what is running and, usefully, that their own review standards become part of the configuration. What is not disclosed is who supplies the substrate.

No foundation model provider, model class or version is named, no hosting arrangement is published, and no sub processor list was located, and custom built with healthcare billing knowledge is a description of tuning rather than of origin, so it does not settle whether a third party base model sits underneath.

The data surface is total rather than sampled, since the platform reads complete clinical documentation for every encounter, and compliance reports are offered on request through a security contact rather than published. Two questions follow from the per client design: whether one client's review criteria or documentation informs another client's implementation, and what retention applies to documentation once a claim is finalised. Ask both, ask whose base model is underneath, and ask for a sub processor list.

Quantified, consistent across sources, and unusually specific about the metric used. The company reports average revenue increases of roughly 11 percent measured as RVUs per patient, with a stated ceiling of 15.2 percent, and reports healthcare organizations across 43 states using the software. Naming RVUs per encounter as the measurement basis is better practice than an unqualified revenue claim.

Buyers should note two things: figures are vendor reported without independent audit, and a platform whose value is measured in upcoded revenue carries an inherent tension that makes the compliance claim the one to verify hardest.

The platform reads complete clinical documentation for every encounter, so the PHI surface is total rather than sampled. The company states HIPAA, SOC 2, and GDPR compliance and offers compliance reports on request via a security contact address.

The design choice that most reduces risk here is architectural rather than policy: because suggestions carry citations back to the source documentation, a fabricated or unsupported code is checkable at the point of review rather than discovered in an audit.

HIPAA compliance is stated directly in the company's published FAQ alongside SOC 2 and GDPR, with compliance reports available on request. No explicit business associate agreement commitment was located, which is the one element that would move this to an A given the platform's total access to clinical documentation.

SOC 2 and GDPR compliance are both stated, and the company publishes a route to obtain the reports by emailing its security address. That is a functioning if manual trust process. The SOC 2 type is not specified, which matters because Type 2 tests operating effectiveness over time while Type 1 assesses design at a point in time, and no self serve trust center exists.

Converted from Not Rated. The prior note carried the correct analysis with no grade attached.

The scoping holds. No device pathway applies, since the platform operates on billing and coding rather than diagnosis or treatment, and the company's own site carries a disclaimer to that effect. What governs is payer audit exposure and False Claims Act risk attached to coding accuracy, and the liability sits with the practice submitting the claim rather than with the software supplier.

Where no vendor level regulator exists, this index grades on what the vendor publishes in its place: accuracy, audit trail, and whether a credentialed human verifies before submission. Two and a half of the three are answered.

The strongest is oversight, and it is configurable by the customer rather than fixed by the vendor. A practice can have the system code autonomously, autocorrect directly in the record, or queue charts for human intervention. Publishing that the level of human involvement is a customer decision is directly responsive to the liability allocation, because the party carrying the False Claims Act exposure gets to choose how much of it to accept.

Traceability exists too: reviews produce detailed reasoning and corrective recommendations rather than a bare code.

Held at B because the accuracy disclosure is relative rather than absolute. A 72 percent improvement over internal teams says the system beats a baseline that is not described; it is not an accuracy rate. And no confidence threshold is published, which is the benchmark a peer in this category sets by stating where machine output stops and a human coder takes over.

Converted from Not Rated. The prior analysis identified the right risk and the company's published metrics now make it concrete.

No AI governance framework, bias evaluation, calibration study or third party review was located.

The question is directional. A model tuned to find missed revenue has an incentive gradient toward coding up. The company frames the same engine as a compliance control that catches errors in both directions, and that framing is credible on its face, since payer compliance and revenue discovery are genuinely two views of the same review.

What settles the question is which direction is measured, and every quantified claim points one way. An eleven percent average revenue uplift. Up to a 15.2 percent increase in relative value units per encounter. Revenue discovery presented as a named product. No published figure describes corrections that reduced a code, no ratio of upward to downward adjustments, and no calibration evidence demonstrating the model is neutral rather than optimised for capture.

The per customer tuning sharpens it further. A dedicated engineer trains each customer's environment on feedback from that practice's providers and revenue cycle team. That is explicit, ongoing adaptation to one organisation's preferences, which is a legitimate implementation model and also the mechanism by which a practice's existing coding intensity, whatever it is, gets learned and applied at scale to every chart.

Ask for the distribution of coding changes in both directions, and what monitoring detects intensity drift over time.

The mechanism here is the right one for coding and it is the sixth instance of the same design in this backfill, which is beginning to look like the category's real differentiator. Suggestions carry citations back to the source documentation, so a fabricated or unsupported code is checkable at the point of review rather than discovered in an audit months later.

That matters because the characteristic failure of a generative coding system is not a wrong code drawn from a plausible reading, it is a code with nothing behind it at all, and a citation makes that immediately visible to the reviewer rather than to a payer. Held at C because nothing measures the system and the customisation model makes measurement awkward rather than impossible.

Per client implementations are built to replicate that client's own review criteria through an onboarding period of feedback, so performance is client specific and no single published number would describe it, and the company publishes none. That is a real explanation and not a reason to skip the disclosure: a vendor could publish a distribution across clients, or the accuracy achieved at the end of onboarding, or the proportion of suggestions accepted unchanged.

No evaluation methodology, accuracy figure, warranty, indemnity or remediation commitment was located. Ask what the acceptance rate is at the end of onboarding, and whether citations are validated against the claim they support rather than merely retrieved.

Integration is bidirectional and the write path is the point: the company states it integrates with the EHR to read all provider documentation immediately after a note closes, and can autocorrect coding mistakes directly in the EHR rather than reporting them elsewhere. It reports successful integration with dozens of commercial EHRs as well as homegrown EMRs, with a built in integration list customers select from, and assigns engineering support to each implementation. No named EHR certifications or marketplace listings were located.

Converted from Not Rated. No hosting, region, tenancy or residency terms were located, and the architecture raises a subprocessor question the published material does not address.

The company states it is compliant with the European data protection regulation alongside the health privacy rule, which implies some position on European handling, but no residency terms, transfer mechanism or regional hosting option is published.

The more consequential gap is specific to how this product is built. The platform is described as based on the most advanced large language models, reading any type of clinical documentation and running custom analysis modules. Whose models, and where they run, is not stated. If clinical documentation is processed by a third party model provider, that provider handles protected health information and is a subprocessor requiring its own agreement, and the buyer inherits a data path it has not evaluated. If models run within the vendor's own environment, that is a different and simpler position. Nothing published distinguishes them, and for a platform reading complete charts this is the first question a security review should ask.

The implementation model adds a further question. Each deployment is a custom integration built by a dedicated engineer who continues to tune the environment afterwards, which implies ongoing vendor personnel access to a live clinical system. What that access covers and how it is controlled is not described.

Ask which models process charts and where, whether any third party provider is in the path, and what standing access implementation engineers retain.

Converted from Not Rated. No pricing is published, and in this category the missing detail is not the number but the structure.

Nothing is disclosed on rate, basis or term. Whether pricing is per chart, per provider, subscription, or contingent on revenue recovered is not stated, and that distinction is the material one. A flat fee and a percentage of found revenue create opposite incentive structures on a product whose core function is deciding whether a chart supports a higher code.

That connects directly to the governance question on this record. Where every published outcome metric measures revenue captured, and the fee may scale with revenue captured, the alignment runs one way and the party absorbing a wrongly upcoded claim is the payer or ultimately the patient, neither of whom is party to the contract. The same structural observation applies to contingency priced payment integrity vendors elsewhere in this index, in the opposite direction.

What is published is the return side in detail: an average eleven percent revenue uplift, up to 15.2 percent more relative value units per encounter, and 10.6 times the chart coverage per full time employee. Precise numerator, no denominator, which is the same asymmetry this pass has now found in five categories.

One genuinely buyer favourable term is offered and deserves credit: a defined free trial period agreed with the customer before any long term commitment, which lets a practice test the claims on its own charts rather than take them on trust.

Ask for the pricing basis before the return model, and specifically whether the fee scales with revenue found.

Focused on outpatient and ambulatory fee for service settings across specialties, with stated buyers including provider organizations, health systems, payers, and managed services organizations. The company positions itself for high volume lower margin specialties and cites primary care clinics operating on very thin margins as a core case, alongside an urgent care association partnership. Inpatient coding is not the target, which distinguishes it from vendors built for facility coding.