RAAPID

The model is the engine of the core modules, and the platform carries other value. RAAPID's neuro symbolic AI, machine learning combined with a clinical knowledge graph, reads charts, suggests condition codes with MEAT evidence and flags unsupported ones across retrospective, prospective and RADV audit workflows, and those capabilities do not exist without it.

The rest can be sold without it: the retrospective page offers a fully managed service in which the company's certified coders handle a plan's entire retrospective program, the RADV solution generates export files, manages audit responses and adds expert auditor oversight, and the platform carries the coding and validation workflow a team works in. A buyer is purchasing an AI engine inside a coding and audit operation, which the company can also staff.

Coder validation is built into the workflow and the company publishes what it contributes, reporting 92 percent accuracy before human review and 98 percent after. Publishing both numbers tells a buyer how much work the human is doing. The RADV solution adds expert auditor oversight, and the prospective workflow brings suggestions to the clinician before and during the visit rather than submitting them.

The missing part is the threshold. The company announced its platform in 2025 as the first autonomous clinical AI platform for risk adjustment while describing coder validation as the step that produces final accuracy, and nothing on the home, security, AI as a service or terms pages says which outputs, if any, pass without a coder, at what confidence a suggestion is suppressed, or how the two way removal of unsupported codes is reviewed before a code is withdrawn.

The approach and the suppliers are named, without version discipline. The company describes neuro symbolic AI, machine learning and deep learning combined with a clinical knowledge graph, plus document AI and OCR for unstructured charts, and says every suggested condition code links to specific MEAT evidence in what it calls a glass box audit trail. Its April 2025 marketplace announcement names Azure AI Foundry, including Azure OpenAI Service, and Azure Machine Learning for core model development and continuous improvement.

What is missing for A: which models and versions sit behind extraction and suggestion, what the knowledge graph is built from, independent validation of the published 92 percent and 98 percent figures, and a version or change log. "Continuous improvement" means the models change, and in risk adjustment a customer needs to know when the model that suggested a code changed.

The model layer and host are named, without the full party list or the terms that govern data once it arrives. The company's April 2025 announcement of its Microsoft Commercial Marketplace offer says the platform uses Azure AI Foundry, including Azure OpenAI Service, and Azure Machine Learning for core model development and continuous improvement, on a multi region Azure deployment; its security page names Microsoft Azure and Amazon Web Services as hosting partners. Naming the model service places this in B.

One deployment option shortens the chain further: the platform can run inside the customer's own Azure tenant, where the company says PHI never leaves the organization's control. Under the RAAPID hosted option the vendor holds the documentation. No subprocessor list, and no retention or training terms governing charts sent to Azure OpenAI or used for continuous improvement, appear on the home, security, AI as a service or terms pages, and multi region deployment leaves open which regions process a given customer's charts.

Claims are specific and quantified but entirely vendor reported, with no independent validation behind them. Published figures include 92 percent out of the box accuracy rising above 98 percent with coder validation, 60 to 80 percent reduction in manual review burden, and 3 to 10 times return on investment. The company also asserts industry accuracy of 20 to 30 percent as its comparison baseline, a striking claim about rivals that no cited source supports. Backing from a major corporate venture arm is a signal, but a limited one. Founded in 2022, so there is no longitudinal track record.

The surface is broad, one deployment option is a real mitigation, and the handling terms are unpublished. The platform processes complete clinical documentation for health plans and providers, and the company reports more than 8 million patient records a year. Deployment inside the customer's own Azure tenant keeps charts behind a perimeter the plan controls, which the company describes as PHI never leaving the organization; the hosted and managed options put the documentation in the company's hands.

The company's marketplace announcement says Azure Machine Learning and Azure AI Foundry serve core model development and continuous improvement, and a platform built on a clinical knowledge graph improves as that graph is refined. Nothing on the home, security, AI as a service or terms pages says whether customer charts contribute to either, whether learning from one customer reaches another, how long documentation is retained under the hosted options, or what is returned at contract end. No safety engineering is published either: how an unsupported suggestion that slipped past review is traced and corrected across other charts.

Compliance is asserted and the instrument is not published. The home and security pages display a HIPAA compliant badge and describe the platform as HIPAA compliant; the AI as a service page says the APIs are designed to meet HIPAA's requirements. No business associate agreement, subcontractor flow down or breach notification term appears on the home, security, AI as a service or terms pages.

The role is ordinary on the hosted and managed options: a vendor reading a plan's or provider's charts to identify and validate diagnosis codes is a business associate. The customer tenant option changes the shape. Where the platform runs inside the customer's Azure environment and PHI does not leave it, the agreement may cover support access rather than data custody, and support access for implementation, tuning and troubleshooting is what it must then govern. The managed coding service, in which the company's own coders work the charts, is custody in the fullest sense.

Recognized certifications are named in the company's own material, without the artifacts. The security page says RAAPID holds HITRUST r2 certification, which it describes as independently validating the controls behind every clinical AI workflow, alongside SOC 2 Type 2, and its announcements say HITRUST certification has been held for two consecutive years. The page names Microsoft Azure and Amazon Web Services as hosting partners and describes their own compliance programs, which are theirs rather than RAAPID's.

No trust portal, certification letter or SOC 2 report request path is linked, and no scope statement says whether the customer tenant deployment, the hosted option and the managed coding service all sit inside the certified environment. SOC 2 produces an attestation report, and neither its period nor its auditor is published.

No device position is needed for what is marketed, and the governing regime is CMS risk adjustment. The platform suggests diagnosis codes with the chart evidence behind each, assessed against MEAT, for a coder to validate retrospectively or a clinician to address before and during a visit, and flags codes the record does not support. Surfacing possibly undocumented conditions with the evidence shown for a person to confirm is a documentation function with a reviewable basis, not a diagnostic claim, and building a product around the audit standard is not a regulatory pathway.

What governs is Medicare Advantage risk adjustment law: submitted diagnoses must be supported by the record, RADV audits test that and CMS now extrapolates findings across a contract, and the prospective workflow's in visit prompts sit close to the line between documentation support and steering diagnosis. The two way design, removing unsupported codes as well as adding missed ones, is the right orientation for that regime.

No governance program or bias evaluation is published, but the explainability architecture functions as a governance control in practice, since a code that must be traced to specific documented clinical evidence cannot be generated from statistical association alone. That structurally constrains one way automated coding can go wrong. No analysis addresses whether performance varies by documentation quality, specialty, or population, which is the live equity question when documentation density differs across patient groups.

Mechanisms let a reviewer challenge each output, and the one commitment published does not say what it pays. Every suggested condition code links to chart evidence assessed against MEAT, the documentation standard CMS auditors apply; coding is two way, adding missed diagnoses and flagging unsupported ones for removal; and coders validate suggestions before submission, with the company reporting 92 percent accuracy before that review and 98 percent after. That is source traceability tied to the auditor's own test, and review before commit.

The retrospective risk adjustment page states a guaranteed 10 to 1 return on investment. Nothing on the home, security, AI as a service or terms pages says how that return is measured, over what period, or what the customer pays or recovers if it is not reached, so it is not yet a commitment a buyer can hold the company to. The terms of service disclaim any warranty of the accuracy or reliability of the materials, and no remediation commitment attaches to a code that fails audit.

An interface exists and no integration is named. The home page describes the prospective solution as EHR integrated and lists API integration and custom workflows for AI as a service; the company's marketplace announcement describes an EHR integrated solution on Azure, and the AI as a service page says the APIs integrate with existing health IT systems. The prospective product implies deeper reach than retrospective review, since pre visit summaries and in workflow prompts have to reach a clinician inside the record system at the point of care.

No record system, supported version, standards based interface, interface documentation or write back path is named on the home, security, AI as a service or terms pages. Where the platform runs inside the customer's own Azure environment, integration is a problem the customer's own team can solve against its own systems, which is a different and often better position than a vendor connector list.

Options and residency are stated, with the processing path and tenant isolation left open. The company describes deployment via API, Azure or SaaS on its retrospective page, and its announcements describe running natively in a customer's own Azure environment with full data sovereignty and identity control, or on RAAPID hosted Azure; customers with Azure consumption commitments can apply them to the platform, and the company says this removes the need for additional security approvals. For a health plan, deployment inside its own tenant means the charts need not leave a perimeter it already controls.

What A would need is not published on the home, security, AI as a service or terms pages: how customers are separated in the hosted option, which regions process data in a multi region deployment, whether model calls to Azure OpenAI stay in region, and backup and retention posture.

How the product is bought is published, but not what it costs. The company sells four solutions (retrospective and prospective risk adjustment, RADV audit, and AI as a service by API) in three service models: a fully managed program in which its certified coders do the work, the platform used by the customer's own coders, or the AI connected to existing systems. It is transactable on the Microsoft Commercial Marketplace, and customers with Azure consumption commitments can apply them toward it. The retrospective page states a guaranteed 10 to 1 return on investment, with average revenue of $2,500 per member, and implementation in four to six weeks.

Value is framed as return on revenue captured. What drives the cost (per chart, per member, per coder seat or per API call) is not stated, though the marketplace listing may carry a price. Whether any fee rides on codes captured matters here, because in risk adjustment a fee tied to capture points the incentive the same way as over coding.

Coverage spans the full risk adjustment workflow rather than one phase: prospective documentation improvement before and during the visit, retrospective coding with two way validation, RADV audit response, and AI by API for technology and service partners. Buyer types named are health plans, providers including ACOs and health systems, and technology and service partners, with program coverage centered on Medicare Advantage, ACA and at risk arrangements. Investors include a payer (CareFirst's Healthworx) and a health system (UPMC Enterprises), and KLAS has featured the company in an emerging company spotlight based on customer ratings.

The named validation is general. Customer quotes on the home page are unattributed by organization, and no named deployment is tied to the prospective workflow specifically. It is not a clinical product.