Maverick Medical AI

The model is the product. mCoder reads each clinical report, assigns CPT, HCPCS and ICD-10-CM codes and sends 85 percent or more straight to billing, and CodeAgent checks the physician's note at the report screen for issues that would block coding or billing before sign off. Since September 2026 both run on what the company calls its Clinical-to-Revenue Foundation Model, with further agents checking payer policy, coding compliance and reimbursement rules before a claim leaves. Remove the model and nothing remains to sell: the dashboards, variance reports and exception queues exist to monitor and complete what the model codes.

What the system may and may not do is published, with the routing conditions and the customer's control over the threshold. The FAQ lists what goes to a person rather than to billing: low confidence predictions, missing or insufficient documentation, ambiguous or conflicting language, procedures that may not meet medical necessity under local or national coverage determinations, National Correct Coding Initiative edits, and complex cases such as interventional radiology. Exceptions reach the customer's coders in the mCoder platform with the supporting passages shown, and the model learns from their corrections.

The customer sets the boundary. At go-live the client reviews 100 percent of cases the engine coded, typically for a week, then decides how many encounters move to direct to bill and what share to keep auditing in a quality assurance queue; every direct to bill encounter remains reviewable. Maverick's compliance lead monitors coder changes daily in the first week, then weekly, then audits 200 direct to bill encounters per client each quarter. The numeric confidence cut-off is not published.

The company's marketing claims: "No humans. No delays. No backlogs." describes the automated share; "100% automation coverage" and "protect every claim before it reaches the payer" (September 2026 release) describe coverage and process rather than correct output; "fully autonomous, complete and compliant coding" reads as end to end coding.

The approach and the training data are described, without version discipline. The FAQ says the engine combines machine learning, deep learning and transformer models, and that Maverick trains and fine-tunes them on medical coding data into a masked language model. The base model was trained on tens of millions of radiology reports, and each client's model is further trained on the past two years of that client's documentation and coding history, so it learns local coding standards, report templates and payer mix. It keeps learning from coders' corrections to the exceptions it routes, and the team builds crosswalks and updates the model when new code sets take effect. In September 2026 the company announced a Clinical-to-Revenue Foundation Model that now powers both mCoder and CodeAgent across specialties.

What is missing is the discipline that tells a buyer when the system changed: no version identifiers, release notes or change notices for the continuous learning or the foundation model switch, and no accuracy by coding element.

A structurally short chain: the model is built in house and hosted on a named cloud. The FAQ says Maverick trains and fine-tunes its own masked language model on coding data and runs the platform on US located Amazon Web Services servers; no outside model provider is described. The September 2026 release calls the Clinical-to-Revenue Foundation Model proprietary.

The chain is not enumerated beyond that. No subprocessor list is published; the privacy policy (updated 28 May 2025) names only categories (a cloud service provider, analytics and marketing providers, a CRM provider) and excludes client data. The product also reaches customers through partners, including Infinx, a revenue cycle vendor that invested in August 2025, and radiology platforms named in earlier announcements, and none is described as a party that receives report text. Nothing states whether any third party model, speech or text service sits in the CodeAgent path.

Named deployments with before and after figures, answering throughput rather than accuracy. RadNet (about 400 sites, over 10 million scans and 500 radiologists a year) moved from 50 to 86 percent direct to bill at go-live and cut coding lag from 12 days to one; its chief information officer is quoted saying all its coding runs on Maverick at 85 percent direct to bill. Steinberg Diagnostic Medical Imaging (over 500,000 studies a year) went from 40 percent baseline automation on high volume studies to 85 percent direct to bill, cut coding lag from 10 to 12 days to one, and stopped outsourcing (case study, November 2025). An unnamed imaging network of over 100 sites reports 87 percent direct to bill from zero, 95 percent coding accuracy and 60 percent budget savings within 90 days. The FAQ says customers have reached up to 93 percent.

The headline claim pairs 85 percent direct to bill with 97 percent accuracy, and the accuracy half has no published method, sample or reference standard; the quarterly 200 encounter audits described in the FAQ would produce one, and none is published. The full case studies sit behind download forms.

Security assurances plus a disclosed training practice, with retention and de-identification unanswered. The FAQ says data is encrypted in transit and at rest, access is role based and logged, networks are segmented, and the infrastructure sits on US located Amazon Web Services servers. It also says how customer data reaches the model: the base model was trained on tens of millions of radiology reports, each client's model is further trained on the past two years of that client's documentation and coding history, and the system keeps learning from coders' corrections to exceptions. That is a statement of training use, and it runs toward using customer data rather than protecting it: it does not say whose reports trained the base model, or whether one client's data or corrections improve the shared model other clients use.

No retention period for reports, coding history or training copies is stated, and nothing addresses de-identification. The company privacy policy (updated 28 May 2025) excludes data processed on behalf of clients. Safety engineering, meaning how a model fault is detected and handled, is described only as accuracy monitoring.

HIPAA compliance is claimed and business associate status is not stated. The FAQ says "We maintain HIPAA compliance, and our platform is SOC 2 certified", and the home and mCoder pages carry the same pairing, with encryption, access controls and audit logging described. The company privacy policy (updated 28 May 2025) covers website visitors, clients and platform users, and states it does not apply to individuals whose data Maverick processes on behalf of its clients, which is where patient reports sit; it does not mention HIPAA or protected health information.

No statement that Maverick acts as a business associate, no agreement, no tier and no subcontractor position is published.

A recognized certification named in the vendor's own material, without the artifact. The FAQ states "our platform is SOC 2 certified" and the home and mCoder pages repeat it, with encryption in transit and at rest, role based access, audit logging, network segmentation and continuous monitoring on US located Amazon Web Services servers.

No report type (Type 1 or Type 2), audit period, auditor or date is given, and the company publishes no trust center, report request route, penetration testing cadence or vulnerability disclosure policy, so the report's type, scope and date are available only on request. The security controls AWS certifies for its own data centers belong to the host, not to Maverick.

The product is scoped as revenue cycle work, and no device claim is made. mCoder assigns billing codes from finished reports, which is an administrative determination, so no clearance is expected and none is claimed.

The exposure sits in claims submission. Codes on a claim are representations to a payer, and inaccurate coding falls under federal false claims law, which lands on the billing provider. CodeAgent adds a documentation question: it prompts physicians before the note is finalized about missing documentation, medical necessity gaps under coverage determinations and payer specific risks. Those prompts concern billing support rather than diagnosis or treatment, and nothing in the company's material markets a clinical recommendation. Still, software that shapes documentation and then codes from it sits close to documentation integrity rules, and the company states no position on them.

A named process for testing model behavior, with no results published. The FAQ describes several rounds of validation before go-live, comparing the model's codes with the customer's historical coding and benchmarking it against expert human coders, with errors in the customer's own historical coding flagged back to its coding team. After go-live, Maverick's Director of Coding Compliance monitors coder changes daily in the first week and weekly after, then audits a random sample of 200 direct to bill encounters per client each quarter. Customers get variance reports comparing the model with their coders and with auditors, code over code, and track direct to bill rate and accuracy on a dashboard. The model's work is open to the customer's own review.

What is not published is any result from those audits, any breakdown by specialty, payer or modality, or any bias or drift analysis. The design also has a loop of its own. CodeAgent prompts physicians about missing documentation and payer specific risks before the note is finalized, and mCoder then codes from that note, so the same system shapes the documentation and codes from it. Nothing says how those prompts are logged in the record. Safety engineering is not addressed.

Mechanisms to challenge an output exist, and the one guarantee published covers automation rather than correctness. The home page advertises an "85% guaranteed direct-to-bill rate"; that commits Maverick to the share of reports coded without people, not to those codes being right, and a vendor could meet it while billing wrong codes. Accuracy is published as 97 percent on the home page and as "95%+" maintained alongside 85 percent direct to bill in the FAQ, with no study, denominator or reference standard behind either figure.

What stands behind the output is review machinery. Built-in explainability shows which parts of the report supported each code, every direct to bill encounter stays available for review, the customer reviews 100 percent of coded cases for about a week at go-live and chooses how many to keep auditing, and Maverick audits 200 direct to bill encounters per client each quarter. No remedy is stated for a shortfall in accuracy found by those audits, and no warranty or indemnity is published. Codes on a claim are the provider's representations to a payer.

Integration is described as a service delivered to specification, with no record system named. The FAQ says Maverick integrates the platform "with your system, according to agreed specifications", taking reports on a continuous live feed or in batches and sending codes to billing, and recommends payer specific billing rules stay in the customer's billing system. The customers named are imaging operators, whose reports come from radiology information and reporting systems, and the company's earlier announcements tied it to a cloud radiology information system, radiology reporting products and a radiology billing vendor; the September 2026 release extends the pitch to hospital revenue cycles.

No record system, radiology information system or billing system is named on the product, FAQ or release pages, no interface standard (HL7, FHIR) is stated, and no marketplace listing is linked. A buyer cannot tell from published material which systems run in production or in which direction data moves.

A single hosted service with its location stated. The FAQ says Maverick's cloud infrastructure is "built and maintained in US located Amazon AWS servers", with encryption in transit and at rest, role based access, audit logging and network segmentation, and the mCoder page repeats that the platform runs on US based AWS infrastructure. Reports arrive on a continuous live feed or in batches.

What stays open is isolation and the processing path beyond the platform: nothing says whether each client runs in a separate environment, where backups and training copies of client coding history sit, or where engineering and support staff with access to production data are located.

No price is published, and the terms of the sale are discoverable. The product pages and FAQ describe what a customer buys: mCoder coding on a live feed or in batches, a guaranteed 85 percent direct to bill rate, implementation in about 90 days led by an assigned project manager, go-live validation, quarterly accuracy audits, a dedicated account manager and a help desk with stated hours. The cost evidence is an unnamed imaging network reporting 60 percent budget savings and a customer statement that it no longer outsources coding.

What drives the fee is not stated: no unit of charge (per report, per study or per volume band), no figure, and nothing on whether CodeAgent is priced separately or whether any fee rides on the direct to bill guarantee. Distribution also runs through partners, including Infinx, which invested in August 2025, so the contracting party may be a revenue cycle vendor rather than Maverick.

Coverage is named, with validation behind radiology. Every named deployment is an outpatient imaging operator: RadNet, Steinberg Diagnostic Medical Imaging and an unnamed network of over 100 imaging sites, with direct to bill rates between 85 and 87 percent and customers reaching up to 93 percent. The FAQ is written around radiology workflows, including interventional radiology cases routed to people.

The September 2026 release extends the claim to the hospital revenue cycle and names cardiology, orthopedics, emergency medicine and pathology among specialties the foundation model can be deployed across. No customer, figure or date backs those settings, and CodeAgent's physician facing checks are described without a named deployment. Nothing addresses coding regimes outside the United States.