AccuCode AI

The engine is the product across both lines. Documentation goes in, codes or abstracted quality measures come out, each cited to the passage that supports it. There is no workflow layer, network or platform underneath that would function without the model, and the company describes both products as configurations of a single engineering stack rather than separate systems.

The coder audit interface and the human review workflow sit on top of model output rather than beside it, so the model stays at the center. The company was founded inside a medical billing services firm, and what was spun out is the software rather than the service.

Oversight is published in detail, and the threshold that routes a chart to a person is not. The coding page says charts move from ingestion through code assignment to billing handoff with no per chart human review; coder review is reserved for initial validation during deployment (typically sixty to ninety days), compliance sampling afterward (typically one to five percent of charts), and charts routed to a coder by facility, provider or insurance specific rules. Codes that fail code set, NCCI or payer rule checks are dropped even at high confidence, and a code with no supporting passage is withheld. The quality product is described as fully automated from ingestion through submission, with submission direct or audited, and its interface shows per field confidence with approve, edit and flag for physician actions.

Neither the confidence level below which a chart or field goes to a person nor what happens after an error is found in production is published. The company argues that human review of every chart reduces accuracy because human coders agree only about 82 percent of the time, which is a case for calibrated sampling, while the routing threshold itself stays unpublished.

The approach is described specifically and no model is named. The home and coding pages say the system processes the entire patient record, including unstructured prose and handwriting, rather than a summary compressed to fit a context window; that roughly eighty percent of clinical information lives in that prose; that a seven stage architecture binds every code or abstracted field to its supporting passage and withholds a code with no passage; and that the models were never trained on patient encounters, customer data or licensed coding content, reasoning from documentation against current code sets, NCCI edits and payer rules instead. A patent is pending on upstream data structuring.

No foundation model, model class or version is named, and no change log is published. The whole chart claim is itself testable in a pilot.

The architecture is described and the host is named; no model provider is named. The security page names Microsoft Azure, restricted to United States regions, and says the company does not route data through third party interfaces hosted outside the United States; the coding page describes a seven stage architecture that binds every code to a source passage and says the models were never trained on patient encounters, customer data or licensed coding content. Vendor and subprocessor management is one of ten monitored control categories, including a vendor register, due diligence and communication of subprocessor changes to customers.

No foundation model or model provider is named on the home, security, clinical quality or coding pages, and no subprocessor list is published, only the commitment that changes to it are communicated. Naming the cloud host records residency and breach scope; it does not say which model reads a chart. The training commitment is strong and carries its weight on AI Safety.

Published evaluation with named auditors and a stated method, short of independent validation and without sample sizes. The coding page reports audits against credentialed coders: cardiovascular surgery by MedAxiom, the American College of Cardiology's cardiovascular collaborative, at 95.6 percent initial agreement and 99.1 percent after joint adjudication against source documentation and coding guidance; E/M by MedAxiom and Professional Consulting Services at 98.4 and 99.6 percent. The page states that the MedAxiom audit was the gating test before MedAxiom signed a partnership that became the joint product AccuCode CV, and that PCS is led by AccuCode's chief operating officer and was founded by its revenue cycle advisor, who led the PCS audits personally. The company describes both as independent organizations; on its own account the first became a partner after auditing and the second is a related party.

No per audit sample size is given (the page says tens of thousands of encounters in total), despite the FAQ promising sample size and panel composition for every cycle. A 100 percent score on a retired certification exam is a capability benchmark, as the page says. Baptist Health Systems is a named quality partner since August 2024.

Training and handling commitments are categorical and specific; the retention schedule is not published. The coding page says the company does not train models on patient data of any kind, whether identified, de identified, aggregated or synthetic, and that its models were never trained on patient encounters or customer data; the security page makes the same commitment contractually binding in every business associate agreement, and enumerates handling controls: encryption at rest and in transit with key rotation, least privilege access with deprovisioning inside one business day, multi factor authentication, audit logging across production, data minimization and a record of processing activity. Customer data, PHI, audit logs, backups and archives stay in United States infrastructure, and only United States based staff can access PHI.

What holds it out of A: retention periods are not stated (the data retention policy is offered by name under NDA), and nothing says how long charts are kept after processing or after a contract ends. Safety engineering is published in one respect: the system withholds a code when no supporting passage exists rather than generating one. How a wrong but cited code is detected across a customer's volume is not described.

Business associate status is stated with unusually specific contractual terms, and the agreement itself is not published. The security page states full HIPAA compliance across the Privacy, Security and Breach Notification Rules, business associate agreements executed with every covered customer and monitored as a control, and four commitments written into every agreement: United States data residency, United States based personnel for PHI access, no use of customer data to train any model, and breach notification materially faster than the 60 day statutory floor. A pilot begins with a signed agreement before any chart is sent.

A needs the agreement readable before contracting and a statement that it binds subcontractors who handle PHI. Neither is published: the terms are described rather than shown, and the vendor management controls describe privacy and security requirements in third party agreements without saying the business associate agreement flows down to them. The notification window is described as materially faster without a number.

A certification named with its type, and a live trust report anyone can open without asking. The security page links publicly to a daily updated Drata security report covering 244 continuously monitored controls across infrastructure, code repositories, identity provider, endpoints and HR systems, and says it is the same report shared with the company's SOC 2 auditors. SOC 2 Type II is named with its scope (security, availability, processing integrity, confidentiality and privacy over a period of operation), and the page distinguishes annual attestation from daily verification.

Ten control categories are enumerated, including penetration testing, continuous vulnerability scanning with risk based remediation, automated patching, static application security testing, software composition analysis, background checks and phishing simulation, with any policy offered by name under NDA. The Drata report itself may show whether the SOC 2 report can be downloaded from it or is supplied on request.

No device pathway applies to either product and none is claimed. Assigning billing codes and abstracting quality measures from documentation are administrative determinations rather than clinical ones.

The coding side carries the usual exposure: codes on a claim are representations to a payer, and error is governed by federal false claims enforcement landing on the billing provider. Source citation for every code and an integrated audit workflow are the right posture against it.

The quality side carries a different exposure. The product calculates and submits measures to CMS programs including Value Based Purchasing, the readmissions reduction program and MIPS, to The Joint Commission and to specialty registries, and the clinical quality page describes submission as direct or audited and the pipeline as fully automated. Submitted quality data drives payment adjustment, public reporting and accreditation standing, and an incorrect automated submission is a different kind of event from a miscoded claim. Nothing on the home, security, clinical quality or coding pages says who attests to a submitted measure, whether a person signs off before transmission when direct submission is chosen, or how corrections and resubmissions are handled.

Governance commitments with named process behind them, and no subgroup result. The company states four evaluation practices, each specific enough to check: it reads the whole chart even when an answer sits in a structured field, because fields are sometimes wrong; it flags disagreement rather than resolving it silently; it benchmarks against adjudicated ground truth rather than its own prior results; and every audit cycle reports what the system missed to the customer. Audit panels, adjudication method and the specialties tested are described on the coding page, and processing the whole record rather than a summary removes one place where a model silently decides what it never sees.

Nothing compares performance across groups: no breakdown by specialty beyond the two audited, by payer, by physician documentation style or by facility, and no external audit of coded output by an unaffiliated party. Documentation quality varies across settings, and whether accuracy holds on sparse or handwritten charts is the question a comparison would answer.

Mechanisms let a reviewer challenge each output, an error rate is published without its denominator, and nothing stands behind the output. Every code is bound to a cited passage and the system withholds a code when no passage supports it; compliance sampling of one to five percent of charts continues after deployment; and the company commits that each audit cycle reports its misses to the customer. The coding page now states an error rate: less than one percent on the hardest codes after adjudication, from audits reported at 99.1 and 99.6 percent.

The band credits a published error rate with its method and denominator. The method is described (initial agreement, then joint adjudication of each disagreement to source); the denominator is not, beyond tens of thousands of encounters across all audits, and the auditors include a firm led by the company's own chief operating officer. The contractual commitments the security page lists (residency, personnel location, no training, faster breach notice) concern data handling, not the correctness of codes. No warranty, remediation or reversal commitment attaches to a wrong code or a wrong quality submission.

One record system credential and one named standard, with a narrow footprint. The home page and site header say an AccuCode app is available on Epic and name HL7 FHIR as the integration standard. Coding output hands off to the billing system, and the quality product submits directly to CMS programs, The Joint Commission and specialty registries without rekeying, so outbound integration reaches past the record system.

The gaps are in breadth and direction. Only one record system is named on the home, security, clinical quality or coding pages, nothing states whether integration writes back to the record, and the quality product needs integration work before it runs on real data while the coding pilot runs on exported charts.

Residency is stated in detail and tenant isolation is not. The security page says the platform runs in United States only Azure regions with network segmentation, restricted public access, a web application firewall, private endpoint access to storage, zone redundancy and autoscaling, and that customer data, PHI, audit logs, backups and archives are stored only in United States infrastructure, with backups and archives in region and no cross border replicas. Every person who can access PHI works in Little Rock or Seattle, with no offshoring or third party outsourcing of access, and the commitment is written into every business associate agreement. Business continuity is covered by documented backup, restore testing and recovery exercises.

It is not stated whether customers share infrastructure or are isolated from one another, or whether any processing, including calls to a foundation model service, leaves the named Azure regions. One hosted option is described.

The evaluation path is specified precisely and the economics are not published at all.

On the pilot the company is concrete: send one hundred of your own charts, receive every code cited to source documentation, see the disagreements your current coders will register, four weeks end to end, beginning with a signed business associate agreement. For quality abstraction the first conversation covers registries and measure scope, with the interface demonstrated on synthetic data before any real deployment. The first engagement is laid out in full.

On cost there is nothing. No pricing page, no unit of charge, no range, no implementation or onboarding fee position, no minimum commitment, no return calculator, and no percentage saving against existing coding spend. Whether the pilot itself is free is not stated.

The pilot terms are specific enough to count as a real commercial disclosure, but no economic figure of any kind exists.

Coding coverage is asserted broadly and quality coverage is enumerated precisely. Coding is described as specialty agnostic with cardiovascular optimization built with MedAxiom, the American College of Cardiology's cardiovascular collaborative, and sold jointly as AccuCode CV; audits cover E/M and cardiovascular surgery, and the company says every other specialty scores higher without showing it.

Quality coverage is listed as live on the clinical quality page: CMS eCQMs, Hospital IQR, Value Based Purchasing, the readmissions and hospital acquired condition reduction programs, MIPS and MVP; Joint Commission ORYX measures; NCDR CathPCI, ICD, AFib ablation and chest pain registries; STS adult cardiac, congenital and general thoracic; and AHA Get With The Guidelines. That is where the value based care tie sits: the measures that set hospital payment adjustments and clinician MIPS scores.

Demonstrated coverage rests on one named health system partner (Baptist Health Systems) and one specialty partnership, and no named customer is tied to the quality programs listed.