Azra AI

The models are load bearing. Without them the product is a team of people reading every pathology and radiology report a hospital produces, which is the manual case finding it replaces: the company describes clinically trained AI turning unstructured reports into identified patients at the moment a diagnosis is confirmed, and publishes a headline precision figure of 99 percent. Its oncology product surfaces newly diagnosed patients to navigators in real time and ranks their tasks by acuity, distress risk and care gaps.

Much of what a customer buys sits around the model rather than inside it: navigation workflow, multidisciplinary meeting and tumor board management, registry automation, enterprise analytics, and a services arm offering clinical and non clinical navigation and registry case finding staffed by people. The company's own positioning has moved toward care orchestration and platform language, and a growing share of the findings it routes now come from partners' imaging models rather than its own. The model is the trigger; the platform and services are the product.

The structure is described, and it automates more than a navigator's worklist. For oncology, the system surfaces a patient at the moment a diagnosis is confirmed, ranks navigator tasks by acuity, distress risk and care gaps, and triggers follow up such as distress screening and consult scheduling. For incidental findings, the page describes reports streaming in and patients being scored and automatically enrolled into surveillance or treatment pathways, all without manual intervention, with human in the loop review described as keeping clinicians in control. For the registry, registrars review and confirm or flag each case the model finds before it is used, with structured feedback returned to the model.

So the routes to a person exist and are named, and one part is missing: the thresholds. Nothing published says at what risk score a finding surfaces immediately, what a clinician reviews before an automatic enrollment takes effect, how the worklist ranking is built, or what happens to a report the models do not surface. For incidental findings, where the only signal is a line in a report written for another reason, a case the system never raises is a patient nobody follows up, and there is no published path by which that would be noticed. Whether a customer can tune how many cases surface is not stated.

Provenance is stated, and the performance figure published is the wrong half. The company has stated that its natural language models were trained on more than 100 million pathology and radiology reports through an exclusive partnership with a large health system, which says where the data came from; the home page describes clinically trained AI and says the models are continuously evolving. No architecture, validation method, publication or independent evaluation is published, and continuous evolution with no versioning or change notice means a customer cannot tell when the model reading its reports has changed.

The specific gap is the published metric. The home page states 99 percent AI precision, and precision describes how many flagged cases were correct. For a product whose purpose is to stop patients being missed, the number that matters is recall, which describes how many real cases were never flagged at all. Publishing precision without recall reports the errors a hospital will notice and omits the ones it never will. Nor is it stated which findings now come from partner imaging models rather than the company's own.

Part of the model chain is named, because part of it now belongs to partners. The press page names two integrations announced in August 2026: Blackford, whose portfolio of more than 130 medical imaging AI applications now triggers the company's navigation and follow up workflows when a finding is detected, and 4DMedical, whose quantitative lung imaging analytics are paired with the platform under an OEM and reseller partnership. The change log also records an integration of RevealDx's lung lesion score. So for imaging findings, a buyer can identify the parties whose models produce the signal that starts the workflow.

For the company's own layer, the natural language models reading pathology and radiology reports are described as its own and clinically trained, on a corpus the company has said came from an exclusive partnership with a large health system; no supplier is named behind them, and whether any third party foundation model sits underneath is not stated.

What is missing is the rest of the list and the terms. No hosting provider, no subprocessor register, and nothing on what the imaging partners receive, retain or use for training when their findings and the associated patient context pass through the integration, and the privacy policy covers only website contact information.

Scale is claimed, outcomes are headlines, and there is no evaluation. The company has described deployment at hundreds of hospitals, including five of the ten largest United States health systems, and real time processing of more than half a billion clinical reports and messages a year, with MultiCare Health System named among customers. The home page carries outcome counters for reduced manual case finding time, more cancer patients identified and more nurse time with patients, with no denominator, period or comparison behind any of them. The oncology page frames the problem with cited industry benchmarks, such as the share of newly diagnosed patients who never reach a navigator, which describe the gap rather than the product's effect on it.

No peer reviewed publication, independent evaluation or registered study appears. Scale of use is not evidence of benefit: a system reading half a billion reports says nothing about how many patients reached treatment sooner because of it.

Nothing published on retention, minimization or secondary use, for a product that streams clinical free text out of hospital record systems continuously. The privacy policy, last updated March 2026, is a short notice about contact information collected from website visitors and prospective customers, with a generic data security paragraph; it does not address pathology reports, radiology reports or patient data at all.

The training question is specific and unanswered. The company has said its models were trained on more than 100 million pathology and radiology reports obtained through an exclusive partnership with a large health system. Training on that corpus requires a lawful basis, and nothing published states whether the reports were de identified and by what method, whether patients were informed, whether the arrangement was research or commercial, or who owns the resulting model. The consequence follows from the disclosed facts rather than from any allegation: a model built on one health system's clinical text is sold to other systems, and the originating patients are not party to it. Nothing says whether current customers' reports contribute to further training, which the home page's continuously evolving models make a live question.

No statement of status and no privacy document that reaches the product. The site's only legal page, linked from the footer, is a privacy policy last updated March 2026 that covers contact information from website visitors and prospective customers: names, email addresses, job titles and organization names. It mentions the company's services in passing and says nothing about protected health information, patient records, pathology or radiology reports, or business associate status, and no agreement, template or tier is published anywhere.

Agreements very likely exist, since the platform receives live feeds inside hospital record systems; the gap is what a buyer can read before contracting, for a product whose input is clinical free text about newly diagnosed cancers and incidental findings. Nothing says whether an agreement would bind the imaging AI partners whose findings now flow through the platform.

Nothing independent is published. The footer carries only a privacy policy, whose data security paragraph says the company uses administrative, technical and physical measures and that no method is completely secure. No SOC 2, HITRUST or ISO 27001 attestation is named anywhere, there is no trust center, no penetration testing statement and no vulnerability disclosure policy.

The profile raises the stakes of that absence. The company describes live interface feeds inside hundreds of hospitals, including some of the largest systems in the country, carrying pathology reports, radiology reports and admission and discharge messages, and it now connects outward to partner imaging platforms as well. Concentration of sensitive clinical text and live connections into hospital networks are the two properties that make a vendor attractive to attackers, and nothing published describes how either is managed.

No clearance, authorization or submission is claimed. What matters is what a flag is built from, and the company's pages describe two different things.

Much of the product restates what a signed report already says. For oncology, patients are surfaced at the moment a pathologist's diagnosis is confirmed; for incidental findings, natural language processing extracts what the radiologist wrote, including location, size, density, margins and nodule count. Where the flag carries the report's own language and the report sits behind it, the basis is reviewable by the navigator or clinician who receives it, which is the position operational software normally occupies.

The incidental findings programs go further. Each finding is scored with evidence based risk calculators and predictive AI models so that high risk patients surface first; Lung-RADS, LI-RADS and Bosniak classifications and Fleischner adherence are applied automatically; and patients are automatically enrolled into surveillance or treatment pathways, with human review described as keeping clinicians in control. A category or risk score assigned by the system where the radiologist did not assign one is an inference beyond the report's own language, and the page is titled around early cancer detection. The cardiology line is likewise described as detection and acuity stratification.

The reviewable basis holds for the report derived findings; the risk scoring, automatic classification and pathway enrollment are the part a regulator reading the marketing would examine.

Nothing published on how the models are governed or tested: no per site validation process, no drift monitoring, no subgroup performance, no model card and no bias testing method.

The architecture makes one gap specific and predictable. Natural language models over clinical reports inherit the documentation practices they were trained on, and pathology and radiology reporting varies by institution, subspecialty and individual reporter in vocabulary, structure and what is left implicit. A model trained on one large health system's reports and deployed at many others faces exactly that shift, and a report style the model handles less well produces silently missed patients rather than visible errors. The home page's statement that the models are continuously evolving adds drift from the vendor's side to drift from the customer's.

The worklist ranking carries a second exposure: ordering navigator tasks by distress risk and care gaps decides who is contacted first, and nothing states whether that ordering has been checked across age, language, insurance or race, groups for whom navigation access is already uneven.

Review mechanisms exist and catch one kind of error; nothing stands behind the output and nothing catches the other kind. Registrars review and confirm or flag every case the registry product surfaces before it is used, with structured feedback returned to the model, and the incidental findings workflow describes human in the loop review before care coordination proceeds. A person can therefore reject a wrong flag before it is used.

The failure those steps cannot catch is the one that matters most. The published figures, 99 percent precision on the home page and 2 percent false positives for lung findings, describe how often a surfaced case is wrong. A patient with a new cancer finding or a suspicious incidental finding whose report is never surfaced generates no artifact anywhere, so a hospital sees the cases the system caught and cannot see the ones it missed.

No warranty, service level, indemnity or remediation commitment is published, and there is no customer agreement or terms page on the site. Nothing describes who is responsible when a follow up delay traces to a report the model did not raise, or when an automatically assigned category sends a patient down the wrong surveillance pathway.

The integration runs deep and is now named, on the vendor's own account. The incidental findings page says radiology reports stream in over HL7 and FHIR from Cerner, Epic or any source system, are processed in real time and routed into care pathways, and that every action is written back so clinicians do not document twice. The oncology product surfaces patients from the record at diagnosis and sends documentation back into it. The registry product writes captured case information directly into the customer's registry database, and the company has described its registry output in the national reporting format. It now also receives findings from partner imaging platforms, including more than 130 imaging AI applications through Blackford and quantitative lung imaging through 4DMedical.

Standing inside a hospital's interface layer in real time, and staying correct as it changes, is a harder problem than a nightly extract, and the registry output lands in a regulated downstream system rather than a dashboard. The company has named MultiCare Health System among customers and describes deployment at hundreds of hospitals.

Verification is the gap: the Epic and Cerner connections are the vendor's own statement, with no marketplace listing, integration documentation or named customer tied to a named system, and nothing shows what the write back contains or where it lands.

Nothing published says where the platform runs or where reports rest. The product is described as receiving live interface feeds from customer record systems and routing findings to navigators, which establishes a hosted service, and no page names a hosting provider, region, residency commitment, tenancy model, or backup and continuity position. The privacy policy covers contact information and says nothing about the service's data.

The notable feature of this deployment is volume and direction rather than location: clinical free text is streamed continuously out of hospital record systems to a single vendor environment, and now findings arrive from partner imaging platforms as well. Whether reports are retained after processing, and for how long, is the question that decides the size of that exposure, and it is unanswered. So is availability, for a system whose value is surfacing a newly diagnosed patient within hours rather than days.

The offer is enumerated in detail, and nothing in it is priced. The solutions menu lists closed loop patient management, oncology navigation, incidental findings and screening, high risk screening, multidisciplinary case review and tumor boards, emergency to outpatient continuity, registry automation, clinical trials and enterprise analytics, organized by service line: oncology, radiology, cardiology, emergency medicine and clinical research. A separate services menu sells people as well as software: strategic execution, clinical and non clinical navigation, tumor board support and cancer registry case finding. The route in is an enterprise review, so this is sold as a platform plus managed services to health systems.

What no page answers is the unit of charge, and on this product the candidates diverge: per report processed, per patient surfaced, per service line, per bed or facility, or a platform fee, with the staffed services priced separately by nature. Nothing states implementation fees or term, and nothing says whether the partner imaging integrations carry their own charges.

Started in oncology and has broadened quickly, most of it this year. The core is the service line from the report that carries a new cancer diagnosis through navigation, tumor boards, follow up and registry reporting. The April 2026 acquisition of Thynk Health added lung cancer screening and incidental findings programs, now running across lung, breast, pancreas, liver, renal, gynecological and neurological findings; the current menu adds cardiology, emergency department discharge continuity and clinical research, and the August 2026 imaging partnerships bring partner findings into the same workflows.

The registry work is sold as a product in its own right: the Registry Automation solution automates cancer case finding across disease sites in real time, lets registrars review and confirm or flag each case with structured feedback to the model, writes the captured information into the registry database and produces the data accreditation surveys require. The same work is sold as a staffed service, cancer registry case finding, and the company names registry partnerships for pre populating national registry submissions. That is registry abstraction sold to providers with results attached.

The setting is the United States hospital and health system, with navigators, registrars and tumor board teams as the named users; no ambulatory or international deployment appears. The breadth outside oncology arrived within months, by acquisition, launch and partnership rather than demonstrated deployment.