Azra AI

The natural language 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 process it replaces. The models are stated to be trained on more than 100 million reports through an exclusive partnership with a large health system, and they identify and classify cancer diagnoses and characterise incidental findings in real time rather than in overnight batches, which is what allows a navigator to reach a patient within hours.

Held at B rather than A because a large amount of what a customer buys sits around the model rather than inside it: navigation, tumour board management, registry abstraction, analytics and consulting, and the company's own positioning has moved toward care orchestration and platform language. The model is the trigger; the platform is the product.

Nothing clinical is decided automatically. The system surfaces a case into a work queue and a navigator, registrar or clinician acts, so the model changes who gets attention and when rather than what the diagnosis is. One configuration detail deserves more scrutiny than it usually receives, because it moves an important decision to the buyer: the case queue can be tuned by each health system to raise or lower the volume of reports surfaced.

That is a sensitivity setting, and a hospital tuning volume down to protect navigator capacity is directly choosing to see fewer patients, including some real ones. Nothing published describes what is lost at each setting, so the tradeoff is being made without the numbers that would inform it, and it is being made by the customer rather than the vendor.

More provenance than most and the wrong half of the performance. The company names the scale of its training corpus, at more than 100 million pathology and radiology reports, and names its origin as an exclusive partnership with a large health system, which is more than most vendors in this index disclose about where their data came from. No architecture, validation methodology, publication or independent evaluation was located.

The specific gap is the published metric: precision is stated at up to 98 percent, and precision is the number that describes how many flagged cases were genuine. For a product whose purpose is to stop patients being missed, the number that matters is recall, which describes how many genuine cases were never flagged at all.

Publishing precision without recall reports the errors a hospital will notice and omits the ones it never will, and that is the same structural gap this index recorded against a cancer prediction platform's sensitivity claim and answered well by a colonoscopy vendor's miss rate trial.

The corpus is named at both scale and origin, which is more than most of this index discloses, and the origin is what raises the question. The models are stated to have been trained on more than a hundred million pathology and radiology reports obtained through an exclusive partnership with a large health system. Naming the source institution rather than describing a proprietary dataset lets a reader locate where the material came from.

What is unstated is everything that would make it assessable: whether the reports were de identified and by what method, whether patients were informed, whether the arrangement was research or commercial, what lawful basis applied, and who owns the resulting model.

The commercial consequence is worth naming plainly and without accusation, because it follows from the disclosed facts rather than from any allegation: a model built on one health system's clinical text is now sold to that system's competitors, and the originating patients are not party to any of it. That arrangement may be entirely lawful and properly consented, and nothing published lets a reader establish either way.

Two passes located no retention position, minimisation statement or secondary use policy, which is a significant absence for what may be the largest real time flow of clinical free text in this index. Ask for the lawful basis and de identification method behind the corpus, model ownership, retention, and whether customer reports contribute to further training.

Operational scale is real and independently improbable to fake; clinical evidence is absent. The company states deployment at hundreds of hospitals including five of the ten largest United States health systems and real time processing of well over half a billion clinical reports and messages annually, and names customers including MultiCare Health System with stated gains in time to treatment, navigator time with patients and oncology patient retention.

Those outcome figures are vendor published and carry no denominators or comparison periods. Two retrieval passes located no peer reviewed publication, no independent evaluation and no registered study. The scale of use does not substitute for evidence of benefit precedent applies directly: knowing that a system reads half a billion reports says nothing about how many patients reached treatment sooner because of it, and time to first treatment is a measurable outcome with an established literature, so the study is available to be run.

Two passes located no retention position, minimisation statement or secondary use policy, which is a significant absence for what may be the largest real time flow of clinical free text in this index. One question is specific and unanswered: the models are stated to have been 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, whether patients were informed, whether the arrangement was research or commercial, or who owns the resulting model. The commercial consequence is worth naming plainly and without accusation: a model built on one health system's clinical text is now sold to that system's competitors, and the originating patients are not party to any of it.

Two retrieval passes located no HIPAA statement, no Business Associate Agreement terms and no privacy or legal page. Graded on published posture. Agreements certainly exist, since the platform receives live interface feeds inside hundreds of hospitals and could not operate otherwise, so this describes what a buyer can read before contracting rather than what is contractually true.

A vendor with this install base is also the one for which published terms would do the most good, because it is being evaluated by many procurement teams at once and each is reconstructing the same answers privately.

Two passes located no SOC 2, no HITRUST, no ISO 27001, no trust centre and no vulnerability disclosure policy. The scale is what makes the absence notable rather than routine. A system taking live interface feeds inside hundreds of hospitals, including five of the ten largest health systems in the country, and processing over half a billion clinical reports and messages a year, is both a concentration of sensitive clinical text and a set of live connections into hospital networks. Concentration and connectivity are the two properties that make a vendor attractive to attackers, and this vendor has both without publishing anything about how it manages them.

No clearance, authorisation or submission located and none claimed. Identifying already diagnosed patients from reports and routing them into navigation and registry work is administrative, and sits comfortably outside device regulation. Two developments move the product closer to a boundary it has not yet crossed, and a buyer should watch them rather than assume the original position holds.

Incidental findings management, expanded through the 2026 acquisition, is a patient safety function rather than a clerical one, since the failure it prevents is a suspicious finding noted in a report and never followed up. And the 2026 research platform describes real time models detecting and characterising disease across cardiology and neurology, which is a broader claim than classifying a diagnosis somebody else has already recorded. Characterising disease is closer to the line than finding the patient who already has it.

Nothing published, and 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, by subspecialty and by individual reporter in vocabulary, structure and what is left implicit.

A model trained on the corpus of one large health system and deployed at hundreds of others faces exactly that distribution shift, and a report style the model handles less well produces silently missed patients rather than visible errors. No per site validation process, no drift monitoring, no subgroup performance, no model card and no bias testing methodology was located.

This index has recorded the same documentation variation problem in trial matching, and it is a stronger concern here because the volume is larger and nobody downstream sees the reports that were never surfaced.

The published metric is the wrong half of the performance, and this is the clearest example of a pattern running through much of this backfill. Precision is stated at up to ninety eight per cent, and precision describes how many flagged cases were genuine. For a product whose entire purpose is to stop patients being missed, the number that matters is recall, which describes how many genuine cases were never flagged at all.

Publishing precision without recall reports the errors a hospital will notice and omits the ones it never will: a false positive arrives as a case a navigator reviews and discards, visible and countable, while a false negative is a patient with a new cancer finding who is never entered into the pathway and who generates no artefact anywhere. A hospital running this for a year sees the cases it caught and cannot see the cases it did not.

The figure also uses the ceiling construction this index treats as an absence of a number, since up to ninety eight per cent is satisfied by any lower value. No architecture, validation methodology, publication or independent evaluation was located, and no warranty, indemnity or remediation commitment. This index has recorded the same structural gap elsewhere and has also recorded the vendor that answered it properly with a published miss rate trial. Ask for recall against a manually reviewed sample, the false negative rate by finding type, and how missed cases would ever be detected.

The deepest live integration footprint on this list and among the deepest in the index. The platform consumes pathology reports, radiology reports, admission, discharge and transfer feeds and other clinical data in real time as they enter the record, which is a materially harder engineering and operational problem than a nightly extract, because it means standing inside the hospital's interface layer and staying correct as that layer changes.

It is stated to run at hundreds of hospitals including five of the ten largest United States health systems, and to process well over half a billion clinical reports and messages annually. It also writes outward into a cancer registry in the national reporting format, so the integration is bidirectional and lands in a regulated downstream system rather than in a dashboard. Very little in this index operates at that combination of breadth, depth and latency.

Delivered as a hosted platform receiving live interface feeds from customer systems, with no published architecture, named hosting region or residency commitment located in two passes. The unusual feature of this deployment is volume rather than location: clinical free text is streamed continuously out of hundreds of hospitals to a single vendor environment, which is a different exposure profile from a product that queries a system periodically or runs inside it. Whether reports are retained after processing, and for how long, is the question that determines the size of that exposure, and it is unanswered.

No pricing, pricing mechanism or contracting model located. One bookkeeping observation is recorded neutrally because a buyer assessing supplier durability would want it: venture databases show total funding of roughly 8 million dollars across two rounds, with the most recent under 2 million in mid 2025, which is difficult to reconcile with an install base of hundreds of hospitals, a claim to the largest position in its market segment, and an acquisition completed in April 2026.

The most likely explanations are that the company is substantially revenue financed, or that its capital structure sits outside what venture trackers capture, which its former identity as a holding company would be consistent with. Neither is a criticism; the point is that the public record does not explain how the business is funded, and a health system placing a core clinical workflow on a supplier should establish that directly.

Started narrow and has broadened quickly. The original scope was oncology across the whole service line, from a suspicious report through navigation, tumour boards and registry reporting to survivorship. The April 2026 acquisition of Thynk Health added lung cancer screening programme management and incidental findings follow up, which extends the product beyond patients who already carry a diagnosis to patients whose only signal is a line in a report written for another reason.

The May 2026 research platform adds trial matching and extends report level detection into cardiology and neurology. Setting is the United States hospital and health system, with no ambulatory or international presence located. Held at B because most of the breadth arrived within the last four months, by acquisition and by launch rather than by demonstrated deployment.