Rad AI

Rad AI's generative models are the engine of its core module, and the portfolio carries value beyond them. Rad AI Impressions generates the report impression from the radiologist's dictated findings with the company's own generative models and would not exist without them.

Rad AI Reporting is described as AI first, cloud native reporting, a radiology reporting system with templates and real time quality checks before sign off, and Rad AI Continuity manages follow up of recommendations through the EHR and RIS; both carry workflow value a practice uses whatever the model drafts.

The workflow puts a radiologist on every report, and the limits of the checks around it are not published. The company's FAQ states that a licensed radiologist must always review the AI generated impression, edit it as needed and sign the final report; Rad AI Reporting runs real time quality checks before sign off, and the security page describes human in the loop quality assurance since 2019.

The company also states the tool catches clinically significant errors in the radiologist's dictation in about 5 percent of reports. The oversight structure is described, and the review step is universal. What is missing is the rest of the structure: what the quality checks look for and when they flag, what the system is not used for, and what happens when a generated impression or a finding carried forward from a prior report (Omni Unchanged) proves wrong, a failure the signing radiologist can miss because the text reads like an ordinary prior finding.

Rad AI Continuity is sold as follow up management 'so every recommendation is tracked, communicated and completed'; the company's own figure, follow up completion rising from about 30 percent to over 75 percent, shows that the result is not every recommendation.

The approach is named and the data behind it is described in general terms. The company states it runs its own generative models trained for radiology rather than general purpose models; its FAQ says Rad AI Impressions is trained on a large, diverse corpus of de identified historical radiology reports, with no images, and that for each site three to five years of de identified historical reports are used to customize the model to individual radiologists' language, style and reporting patterns. No architecture, model version or update practice is published, and no benchmark or validation method accompanies the models.

The model layer is short. Rad AI runs its own generative models trained on de identified radiology reports rather than calling an outside model provider, so no third party model vendor receives report text. The product is cloud native, and no infrastructure provider or region is named on the company's pages; the security page shows a set of standards marks under 'We Comply With These Standards' without saying which belong to Rad AI and which to its cloud host. The company says each site's de identified historical reports are used to customize the model for its radiologists, and whether those reports also enter the shared training corpus is not stated.

The figures published are vendor reported without method. The company reports that health systems using Continuity raised follow up exam completion from about 30 percent to over 75 percent, a median of one hour saved per radiologist shift, up to 35 percent fewer dictated words, reports generated up to 50 percent faster, and clinically significant dictation errors caught in about 5 percent of reports; none names the site, period, denominator or method behind it.

The company names more than 200 health organizations as users, including AdventHealth, Advocate Health, Emory, Inland Imaging, LucidHealth and Trinity Health, and reports more than 11,000 radiologist users; that scale of use is not evidence of benefit.

De identification and training use are stated in specific terms; retention is not. The security page describes a de identification pipeline specialized to radiology reports, and the FAQ says the models are trained on de identified historical reports, with no images, and customized for each site from three to five years of its de identified reports.

No retention period for report text or derived data is published, no standard for the de identification (safe harbor or expert determination) or independent check of it is named, and whether a site's reports are pooled into models serving other customers is not stated. A pipeline built at this volume indicates routine secondary use of customer reports, an inference the company's own statements support without quantifying. Safety engineering, meaning guardrails and how a safety event is handled, is not published beyond the real time quality checks the reporting product runs before sign off.

The company claims HIPAA compliance and publishes no business associate position or instrument on its open pages. The FAQ says Rad AI is HIPAA compliant with enterprise grade security controls, and the security page describes the company as SOC 2 Type II 'HIPAA+ certified'; HIPAA has no certification scheme, so the phrase describes a compliance program rather than a credential.

A cybersecurity role advertised on the company's careers pages includes managing business associate agreements and coordinating HIPAA risk analyses. The privacy policy linked from the footer is a general website document. The company's trust center (trust.radai.com, on Vanta) builds its content in the browser and is where agreement templates and compliance documents would be listed.

A recognized attestation is named, and the report sits behind a request. Rad AI's security page states SOC 2 Type II covering security, confidentiality and availability, continuing twelve month third party audit cycles, real time control monitoring with more than 130 tests a day, alerting, logging and anomaly detection, and formal security policies available on request. Its trust center (trust.radai.com, on Vanta) may hold the report itself.

The security page also carries two wording problems. It calls the company SOC 2 Type II 'HIPAA+ certified', when SOC 2 is an attestation and HIPAA has no certification, and it shows a set of standards marks under 'We Comply With These Standards' without separating Rad AI's own attestation from its cloud host's program.

Rad AI states no regulatory position, and none of its products holds an FDA clearance or authorization. The products are reporting and follow up tools: the impression is drafted from the radiologist's own dictated findings, which the radiologist reviews, edits and signs, so the basis of the output is in front of the clinician. Two features sit closer to the device line and are worth the company addressing.

The impression is the diagnostic conclusion of the report, and the company markets catching clinically significant errors in the radiologist's dictation in about 5 percent of reports, a function stronger than reformatting language. Whether both stay outside the device definition turns on the clinical decision support criteria, including whether the clinician can independently review the basis of the output, and the company does not discuss them. The FDA cleared NinesAI and NinesMeasure algorithms were acquired by Sirona Medical in 2022 and have no connection to this company.

No subgroup analysis, bias assessment, fairness testing or model card is published. Two features of the product make the gap matter. The model is customized to each radiologist's historical reports, and a model fitted to an individual's habits reproduces that individual's systematic tendencies, including habitual omissions and hedges, in the report's conclusion.

And the measurement published runs one way: the company reports catching clinically significant errors in about 5 percent of dictations, with no rate for generated impressions that omit a finding present in the dictation, or for stable findings carried forward from a prior report after they have changed.

Mechanisms exist that let a radiologist catch an output, and nothing stands behind it. A licensed radiologist must review, edit and sign every report, and the reporting product runs real time quality checks before sign off. The company's figure, clinically significant errors caught in about 5 percent of reports, has no definition of clinically significant, no denominator and no independent check, and it measures errors in the radiologist's dictation rather than errors the model makes, so it is not an error rate for the product. No warranty, indemnity or remediation commitment is published.

Named reporting systems and standard interfaces are published. The FAQ says Rad AI Impressions integrates with radiology reporting systems such as PowerScribe and Fluency through a desktop client, that Rad AI Continuity integrates with the EHR and RIS through standard HL7 results, orders and ADT messages, and that the company uses DICOM, HL7 and FHIR; more than 200 health organizations are named as users. No EHR vendor is named and no marketplace listing is cited, so bidirectional integration with a named record system is not verifiable.

A single hosted option is described and its location is not. Rad AI Reporting is cloud native, and the impressions product runs through a lightweight desktop client with single sign on and no servers or virtual machines to set up at the site, so report text leaves the institution with each study. No cloud provider, region, residency commitment or tenant isolation is published.

No price, tier or unit of charge is published. Rad AI sells through agreements with radiology practices and health systems, and inquiries go through a demo request.

Radiology reporting and follow up, scoped clearly by function and buyer, with the modality scope unstated. The products cover report generation, impression drafting and follow up of incidental findings, sold to hospital radiology departments, physician owned practices and teleradiology groups; named users include AdventHealth, Advocate Health, Emory, Inland Imaging, LucidHealth and Trinity Health, and four health systems are strategic investors.

The models learn from reports rather than images, and nothing states which modalities, subspecialties or body regions they are trained and validated on, although report language varies sharply between them.