Behold.ai

The model is the product and the business model depends on it. red dot classifies chest X-rays, marking those it judges normal with very high confidence for autonomous reporting and those suggestive of lung cancer for urgent reporting, and a companion product triages CT head scans.

The company positions the service as an alternative to outsourced radiology reporting, claiming a 50 percent cost saving against outsourced reporters, so the AI substitutes for purchased reporting capacity rather than decorating another product. There is no services layer beneath it; remove the model and nothing is left to sell.

red dot acts alone on one class of study. Chest X-rays it identifies as High Confidence Normal are reported autonomously and removed from the reporting workload, about 15 percent of studies according to the product page; those suggestive of lung cancer are prioritized for urgent reporting and same or next day CT; the rest go to radiologists through the trust's RIS or PACS worklist, which red dot updates with a normal or abnormal notification.

The company cites an error rate of 0.33 percent on the autonomously reported studies against 13.5 percent for consultant radiologists, from work published in Clinical Radiology. That is the autonomous channel named and routed, with an error rate attached.

The rest of the structure is not published: the confidence threshold, the studies excluded from autonomous reporting (age, projection, inpatient or portable films), any override path, and any audit of autonomously reported studies after the fact. Because no clinician opens an auto reported film, those controls are the safety case. Who holds clinical responsibility for an auto reported study that proves abnormal is not stated either.

Behold.ai describes what red dot does, not what produces the output. The product and how it works pages describe the classes it assigns (High Confidence Normal, suspected lung cancer), the notification it sends back to the RIS or PACS and the share of studies auto reported, and localization has been shown through heatmaps.

No model architecture, training data (size, sources, trusts, equipment makers, time period), version or update practice is published, and the 2020 FDA 510(k) summary for the pneumothorax triage device describes it only as AI based software. Performance figures, however specific, describe validation rather than the model. For a product that reports studies no clinician opens, knowing which model version made the call, and when it changed, is part of the audit trail a department needs.

The chain is short. red dot is Behold.ai's own model, cleared by the FDA in 2020 as a pneumothorax triage device (K191556) and sold as the company's own classifier, so the model layer is not a third party's. The how it works page states that chest X-ray images and messages reach the red dot platform over an encrypted VPN and that all patient examination data resides in fully NHS accredited data centers.

What is not named is the operator of those data centers, the cloud provider if any, or any subprocessor, and after the 2025 asset purchase it is not stated whether the hosting arrangement carried across from the original company.

Published research and named NHS deployments, tempered by what happened to the product. Two papers in Clinical Radiology underpin it: one on autonomous diagnosis of normal chest radiographs, reporting 15 percent auto reported at a 0.33 percent error rate, and one reporting a 60 percent reduction in missed lung cancers when triage ran alongside consultant radiologists on a tumor enriched dataset.

The company has published real world results from NHS trusts, including a Somerset evaluation of 3,794 images, a Calderdale and Huddersfield report of shorter lung cancer waiting times, and Basildon processing 4,415 chest X-rays in 3.5 days against 4 to 8 weeks by outsourcing, and a Health Research Authority registered evaluation of red dot version 2 was planned.

For CT head the company reports removing more than 60 percent of scans from the workload at a 93 percent negative predictive value, without a source. The counterweight: NICE withdrew red dot from its recommendations in February 2025, and the original company entered administration. A prospective evaluation of autonomous reporting in live use is the evidence the claim calls for.

General assurances, and nothing on what happens to the studies. The how it works page states that images reach the platform over an encrypted VPN and that all patient examination data resides in fully NHS accredited data centers. No retention period for images or classifications, no statement on whether customer studies are used to train or improve the model, and no de identification practice is published.

The retention question matters more here than for an assistive tool: for an auto reported study, the platform's record of what it saw and at what confidence may be the only assessment the examination received, and a later review of a missed finding depends on it being kept. Safety engineering, meaning guardrails and how a safety event is handled, is not published; the confidence threshold is an oversight control, described under autonomy.

The company states no business associate status; its product and technology pages describe NHS accredited data centers and an encrypted VPN rather than a regulatory position. The site also carries a privacy policy, the document where its data protection role for the service and any position on US patient data would be stated.

The deployments on record are in the NHS, where UK data protection law applies, but the home page describes the company as focused on radiologists in the USA as well as the NHS, and its only US clearance is for pneumothorax triage, so the United States position matters.

The contracting entity is the other open question: the original Behold.ai Technologies Limited entered administration in January 2025, a successor acquired the assets, and agreements signed by the original company do not bind the successor automatically.

The product and technology pages describe controls, an encrypted VPN for images and messages and hosting in NHS accredited data centers, and name no information security certification of the company's own. The NHS accreditation described belongs to the data centers, not to the company's operations. Registration with the Care Quality Commission, which the successor has cited, concerns whether care is safe and well led and is not an information security attestation.

NHS suppliers commonly hold a Data Security and Protection Toolkit submission or Cyber Essentials certification; neither is named on the pages describing the product, and the company's privacy policy is the published document where either would most likely appear. The corporate history adds a point: assurance work done by the original company does not carry to the successor automatically.

The company markets a clinical capability beyond any regulatory position it states. What it sells is autonomous reporting of chest X-rays classified as High Confidence Normal, removed from the worklist without a radiologist's read, alongside lung cancer triage, and its home page says it is completely focused on helping radiologists in the USA and the NHS. The pages describing the product state no regulatory position for the seller.

The only FDA clearance for the product is K191556 (28 February 2020, 21 CFR 892.2080, product code QFM), granted to the original Behold.ai Technologies Limited for red dot as a triage and notification tool for adult chest X-rays with features suggestive of pneumothorax: passive notification for worklist prioritization, not for stand alone diagnosis, and not removing cases from the standard workflow.

The original company entered administration in January 2025 and a successor acquired the assets; the successor does not name the clearance, and which authorizations transferred is not stated. The product has also carried CE and UKCA marking, and NICE withdrew it from its recommendations in February 2025. Autonomous normal reporting offered to US radiologists sits outside the only US clearance, which excludes removing cases from the workflow, and nothing published says it is withheld from US sites.

No results by patient group are published for a product that removes studies from human review. The 2020 FDA summary for the pneumothorax triage device reports 888 cases from four sites in the US and UK without subgroup results, and the published error rate for autonomous normal reporting is given for the whole population.

If performance varies by age, sex, body habitus, projection or equipment, patients in the weaker group are the ones whose abnormal film is reported normal and never opened. Chest X-ray classifiers trained on public datasets have underdiagnosed female, Black, Hispanic, younger and Medicaid insured patients (Seyyed-Kalantari and colleagues, Nature Medicine, 2021).

The confidence threshold is a control on how much the model decides, described under autonomy, not an evaluation of how it behaves across groups, and no governance framework or post market monitoring commitment is named.

An error rate with its denominator is on the record in the product's own FDA filing. The 510(k) summary for red dot (K191556, February 2020), filed by the original Behold.ai Technologies Limited whose assets the current company acquired, reports standalone performance on 888 adult chest X-rays from four sites in the US and UK (299 with pneumothorax, 589 without), against a consensus of at least two ABR certified radiologists: sensitivity 94.65 percent (91.46 to 96.91) and specificity 87.95 percent (85.04 to 90.46).

That figure covers pneumothorax triage only, not autonomous normal reporting, and it is a test set result, not performance in use. For autonomous reporting the company cites an error rate of 0.33 percent on the 15 percent of studies auto reported, from work in Clinical Radiology, without the denominator on the product page.

No warranty, indemnity or remediation commitment attaches to the output, and who answers for an auto reported study that proves abnormal is not stated: the radiologist did not read it, the referrer received a normal report, and the company supplied the classification.

red dot writes back into the reporting workflow. The how it works page states that after analysis it sends an electronic notification back to the trust RIS, or to PACS where reporting is PACS driven, marking each examination as normal or abnormal and prioritizing it in the existing worklist; for auto reported studies that notification is the report.

That is one directional writing into the record of the examination, deeper than a read only overlay, and it runs at named NHS trusts, including Somerset, Calderdale and Huddersfield, Basildon and Southend, where the company has published deployment results. No RIS or PACS vendor is named and no HL7 or FHIR interface or connector list is published.

One hosted option with its residency committed. The how it works page states that images and messages reach the red dot platform over an encrypted VPN and that all patient examination data resides in fully NHS accredited data centers, so for NHS customers the data stays in the UK health service's accredited estate.

What is open is the operator of those data centers, tenant isolation, whether a local deployment is offered, and where US studies would be processed given the company's stated US focus. The corporate history makes continuity part of this: the processing location is where the only assessment of an auto reported study was made, and a department needs its own record of those studies if the service ends, as the original company's administration showed can happen.

No price is published, and the pricing basis is stated only relatively: the company positions red dot as an alternative to outsourced reporting, claiming a 50 percent cost saving against outsourced reporters and up to 70 percent lower outsourcing costs. That gives a buyer a benchmark, the per study outsourcing rate, without a figure a buyer can apply. The dominant commercial question is continuity rather than price.

The original company entered administration in January 2025 after failing to win NHS AI Diagnostic Fund contracts, NICE withdrew the product from its recommendations the following month, and a successor acquired the assets and relaunched it.

Two narrow, high volume targets, named and partly validated. On chest X-ray red dot reports normal studies autonomously and triages suspected lung cancer, with published research in Clinical Radiology and deployment results from named NHS trusts behind it; on CT head it triages acute findings and rules out normal scans, for which the company reports removing more than 60 percent of scans from the workload at a 93 percent negative predictive value without a published source.

The settings on record are NHS trust radiology departments. Depth over breadth is a deliberate choice and not a gap. The home page now describes the company's AI as covering all modalities, which goes beyond the two products described. Which referral sources, patient groups and projections the autonomous channel is validated for is not stated.