Overjet

Vision AI on dental radiographs is the engine, and the platform now sells more around it. The core is FDA cleared radiograph analysis that detects and outlines disease and anatomy: caries, periodontal bone level, calculus, periapical radiolucency, charting and, since December 2025, CBCT, each cleared as its own device.

Around it the home page now sells IRIS, imaging software for capturing and enhancing radiographs; Insurance Verification, automated eligibility and benefits checks; a Voice AI suite for documentation; and DSO Analytics; and the company runs a separate claims review product for dental insurers. Several of those are AI branded, and some (imaging capture, eligibility checks, analytics dashboards) carry value beyond the detection models.

Positioned as concurrent read decision support with the dentist retaining diagnosis. Overjet Caries Assist is cleared as a concurrent read computer aided detection device (K222746): it outlines or fills suspected lesions on the radiograph the dentist is already reading, and the dentist reviews and can edit them, so the output is inherently reviewable against the clinician's own read.

The clearance's reader study shows the effect of that design: with the aid, 28 dentists' sensitivity rose from 64.6 to 78.5 percent on bitewings and from 65.6 to 79.0 percent on periapicals. Two pressures sit on that design. The company markets case acceptance as a use, and an AI overlay shown to a patient during a treatment conversation carries persuasive weight; and the same company's payer product reviews the claims that follow, which creates a parallel incentive question. What is not published is how an edited or rejected finding is recorded, or what a practice should do when an overlay conflicts with the clinical exam.

The regulatory record names the approach; training data and update practice go unpublished. The caries clearance (K222746) describes a machine learning pipeline of four modules: image preprocessing, tooth number assignment, a caries module built as an ensemble of three instance segmentation models, and post processing; it shows the device as a new version of an earlier clearance (K212519), extended to periapical images and more sensor manufacturers.

The company states its models were trained on millions of expertly annotated radiographs and publishes commentary that accuracy varies across pathology types, imaging modalities and practice conditions. Training set composition by source, sensor and population, a version history across the eight cleared modules, and how model updates reach practices are not published.

A public subprocessor list names the host and the model providers, and the terms that govern data once it arrives are not published. The trust center lists sixteen subprocessors, including Google (Google Cloud Platform, the infrastructure provider), OpenAI and Anthropic (large language model providers), Deepgram (speech recognition), LabelBox (data labeling), and operational tools such as PagerDuty, Slack, Salesforce, Gainsight, Auth0, New Relic, Sigma Computing, Notion, Intercom and Good Methods Global, with US and UK locations.

The detection models themselves are Overjet's own cleared devices. What is missing is the terms: nothing states what reaches OpenAI, Anthropic, Deepgram or LabelBox (whether radiographs, transcripts or chart text), for what purpose, under what retention and whether they may train on it.

Regulatory studies with real denominators, company run and retrospective. The caries clearance (K222746) reports a standalone test on 1,293 bitewing and 1,314 periapical images (27,920 and 16,254 tooth surfaces) against the consensus of three US licensed dentists with oral radiologist adjudication, with sensitivity of 76.6 and 79.4 percent and specificity of 99.1 and 99.4 percent, and a multi reader study in which 28 dentists read 330 bitewing and 330 periapical images with and without the aid: AUC rose from 0.729 to 0.785 on bitewings and from 0.799 to 0.848 on periapicals, and sensitivity by about 14 points.

The company reports validation showing accuracy that meets or exceeds general dentist performance and use across large dental groups and insurers covering more than 75 million Americans, which is scale of use rather than evidence of benefit. No independent, peer reviewed evaluation of the deployed product measuring an effect on treatment decisions or patient outcomes is on the company's pages.

Assurances are published, and the questions artificial intelligence raises are not answered. The trust center says Overjet does not sell customer data and does not mine or access it for advertising, encrypts data at rest and in transit, and operates as a business associate.

It does not say whether customer radiographs are used to train, tune or evaluate Overjet's models, how long images are retained, or whether they are de identified before any such use; and its subprocessor list includes a data labeling vendor (LabelBox) and large language model providers (OpenAI, Anthropic) alongside a speech recognition provider (Deepgram), without stating what data reaches them. The boundary specific to this company sits between its two customer groups.

Radiographs captured in a practice are that practice's clinical records, and the same company runs a claims review product for dental insurers. A payer legitimately receives radiographs submitted with a claim; the narrower question is whether images or model outputs generated inside a practice deployment, on patients or procedures never submitted to that insurer, can reach the payer product, and nothing published rules it out. Safety engineering, meaning guardrails and how a safety event is handled, is not published.

Overjet states it acts as a business associate and publishes a HIPAA policy; the agreement itself is not published. The trust center says the Overjet platform is HIPAA ready and that Overjet operates in its capacity as a business associate, complying with the Security, Breach Notification and Privacy Rules, and the site footer links a HIPAA Policy and a privacy policy.

The dual customer model means two distinct business associate relationships: a dental practice is a covered entity whose radiographs Overjet analyzes, and a dental insurer is a covered entity for which Overjet performs claims review. The company therefore holds obligations running to two parties who are adverse in the ordinary course, and the risk to address is not disclosure to an outsider but disclosure across an internal boundary between two legitimate customers. No agreement template, breach notification window or subcontractor flow down is published.

A HITRUST certification is named on the company's own trust center, without the report. The trust center says Overjet completed an independent HITRUST audit and obtained HITRUST certification, and describes annual third party risk assessments, third party network and system penetration tests, vulnerability management with regular patching, a zero trust architecture, audit logs and encryption at rest and in transit, on Google Cloud Platform.

The certification's type, date and scope, and whether the report can be opened without asking, are not stated. The FDA clearances establish that the software performs as labeled; they do not examine encryption, access control or incident response, which is what this certification covers. The dual customer model raises the stakes, since a breach would expose both practice clinical records and payer claims material.

Each function carries its own clearance, and the list can be checked. FDA records list eight Overjet 510(k)s: K210187 Overjet Dental Assist (19 May 2021, periodontal bone level measurement, 21 CFR 892.2050, LLZ), K220928 Calculus Assist (16 December 2022, 892.2070, MYN), K222746 Caries Assist (27 March 2023, 892.2070, MYN; concurrent read detection and segmentation of caries on bitewing and periapical radiographs, ages 12 and over, permanent dentition; predicate the earlier Caries Assist, K212519), K231678 Periapical Radiolucency Assist (21 September 2023), K233590 Charting Assist (23 February 2024, QIH), K241681 Image Enhancement Assist (9 September 2024), K251514 CBCT Assist (5 December 2025) and K253930 Iris Intelligent Imaging System (10 April 2026).

An August 2026 announcement reports a further clearance for real time image quality checking in IRIS; the April 2026 Iris clearance is the latest in the FDA's 510(k) database. Clearances are function by function, so each module deployed maps to its own number and indication. Insurance Verification, the Voice AI suite, DSO Analytics and the payer claims product are not presented as devices.

The lead clearance reports results by lesion type and names its patient and equipment subgroups. The 510(k) summary for Overjet Caries Assist (K222746, March 2023) gives standalone sensitivity by image type and lesion class: on bitewings 79.9 percent for primary and 60.9 percent for secondary caries, 74.4 percent in enamel and 79.5 percent in dentin; on periapicals 79.8 and 77.9 percent, 67.9 and 84.9 percent.

It states subgroup analyses by age, sex, sensor manufacturer and restoration status, with the per group figures for age, sex and sensor not in the summary. The lesion results are the useful part for a buyer: detection of secondary caries on bitewings and of enamel lesions on periapicals is materially weaker than the headline. The company also publishes commentary that dental AI accuracy varies across pathology types, imaging modalities and practice conditions. No governance program, model card or post market monitoring description is published.

A published error rate with its method and denominator is on the record for the lead module, and nothing contractual stands behind it. The 510(k) summary for Overjet Caries Assist (K222746, decision 27 March 2023) is Overjet's own filing for a machine learning detector (an ensemble of three instance segmentation models) that outlines suspected caries on bitewing and periapical radiographs for patients aged 12 and over.

Tested on 1,293 bitewing and 1,314 periapical images against the consensus of three US licensed dentists with oral radiologist adjudication, it reports sensitivity of 76.6 percent (73.8 to 79.4) on bitewings and 79.4 percent (76.1 to 82.8) on periapicals, and specificity of 99.1 and 99.4 percent, falling to 60.9 percent sensitivity for secondary caries on bitewings.

It is a test set result for the cleared indication, not performance in use, and it covers caries only; the bone level, calculus, periapical, charting, enhancement and CBCT modules each have their own clearance. No warranty covering model output, accuracy commitment, indemnity or remediation route is published.

Integration is claimed, and no system is named. Overjet is described as sitting inside existing dental workflow, connecting with imaging systems and practice management software so analysis appears on the radiograph during the exam, and its IRIS imaging software now captures radiographs directly. The payer claims product is a second, separate integration surface. No practice management system, imaging software vendor, interface standard or marketplace listing is named, so which systems are supported in production cannot be verified.

The host is named, and residency is not committed. The trust center names Google Cloud Platform as the infrastructure provider and lists subprocessors with US and UK locations, and describes encryption at rest and in transit, a zero trust architecture and activity audit logs. Real time chairside analysis means images are captured in the practice and results return while the patient is in the chair, consistent with cloud inference.

Nothing published states the Google Cloud region where Overjet processes and stores radiographs, whether processing stays in the country of care, whether tenancy is separated between practices, or how long images persist after a result is returned.

One separation question matters here: the practice and payer product lines serve counterparties in claims disputes, and whether they share infrastructure or a data store, and what technically prevents material from one reaching the other, is an architecture question with contractual consequences.

No pricing is published for either product line, and the structure is discoverable. Overjet sells radiograph analysis, imaging software, insurance verification, a voice suite and analytics to dental practices and DSOs, and a separate claims review product to dental insurers.

Those two customer groups are counterparties in a claims dispute: the practice submits a claim asserting treatment was warranted, the insurer reviews it, and the same company's models inform both sides, often on the same radiograph. The home page frames this as aligning the objectives of dentists and insurers. There are legitimate readings: one consistent standard applied to both submission and review could reduce disputes rather than create them. What is not published is whether the models are the same on both sides, whether calibration differs by customer type, and what separates the two businesses internally.

Single specialty by design, with validation behind the cleared image types. Coverage is dentistry: bitewing and periapical radiographs for caries (validated in patients aged 12 and over with permanent dentition, K222746), periodontal bone level, calculus and periapical radiolucency, charting and image enhancement, and since December 2025 cone beam CT (K251514), which extends the scope beyond two dimensional intraoral imaging.

Buyers span individual practices, dental groups and DSOs, and dental insurers. Validation is published per module in the clearance summaries; the settings (independent practices, DSOs, payer review) are named without results by setting, and the caries study does not report performance for children with mixed dentition, who fall outside the indication.