AI Act Transparency

Under Annex III of the AI Act, AI systems intended to be used as safety components in the management and operation of critical infrastructure, or as medical devices, are classified according to their risk profile. Cavitech AI contains both high-risk and minimal/no-risk AI components:

High-Risk AI Systems (Annex III, Point 5(a))

Our radiograph analysis pipeline, which provides automated detection and classification of dental pathologies on X-ray images, falls within the scope of Annex III, point 5(a) as an AI system intended to be used as a component of a medical device or as a medical device in its own right, specifically a clinical decision support tool for dental diagnostics. This classification applies because the system processes patient radiographic data and produces findings that inform clinical decision-making by dental professionals. We treat this component with the full rigour required by the AI Act for high-risk systems, including conformity assessment, risk management, technical documentation, and ongoing monitoring.

Minimal/No-Risk AI Systems

Several operational features within Cavitech AI are classified as minimal or no risk under the AI Act. These include: appointment scheduling and calendar management; ambient clinical note transcription (Scribe), which converts speech to text and generates structured SOAP notes for review; conversational AI assistants that provide general dental knowledge retrieval; and treatment plan formatting tools that organise approved clinical findings into structured documents. These systems do not make autonomous clinical decisions, do not process patient data for diagnostic purposes, and operate solely as productivity and administrative aids under the direct supervision of dental professionals.

Cavitech AI deploys several distinct AI systems, each with a defined scope of capability and intended purpose. This section describes what each system does, how it operates, and the boundaries of its function.

Radiograph Analysis

The radiograph analysis system is designed to assist licensed dental professionals in identifying potential pathologies on dental X-ray images. The system is capable of detecting up to 28 distinct pathology types, including but not limited to: caries (incipient, moderate, and deep), periapical lesions, periodontal bone loss, root resorption, impacted teeth, retained roots, calculus, widened periodontal ligament space, root fractures, and dental anomalies. For each detected finding, the system provides a confidence score expressed as a percentage, indicating the model's statistical certainty. Findings are localised to specific teeth using FDI (Federation Dentaire Internationale) notation and annotated directly on the radiographic image. The system operates as clinical decision support only. It does not render diagnoses, prescribe treatment, or replace the clinical judgement of the treating dentist.

Ambient Scribe

The Ambient Scribe system performs real-time speech-to-text transcription of clinical encounters and generates structured SOAP (Subjective, Objective, Assessment, Plan) notes from the transcribed content. The system supports 99+ languages, including Afrikaans and Arabic. The Scribe system does not interpret clinical content, make clinical suggestions, or generate diagnostic conclusions. It is a documentation aid that produces draft notes for the dentist's review, editing, and approval before they are finalised. No clinical decisions are made or influenced by this system.

Treatment Planning

The treatment planning system generates phased treatment plans based exclusively on findings that have been individually reviewed and approved by the treating dentist through the mandatory approval workflow. The system organises approved findings into logical treatment phases, suggests sequencing, and formats the plan for clinical documentation and patient communication. The dentist retains full control over all aspects of the treatment plan, including the ability to add, remove, reorder, or modify any element. The system does not autonomously determine treatment approaches or clinical priorities.

In compliance with Article 14 of the AI Act, Cavitech AI has been designed and developed to ensure effective human oversight throughout the entire clinical workflow. The platform implements the following safeguards:

Mandatory Approval Workflow

No AI-generated finding is acted upon without explicit human approval. Every finding produced by the radiograph analysis system enters a "pending" state and must be individually reviewed by a qualified dental professional. The dentist must explicitly approve or decline each finding before it can progress through the clinical workflow. Only approved findings are included in clinical reports, treatment plans, or patient-facing documentation. Declined findings are excluded from all downstream processes.

Dentist Override & Control

At every stage of the workflow, the treating dentist maintains the ability to: override any AI-generated finding or suggestion; modify the details, classification, or severity of any finding; reject findings entirely, with or without providing a reason; add findings that the AI system did not detect; edit AI-generated clinical notes, SOAP notes, and treatment plans before finalisation. The AI system is designed to augment, not replace, clinical judgement.

No Autonomous Clinical Action

The AI system cannot autonomously initiate any clinical action. It cannot send reports to patients, submit insurance claims, prescribe treatment, or communicate findings to third parties without the explicit action and authorisation of the treating dentist. Every output of the AI system requires human initiation, review, and confirmation before it has any clinical or administrative effect.

In compliance with Article 10 of the AI Act, Cavitech AI maintains rigorous data governance practices for the datasets used to develop and validate our AI models. The following principles govern our approach to training data:

No Patient Data Used for Training

Patient data uploaded to the Cavitech AI platform is never used to train, fine-tune, retrain, or otherwise improve our AI models. Patient data is processed solely and exclusively to deliver the analysis service to the requesting dental professional. This separation is enforced at the architectural level, and patient data does not flow into any training pipeline. This commitment is absolute and applies regardless of jurisdiction, consent status, or anonymisation.

Curated & Annotated Datasets

Our AI models are developed using purpose-built, curated datasets of annotated dental radiographs sourced from established academic and research datasets. All images in our training data have been annotated by qualified dental professionals and reviewed through multi-stage quality assurance processes. Annotations follow standardised dental nomenclature (FDI notation) and pathology classification frameworks to ensure consistency and clinical accuracy.

Data Quality & Bias Mitigation

We conduct ongoing data quality reviews to identify and address potential sources of bias in our training data. This includes assessment of demographic representation, image quality distribution, pathology prevalence balance, and annotation consistency. Where gaps or biases are identified, we take corrective measures including targeted data augmentation, re-annotation, and model revalidation. We maintain documentation of data provenance, annotation methodology, quality metrics, and known limitations for all datasets used in model development.

Documentation of Provenance

Complete records are maintained for all training datasets, including: the source and licensing terms of each dataset; the annotation protocol and qualifications of annotators; quality assurance procedures applied; known limitations or biases; version history and changes over time. This documentation is available to regulatory authorities upon request as part of our technical documentation obligations under the AI Act.

In the interest of transparency and in compliance with the AI Act's requirements for clear communication of system limitations, Cavitech AI discloses the following known limitations and associated risks of its AI systems:

False Positives & False Negatives

Like all AI-based detection systems, the radiograph analysis pipeline may produce false positives (identifying a pathology where none exists) and false negatives (failing to identify a pathology that is present). Neither outcome should be treated as definitive. All AI findings must be independently verified by the treating dentist through clinical examination and professional judgement.

Image Quality Dependency

The accuracy of AI-generated findings varies significantly based on the quality of the input radiograph. Factors that may reduce accuracy include: low resolution or poor contrast images; motion artefacts or patient positioning errors; overexposure or underexposure; image compression artefacts from file conversion; non-standard radiographic projections or unusual anatomical presentations. The system is optimised for standard dental radiographic formats, including periapical, bitewing, and panoramic (OPG) images captured with modern digital sensors.

Absence of Clinical Context

The AI system analyses radiographic images in isolation. It does not have access to the patient's clinical history, symptoms, physical examination findings, or other diagnostic information that a treating dentist would consider when making a clinical assessment. This means the AI system may flag findings that are clinically insignificant in context, or may not adequately weight findings that are clinically important given the patient's specific situation.

Confidence Scores Are Not Certainty

Confidence scores represent the statistical probability assigned by the model to a particular finding. A high confidence score does not guarantee that a finding is clinically correct, and a low confidence score does not guarantee that a finding is incorrect. Confidence scores should be interpreted as one input among many in the clinical decision-making process, not as a measure of diagnostic certainty.

Optimised for Standard Formats

The system has been trained and validated primarily on standard dental radiographic formats and projections. Performance may be reduced when processing non-standard image types, extraoral radiographs not within the system's intended scope, or images from significantly older or non-digital imaging equipment. Users should exercise additional caution when using the system with image types outside its primary validation scope.