Resources
AI training data by industry
Vertical-specific requirements for AI training data programs -- compliance obligations, annotator qualifications, and operational considerations by industry.
Healthcare
- What compliance requirements apply to healthcare AI training data?
- Healthcare AI training data programs in the US must comply with HIPAA if they involve Protected Health Information (PHI). This requires a Business Associate Agreement (BAA) with every vendor who handles PHI, technical and administrative safeguards, breach notification procedures, and subprocessor BAA coverage. For EU healthcare AI, the Medical Device Regulation (MDR) and IVDR apply for diagnostic models. For FDA-regulated Software as a Medical Device (SaMD), annotation provenance records are part of the device file.
- What annotator qualifications are required for medical imaging annotation?
- Diagnostic-grade medical imaging annotation (radiology, pathology, dermatology) requires clinician participation -- radiologists, pathologists, or trained medical imaging technicians for initial annotation or validation. Peer review by a second qualified clinician is standard for safety-critical labels. General annotators without medical training produce clinically invalid labels for diagnostic tasks regardless of guideline quality.
- What is DICOM and why does it matter for AI annotation?
- DICOM is the standard format for medical imaging data (radiology, cardiology, pathology). DICOM files contain embedded metadata that may include PHI in header fields. Before annotation, a DICOM anonymization pipeline must strip PHI headers. The specific tags stripped should be documented. Vendors without a DICOM anonymization process are not viable for healthcare imaging programs.
Autonomous Vehicles
- What annotation tasks are required for autonomous vehicle perception AI?
- AV perception training requires: LiDAR 3D cuboid annotation (object detection in point clouds), camera bounding box annotation, sensor fusion annotation (consistent object identity across LiDAR and camera in the same frame), lane and road marking segmentation, drivable area annotation, traffic sign and signal classification, and track identity maintenance across sequences. For multi-sensor stacks, all sensors must be annotated with consistent spatial and identity relationships.
- Why does sensor fusion matter for AV training data?
- AV perception models fuse multiple sensor inputs into a unified scene representation. Training data must be annotated consistently across all sensors -- the same physical object must carry the same identity and attributes in every sensor view of the same frame. Single-sensor annotation (labeling LiDAR independently of camera) produces training data with inconsistent object identities that corrupt fusion model training.
- What is edge case mining and why is it important for AV safety?
- Fleet data is systematically biased toward common scenarios. Safety-critical events -- unexpected pedestrian trajectories, unusual vehicle types, construction zones, adverse weather, sensor failure modes -- are underrepresented in fleet data but overrepresented in incidents. Edge case mining is the systematic collection and annotation of these rare-but-critical scenarios. Models trained without adequate edge case representation fail in exactly the scenarios that matter most for safety.
Foundation Models
- What data types are used to train foundation models?
- Foundation model pretraining uses large-scale text (web crawls, books, code, academic papers), with annotation programs focused on quality filtering, deduplication, and content classification. Post-training (instruction tuning and RLHF) requires demonstration data (high-quality instruction-response pairs), preference data (ranked or rated response comparisons), safety data (examples of aligned and misaligned responses), and red-teaming data (adversarial inputs and preferred responses).
- What is RLHF and why is it used in foundation model training?
- RLHF (Reinforcement Learning from Human Feedback) uses human preference annotations to train a reward model that captures what humans find helpful, harmless, and honest. This reward model then guides policy optimization, steering the language model toward outputs humans prefer. RLHF is the standard method for aligning large language models after pretraining and is a core component of production-grade LLM development.
Fintech
- What AI training data requirements are specific to financial services?
- Financial services AI programs typically require: PCI DSS alignment for programs involving payment card data; CCPA and CPRA service-provider addendums for California consumer data; GLBA-compatible data handling for US financial institution data; SOX audit trail requirements for models used in financial reporting; and for EU programs, DORA (Digital Operational Resilience Act) technical requirements. Document AI for financial documents (invoices, contracts, bank statements) requires annotators with financial domain literacy.
- What annotator skills are needed for financial document AI?
- Financial document annotation (invoice processing, contract review, bank statement extraction, KYC document verification) requires annotators with financial domain literacy -- understanding of invoice structure, payment terms, contract clauses, and financial statement formats. For compliance documents, knowledge of regulatory terminology is required. General annotators produce high error rates on financial documents because correct extraction depends on understanding document structure and financial concepts.
Legal
- What are the annotation requirements for legal AI?
- Legal AI training data (contract review, legal research, compliance monitoring, case prediction) requires annotators with legal domain knowledge. Annotation tasks include contract clause identification and classification, legal argument extraction, case outcome labeling, regulatory compliance mapping, and legal entity recognition. At minimum, annotators need paralegal-level legal literacy; for complex contract or case analysis, law school training or bar admission may be required.
- What confidentiality considerations apply to legal AI training data?
- Legal documents -- contracts, litigation files, regulatory submissions -- often contain attorney-client privileged information, trade secrets, and personal data of parties. Annotation of actual legal documents requires: data processing agreements covering attorney-client privilege implications; annotators bound by specific confidentiality agreements covering legal documents; and access controls preventing annotators from seeing document context beyond their specific annotation task.
Robotics
- What training data does robotics AI require?
- Robotics AI training data includes: RGB and depth camera annotation for object detection and grasp planning; point cloud annotation for 3D environment mapping; human pose estimation annotation for human-robot interaction; demonstration data (teleoperation sequences) for imitation learning; and synthetic-to-real adaptation data. Physical manipulation training often requires data collected in controlled environments matching the deployment setting.
Government
- What data handling requirements apply to government AI programs?
- Government AI programs may require: FedRAMP authorization for cloud-based annotation tools used with US federal data; ITAR compliance for defense-adjacent programs; security clearance for annotators on classified programs; NIST AI RMF alignment for risk management documentation; and Section 508 accessibility compliance for public-facing AI. State and local programs may have additional requirements including state-specific privacy laws.
E-commerce
- What AI training data does e-commerce AI require?
- E-commerce AI programs typically require: product attribute extraction annotation (category, brand, size, color, material) from catalog data; product image annotation (category classification, attribute tagging, defect detection); user intent classification from search and navigation data; recommendation system training data (implicit feedback annotation, cold-start data); and review sentiment and authenticity annotation. At scale, product catalog annotation requires domain-specific taxonomy expertise.
Cybersecurity
- What AI training data is used in cybersecurity AI?
- Cybersecurity AI requires: malware classification annotation (static and behavioral features); network traffic annotation (normal vs. anomalous, attack type classification); log annotation for SIEM and threat detection models; phishing and social engineering content annotation; and vulnerability description annotation for patch prioritization models. Annotation of actual threat data requires annotators with security domain expertise and appropriate data handling agreements.
Explore by industry
Ready to scope a pilot?
Tell us your modality, volume, and languages. We'll return an indicative scope, timeline, and cost band.