Data annotation is the process of labeling raw data - images, video, text, audio, LiDAR, documents - with the tags, boxes, or categories a machine-learning model needs to learn from it. It converts unstructured data into supervised training examples. In regulated settings, annotation is also a governance step: the labels, who made them, and how their quality was measured all become part of the evidence a model’s training rests on.
A model learns from examples. Annotation creates those examples by attaching ground-truth labels to raw data: drawing a bounding box around a tumour, tagging a sentence’s sentiment, marking a pedestrian in a LiDAR scan. Annotation quality sets the ceiling on model quality - a model can only be as good as the labels it learned from.
Image (boxes, polygons, segmentation, keypoints), video (tracking, temporal labeling), text/NLP (entity tagging, classification, sentiment), audio (transcription, diarization), LiDAR/3D (cuboids and point-cloud labeling), and document AI (field extraction). Each has its own tools, taxonomies and quality metrics.
The accepted way to prove label quality is Inter-Annotator Agreement - how consistently independent annotators label the same data - scored with chance-corrected statistics like Cohen’s kappa and Krippendorff’s alpha. In regulated programs, agreement above 0.90 is the evidence-grade bar.
When a regulator or auditor asks how a model’s training data was governed, annotation is the answer - or the gap. Documented labels, measured agreement, and an audit trail of who touched the data turn annotation from a cost center into defensible evidence. That is the difference between "data labeling" and audit-ready data annotation.
Agreement between two annotators, corrected for the agreement expected by chance. κ = 1 is perfect; κ = 0 is no better than chance. Regulated annotation QA contracts above 0.90.
Read More
How consistently independent annotators label the same data. High IAA evidences that labels reflect a shared, documented standard rather than individual opinion.
Read More
Reliability for any number of annotators, tolerant of missing data, across nominal, ordinal, interval and ratio labels - the general case where kappa handles exactly two.
Read More
They are used interchangeably. "Labeling" emphasises attaching a tag; "annotation" often implies richer structure (boxes, segments, relationships). In practice, most teams and tools treat them as synonyms.
Increasingly both. AI (including VLM pre-labeling) drafts labels to remove mechanical work; trained humans verify and correct them. In regulated settings, human verification is the norm because auto-labels alone cannot be defended in audit.
It depends on modality, quality bar, and governance burden far more than on raw volume - which is why serious vendors scope a fixed-price pilot rather than quote a single per-label rate.
IAA scored per cohort, audit trails on every action, evidence exports mapped to EU AI Act Articles 10 & 12. See it on your own data in an evidence-grade PoC.
Certifications & readiness
