High-quality labeled data is the foundation of AI. Labeling platform comparison: Label Studio (open source): flexible, supports text/image/audio/video, customizable interface, REST API for integration, excellent for teams. Prodigy (commercial): integrated ML-assisted labeling, excellent for NLP, active learning built-in. Scale AI (commercial): managed annotation workforce, highest throughput, strict QA, expensive. Annotation workflow: 1) Create labeling guidelines document with examples and edge cases. 2) Run training session with annotators. 3) Start with batch of 100 examples, measure inter-annotator agreement. 4) Iterate on guidelines until IAA > 0.7 kappa. 5) Scale to full dataset. Active learning with Label Studio + scikit-learn: train initial model on 200 examples, predict on unlabeled pool, select most uncertain examples (entropy-based), add to annotation queue. Reduces total annotation by 40-60% while maintaining model quality. Quality control: gold standard items (known-correct examples) in every annotation batch to measure annotator accuracy. Remove annotators below 85% accuracy on gold standards. Disagreement resolution: for critical labels, require 3 annotators + majority vote or expert adjudication. Annotation economics: $0.05-0.50 per label for platforms (Scale, MTurk), $15-50/hour for expert annotation. Budget upfront: good dataset requires 3-10K labeled examples, $5K-50K total cost.