What does data annotation actually involve?
Someone, or something, attaches the right answer to raw input: this invoice field is the total, this sentence is a complaint, this pixel region is a defect. Machine learning then generalizes from those examples. The format varies, from class labels and bounding boxes to entity spans and ranked comparisons, but the economics are constant: labels are the part of an AI project that scales with human hours, which is why annotation grew into a global outsourcing industry and why the gig platforms behind it dominate what people search for around the term.
Why do AI models still need annotated data?
Foundation models shifted the work rather than removing it. Companies rarely label millions of images anymore, but they still label what the base model cannot know: their own documents, their own defect types, their own definition of a good answer. Fine-tuning needs labeled examples, evaluation needs graded outputs, and preference training needs human rankings. The label budget moved from pretraining to the last mile, where a few thousand careful judgments from people who know the domain beat a million generic tags.
How do teams keep annotation quality high without labeling everything?
Three levers. Active learning lets the model pick the examples it is least sure about, so humans label the cases that teach the most. Synthetic data stretches a small labeled set by generating variations of it. And review stays a human-in-the-loop step on a sample of everything automated, because bad labels do not fail loudly, they train a confident model to be wrong. Measure agreement between annotators before trusting any label set; where experts disagree, the model has nothing clean to learn.
What should a buyer know before paying for annotation?
That you are buying judgment, not keystrokes. A vendor quote per thousand labels means little without the error rate, the reviewer layer and the domain knowledge behind it, and a cheap label set that needs relabeling costs more than an accurate one. Specify the gold standard first: a few hundred examples labeled by your own experts, used to test every batch the vendor returns. For most mid-sized companies the right split is a vendor for volume, in-house experts for the gold standard and the edge cases.