Data annotation
Definition
Data annotation
Data annotation is the process of labelling text, images, audio, and video so that machine-learning models can learn from them. Every artificial intelligence (AI) model starts with annotated data, and the quality of that labelling caps how well the model can perform.
Human annotators, and increasingly semi-automated tools, tag each data point with the ground truth an algorithm needs — the object in a photo, the sentiment in a tweet, the intent behind a spoken query. Without those labels, most modern AI has nothing to imitate.
Buyers usually outsource the work. Annotation is labour-intensive but repeatable, which makes it a natural fit for offshore providers in the Philippines, India, and Kenya, where trained teams handle millions of tasks a month at a fraction of Western costs.
Two forces pushed annotation from niche vendor to strategic supply chain. First, large language models multiplied dataset sizes. Second, reinforcement learning from human feedback made annotator judgement a competitive moat.
Key takeaways
- Data annotation labels raw text, images, audio, and video so machine-learning models can learn.
- Buyers outsource most annotation work to specialists like Scale AI, Appen, and Sama, or to offshore BPO teams.
- Cost swings from cents per label for simple image tagging to $20 per label for specialist medical or legal work.
- Model-assisted labelling now handles easy cases automatically, while humans verify outputs and resolve edge cases.
- Quality standards from bodies like NIST treat annotation accuracy as a core input for downstream AI model evaluation.
How it works
Data annotation runs through four stages: define the schema, sample the raw data, apply labels, then review for accuracy. Each stage feeds a quality-assurance loop that catches drift before it poisons the downstream machine learning model.
Common techniques include bounding boxes for object detection, polygon segmentation for autonomous-vehicle vision, named-entity recognition for text, and audio transcription for voice models. The mix depends on what the model must predict.
| Stage | What happens | Typical output |
|---|---|---|
| Schema design | Team defines the label taxonomy | Annotation guidelines doc |
| Sampling | Raw data is sliced by class balance | Batches ready for labelling |
| Labelling | Annotators tag each item to spec | Labelled dataset |
| Review | QA reviewers spot-check labels | Inter-annotator agreement score |
Quality standards vary by industry. The NIST AI Risk Management Framework treats annotation quality as a measurement input for downstream evaluation, while healthcare buyers layer on HIPAA and FDA constraints.
In 2024, the Stanford AI Index reported that training a frontier model like Google’s Gemini Ultra cost around $191 million — a large slice of which was the labelling and evaluation of the underlying dataset.
Annotation looks like a commodity job. Look closer and it splits into hundreds of specialisations, from LiDAR point clouds for autonomous driving to radiology mark-up for medical AI and multilingual sentiment tagging for chatbots — each with its own quality bar.
Examples
Data annotation shows up in every major AI product shipped in the past decade. Waymo’s self-driving cars, OpenAI’s ChatGPT, and TikTok’s recommendation engine each rely on billions of human-tagged examples, and most of that tagging happens through outsourced teams.
Scale AI built a $14 billion business by 2024 supplying data annotation to defence and enterprise buyers, including the US Department of Defense and OpenAI, using tens of thousands of contractors in Manila, Nairobi, and Caracas.
Sama, headquartered in San Francisco with delivery hubs in Nairobi and Kampala, ran the data annotation contract that Meta used to train content-moderation classifiers in 2023, employing roughly 3,000 annotators across East Africa.
Appen, an Australian firm listed on the ASX, generated $273 million in 2023 revenue from data annotation for Microsoft, Google, and Amazon, though its share price fell after the same three clients cut spend to move work in-house.
Labelbox and Snorkel AI, both US-based platform vendors, provide the software layer outsourced annotation teams work inside, with Labelbox securing a $110 million Series D in 2022 and Snorkel expanding into enterprise LLM fine-tuning through 2024.
Related terms
- Machine learning: the modelling technique data annotation feeds.
- Business process outsourcing: the broader delivery model annotation sits inside.
- Knowledge process outsourcing: the tier of outsourcing that judgement-heavy annotation belongs to.
- Quality assurance: the review layer that catches mislabels before deployment.
- Content moderation: a downstream use case for the annotated training sets.
- Artificial intelligence: the end product every annotation contract exists to serve.
FAQ
What is data annotation used for?
Data annotation prepares the training sets that supervised machine-learning models learn from. It powers computer vision, speech recognition, and natural-language understanding. Without labelled examples, most current AI systems have nothing to imitate.
How much does data annotation cost?
Prices swing hard by task complexity and location. Simple image tagging runs 5 to 15 cents per label with offshore teams in the Philippines or India, while medical or legal annotation from specialist reviewers can hit $5 to $20 per label.
Who does data annotation work?
Most enterprise buyers outsource it to specialist firms such as Scale AI, Appen, and Sama, or to general BPO providers running dedicated annotation teams. A growing minority use crowdsourcing platforms or bring the work in-house for sensitive data.
Is data annotation being replaced by AI?
Model-assisted labelling now handles the easy cases automatically, though humans still verify outputs and handle every edge case a model cannot resolve confidently.
Compare vetted data annotation partners and outsourcing providers in one place at Outsource Accelerator.
Independent
