Why data labeling is a multibillion-dollar business

AI learns from examples that people have labeled. In this post, I explain what data labeling is, why AI depends on it, how much work it takes, and why the companies that do it are worth billions.
What data labeling is
Data labeling means tagging a piece of data, such as an image, a word, a video or a voice recording, with what it really is. It’s mostly done by people, and it’s how an AI system learns what the truth looks like.
Take the video from a self-driving car. Every frame has to be labeled to mark the people, road markings, signs, buildings, trees and bicycles. All of it is done by hand at first so an AI system can learn from it later. Here’s what a labeled image looks like:

How much labeling one car creates
Imagine the work in that single image. Now imagine it’s one frame from a video recorded at 15 frames per second, and the car drives for two hours:
15 images per second × 60 seconds × 60 minutes × 2 hours = 108,000 images from one camera
If the car has nine cameras, that’s 972,000 images from a two-hour drive.
Each image holds several objects, and a person has to find each one, trace its outline, and label it. There are shortcuts, but at the start, that’s how the work gets done.
Why it’s worth so much
With well-labeled data, AI systems like voice assistants and self-driving cars become extremely valuable, and the market has noticed. Scale AI, one of the largest labeling companies, started with images and video and expanded into text. In June 2025, an investment from Meta valued it at more than $29 billion.
How we use labeling services
We rely on labeling services when we need labeled examples to train a model. We send our preferred labeling partner a few examples of correct labels, along with the hundreds or thousands of items we need labeled. They bring in tens or hundreds of people, and within a few days we have training data that used to take us months to produce.
Good labels are part of having data that’s ready for AI. I wrote about the rest of that work in why so much of an AI project goes to cleaning data.
Where labeling is going
When I first wrote this in 2021, I expected labeling companies to use AI on their own work. If they’re labeling trees in hundreds of thousands of images, it makes sense to train a model to find the trees first, so people only check the model’s work.
Labeling companies are here to stay as long as AI needs examples from people. There’s an irony in it, though: the goal of AI is to make this kind of work unnecessary.
Checking AI’s work never goes away. In every system we put into production, Bennett ONE checks every answer against criteria we set with your team before anyone relies on it.