What is AI training data? Types, sources and prices

What is training data in AI?

AI developers now pay for real companies' records, such as support tickets, work orders, email and chat archives and code (public deals), as they hit the limits of public and synthetic data. If your company keeps records like these, a free valuation takes six questions about the business and no files.

Training data is the set of examples a machine learning model learns from. The EU AI Act defines it as data used for training an AI system through fitting its learnable parameters: the model adjusts its internal settings until its outputs match the examples. A chatbot learns from text, a speech model from audio paired with transcripts, and a robot from recordings of tasks being done.

Developers keep two other sets apart. Validation data is used to tune the training process, and testing data gives an independent check that the finished system performs as expected before it is released. A model that has already seen its test questions can score well without having learned the skill, so evaluation data must stay separate from training data.

Language models are trained in stages, and each stage uses a different kind of data:

• Pretraining: trillions of words of mostly public text and code, so the model learns language and facts. The most powerful language models are primarily trained on trillions of words of human-generated text from the internet, Epoch AI notes. Models count text in tokens; one token is about 0.8 English words on average.
• Supervised fine-tuning (SFT): prompts paired with good written responses, so the model learns to follow instructions.
• Preference training (RLHF, reinforcement learning from human feedback): people rank model answers, and the model is trained toward the ones they prefer.
• Reinforcement learning in environments: the model attempts tasks in a simulated workplace and an automatic grader scores each attempt. See what RL environments are.
• Evaluation: held-out tests with known answers that measure whether training worked.

Source: EU AI Act, Article 3: definitions of training, validation and testing data

Types of AI training data and what each one trains

Each type teaches a model something different. The public datasets in the last column show what each type looks like at scale. Company records can supply many of them.

For company footage, see selling company video for physical AI. For test sets, see licensing data for AI evaluation.

Source: Penedo et al.: The FineWeb datasets, decanting the web for the finest text data at scale (arXiv, June 2024)

Where does AI get its training data?

Five sources supply nearly all of it. Developers rarely publish exactly what they used: for leading open models, very little is known about how their pretraining datasets were created, the FineWeb researchers note.

Demand for closed companies' data emerged suddenly in 2025, as AI training labs hit the limits of public and synthetic datasets, MIT Technology Review reported in October 2026.

Source: Common Crawl: a free, open repository of web crawl data

Synthetic data vs real data: why buyers still pay for real records

Synthetic data is generated rather than collected: a model writes example conversations, or a simulator produces images and sensor readings. It works best where an answer can be checked automatically. In 2024 Epoch AI found it had only been shown to reliably improve capabilities in relatively narrow domains like math and coding.

It does not replace real data. A 2024 Nature study found that indiscriminate training on model-generated content causes irreversible defects, as the rarer patterns in the original data disappear. The authors call this model collapse and expect data about genuine human interactions to become increasingly valuable. A follow-up study found that collapse follows when synthetic data replaces the real data, and is avoided when synthetic data is added alongside it. Either way, the real data has to stay in the mix.

For workplace AI, real records add what generated data tends to miss. Real archives avoid common errors in synthetic worlds, such as references to dates that don't exist, MIT Technology Review reported, and one buyer it interviewed seeks archives that show disagreements and tricky judgment calls. The same report is candid that the size of the gain is unclear: no company has yet released tests comparing training on real data with the alternatives.

If a buyer wants to generate synthetic data from your records, treat that as a separate right in the license. See what a data licensing agreement should include.

Source: Shumailov et al.: AI models collapse when trained on recursively generated data (Nature, July 2024)

Why real company records are in demand

The public web is finite. Epoch AI estimates the stock of quality, human-generated public text at about 300 trillion tokens, and projects that if trends continue, language models will fully use it between 2026 and 2032, or earlier if models are intensely overtrained.

It is also closing. The Consent in Crisis audit of 14,000 web domains found that in a single year, 2023 to 2024, new restrictions made 5%+ of all tokens in the C4 dataset, and 28%+ of its most actively maintained, critical sources, fully restricted from use. Counting terms-of-service restrictions, 45% of C4 was restricted.

Private data is the obvious next source, but it is hard to use at scale: Epoch AI points to legal issues and to data being fragmented across platforms controlled by owners with competing interests. Licensed company records fill that gap. A company that owns its records, can show it may license them and removes personal information offers what scraping cannot. Buyers pay for data they cannot scrape, and pay more when rights are clear and documented, according to Origin's deals tracker.

The work AI is now asked to do points the same way. After math and coding, enterprise workflows such as filing an expense report, building a pivot table or updating customer records in a CRM are a major growth area for RL environments, Epoch AI found from 18 interviews. Tickets, work orders and CRM histories are records of exactly that work.

Spending may follow. Top AI labs have so far spent far more on computing power than on data, Epoch AI notes, and it expects that to reverse if data becomes the bottleneck.

Source: Epoch AI: Will we run out of data? Limits of LLM scaling based on human-generated data (June 2024)

What does licensed AI training data sell for?

There is no public price list. Most sales of company records are private, and buyers quote wildly different numbers because each values records by its own estimate of how many useful training tasks they will yield. The public deals below show the range; every figure is as reported in Origin's tracker.

• For an operating company, AI companies pay from $20K for a small company to $5M for a large one with a long history. Through Origin, nothing is sold outright: the company keeps its records and licenses a copy for a defined use.
• Buyer programs publish their own ranges. In Origin's study of buyer programs, the typical published range for a company was about $50K at the low end and $843K at the top. These are ceilings set by each program, not offers.
• Market researchers put the whole AI training dataset market at $3.2 billion in 2025, growing to $16.3 billion by 2033 (Grand View Research).
• For what moves the price of your records, see how much AI companies pay for company data.

Source: Origin Data Partners: AI data licensing deals tracker, 2023 to 2026

What makes training data usable?

Volume alone does not make data usable. Buyers check these before they pay.

Source: EU AI Act, Article 10: data and data governance for high-risk AI systems

Does your company have AI training data?

Many established companies do without calling it that. If your business has kept any of the records below for several years, they map onto one of the training data types above.

• Before you delete old systems or let a software subscription lapse, export and list what they hold. Cancelling a help desk, chat workspace or code host deletes the training data with it. Checking takes minutes: a free valuation asks six questions about the business and needs no files.
• Describe the records at company level first: the system, the years covered and the approximate volume. The data inventory template gives the columns, and what company data you can sell to AI compares each record type.
• If the company is closing or has closed, its records can still be licensed. See selling company data after closing.

For AI teams: sourcing licensed company data

If you build models or agents and need real workflow records for training or evaluation, start from the task: the workflow, the information available at the start, the expected action or answer and the way an expert would judge it. Origin can assess whether an introduction to a company with matching records is possible; the company owner must authorize any introduction, and both parties agree permitted use and commercial terms directly. Describe your requirements.

Next step

List the systems your company has used for several years and the records in each: tickets, work orders, chats, code, recordings. Match each one to a type in the table above. No export is needed to find out whether they are worth licensing.

Origin Data Partners is an independent data partner that helps companies get their data bought. It is free to the company that owns the data; a buyer may pay Origin when a deal closes. See what your data is worth.