The four sources of training data
- Public web: crawled pages, code and books, increasingly limited by opt-outs and lawsuits.
- Licensed data: archives from publishers, platforms and businesses, used under contract.
- Human-created data: examples written, ranked or corrected by paid specialists.
- Synthetic data: examples generated by models or simulators.
Why licensed business data is in demand
Epoch AI estimates the stock of usable public human text could be fully used between 2026 and 2032 (Epoch AI). Web data also says little about how professional work is done inside companies. Business records — tickets, drawings, logs, reports — fill that gap, which is why labs and their data partners now license them.
Pre-training, fine-tuning and evaluation
Pre-training uses huge, broad datasets. Fine-tuning and reinforcement learning use smaller, higher-quality sets of examples, often from experts. Evaluation sets test whether a model actually works. Business records are useful for all three, especially the last two. See what data AI companies buy.
How a business supplies training data
A business licenses a defined, de-identified copy of its records to a buyer under a written agreement, and keeps the originals. Read how to sell data to AI companies for the full process, or take Calto Data’s free two-minute check to see whether your records fit current requests.
Frequently asked questions
What is LLM training data?
The text and other data a large language model learns from: web pages, books, code, licensed archives, expert-written examples and synthetic text.
Who supplies AI training data?
Web crawls, publishers and platforms, data labeling and expert networks, data licensing partners, and businesses licensing their own records.
Is using business data for AI training legal?
Generally yes when the business owns the data, personal information is removed or has a legal basis, and a licence sets permitted uses. This is general information, not legal advice.