The Public Web Is Running Out
The first generation of AI models learned from the public internet. That supply is close to used up.
Large language models trained on web pages, books, forum posts and code that anyone could read. Each new generation needed more of it. Researchers at Epoch AI estimate that if current trends hold, frontier models will use up the entire stock of human-written public text sometime between 2026 and 2032.
AI labs saw this coming and started paying for access. In 2024 Reddit licensed its posts to Google for a reported $60 million a year. News Corp signed a deal with OpenAI that the Wall Street Journal valued at more than $250 million over five years. In both cases the buyer paid for text that was already public, for the right to use it and for steady access to new material.
Your company's records sit in a different category. No crawler has seen them and no model has trained on them.
What Models Still Can't See
A model can explain how a refund policy should work. It has rarely watched a team process one.
Public text describes work from the outside: articles, documentation, opinions. Inside a business, work looks different. A customer writes in confused. A support agent asks a colleague. The colleague checks the order, finds the mistake and offers a fix, and the customer accepts. Each step leaves a record, and the records connect.
Frontier AI labs want that chain of problem, action and outcome more than anything else right now. They are training models to power agents that answer tickets, qualify leads, draft follow-ups and route approvals. To do that job well, a model has to learn from thousands of real cases where people did the same work, including the cases where the first attempt failed.
The public web holds very little of this. It lives inside companies, in Zendesk, Salesforce, Gmail and Slack.
You Already Paid to Create It
Your business produces this data as a side effect of running.
Your support team writes replies because customers need help. Your salespeople log calls because someone has to manage the pipeline. Nobody on your team created these records to train AI, and you have already covered the cost in salaries, software and years of operating history.
Once the work is done, most companies leave those records in an archive. They come out for the occasional audit and otherwise cost money to store. Licensing turns them into income without changing how your team works.
History counts. A company that has run its support desk for twelve years holds twelve years of edge cases, policy changes and seasonal problems. A startup can't buy that history at any price, which is why company age is one of the two inputs in our calculator.
Why Buyers Pay
Frontier labs use business records in four ways, and each one creates demand.
Training. Labs use real conversations and workflows to teach their models general skills: following up, resolving a dispute, summarizing a negotiation.
Fine-tuning. After the first round of training, labs tune their models for specific kinds of work, such as insurance claims or B2B sales. That takes examples from the exact job. Generic text won't get them there.
Evaluation. Before a model ships, labs need test cases it has never seen. Private records work well because they can't have leaked into public training sets.
Synthetic data. Labs that generate artificial training examples still need real records to start from. Research published in Nature in 2024 found that models trained repeatedly on their own output degrade over time. Real data keeps synthetic data grounded.
One dataset can serve several buyers if the licence allows it. A support archive that helps one lab train a service agent can help another lab test theirs. You set those terms.
What Makes a Record Valuable
Volume matters less than most people expect. Completeness matters more.
A buyer will pay more for ten thousand support threads that each show the problem, the response and the resolution than for a million messages with no context. Connected records let a model learn cause and effect.
When we assess a company, we look at five things: years of history, number of records, whether each case links from start to finish, whether the company holds the right to license it, and how many buyers want that kind of data today. Industry is only one factor. A logistics firm's delivery disputes can be worth as much as a software company's tickets.
Privacy Is Part of the Product
Serious buyers will not accept data that exposes people.
Frontier labs carry legal and reputational risk for what goes into their models. They need names, email addresses, account numbers and other identifying details removed or replaced before training. Data that skips this step is worth less, and many buyers won't take it at all.
Replacing names only starts the job. A message can still identify someone through context: a job title, a location, an unusual event. The preparation has to review what people wrote as well as the fields around it. We tell you this up front because you should know what de-identification can and can't promise before you sign anything.
Passwords, payment details, credentials and access to live systems stay out of every dataset.
You Keep Ownership
A licence grants specific rights for a specific purpose. The data stays yours.
Your agreement defines which records go in, which uses buyers may make of them, which buyers you exclude and when they must delete the raw dataset. You can leave out a system, a folder, a customer or an entire workflow.
One limit needs saying clearly. Once a buyer has trained a model, deleting the dataset doesn't remove what the model learned. Read that clause before you sign, and choose what you license with it in mind.
Why Now
Frontier labs are competing for a limited supply of real business records, and few companies have started licensing theirs.
That balance won't last. As more companies enter the market and synthetic methods improve, the premium for any single dataset will likely fall. Nobody can predict the timing, but companies that license early get to set terms while buyers still compete for supply.
Starting costs you very little. The first step is a conversation and a few screenshots of record counts, with no data leaving your systems.
What We Believe
Most companies hold a valuable asset and don't know it. The records your team creates every day show how real work gets done, and frontier AI labs need exactly that to build better models.
We think companies should earn from that data on their own terms, with privacy handled carefully and the limits stated before anyone signs. We start with counts and screenshots instead of raw uploads, so you know what your data could be worth before you share any of it.
If your company has run for a few years on tools like Slack, Gmail, Salesforce or Zendesk, find out what your records could earn.