The short version
Frontier AI companies have largely used up the free, public internet. To keep improving, their models need examples of how real work actually gets done: support tickets resolved, projects planned, deals closed, invoices reconciled. That knowledge lives inside ordinary businesses, and the AI labs are willing to pay for licensed, de-identified access to it.
Why AI labs need business data now
Early language models learned from public web pages, books and code. That supply is finite, and much of it has been used. Public text also says little about the day-to-day mechanics of business: how a support agent calms an upset customer, how an engineering team breaks a project into tickets, how a finance team closes the month.
Labs now want AI that can do useful work inside companies. To teach it, they need realistic workplace examples. Synthetic data helps, but models trained mostly on machine-made data tend to lose quality over time. Real, human-generated records remain the gold standard.
Why they pay instead of just taking it
AI companies have faced lawsuits and public criticism over data scraped without permission. Licensing is the clean answer: a written agreement, a clear scope, documented rights and payment. Major labs have already signed publicly reported licensing deals with publishers, media companies and online communities. Business records are the next step in the same direction.
A license protects the buyer legally and gives the seller control. That is why the market is moving toward paid, consented data rather than away from it.
Is this a tactic to put businesses out of business?
This is the most common worry, and it is a fair question. The honest answer is no, for a few practical reasons.
Patterns, not your playbook
Labs want general patterns across thousands of companies — how support conversations flow, how tasks are written — not one company's customers, prices or secrets. Data is de-identified before any buyer sees it.
You decide the scope
Your client chooses which systems, which date ranges and what is excluded. Nothing is shared at the inventory stage, and there is no obligation if they don't like the offer.
Contracts limit use
The license agreement sets the permitted use, the rights granted and the timelines. It is a defined deal, not open-ended access.
AI is coming either way
Workplace AI will be built whether or not any one company participates. Licensing lets a business be paid for its contribution instead of watching others capture that value.
Better tools for the seller too
Models trained on realistic business work become better assistants for the same kinds of companies that contributed — including your clients.
Money for dormant records
Old tickets, docs and project history usually sit unused in storage. Licensing turns that cost into cash that can fund hiring, tools or growth.
How a Datumi deal works
- Inventory first, no data. The company lists its systems, rough volumes and date ranges.
- Indicative offer. A price range, usually within 7–14 days, before anything sensitive is shared.
- Agreement. If the number works, a standard license locks in dataset, payout, rights and timelines.
- De-identification and delivery. Data is de-identified before any buyer sees it, then reviewed by the buyer.
- Payment once the buyer accepts.