Data licensing 101: how a company licenses its records to AI buyers

A data license is a contract that lets a buyer use a defined copy of your records, for a defined purpose, for a defined period. It is not a sale. You keep your records, you keep ownership, and you keep the right to license them again unless you agree to exclusivity. The buyer gets a copy and a set of permissions.

The closest analogy for most owners is licensing a photograph or a piece of software. The photographer still owns the image; the magazine gets to print it once. In a data license, the AI lab gets to train on a de-identified copy of, say, six years of support tickets. When the term ends, the supplied copy is deleted. The model that was trained still exists, which is why the terms of use matter so much.

What it is not: it is not selling your company, not giving anyone access to your live systems, and not handing over customer information. Every serious buyer contracts to de-identify records before training, and a properly structured deal moves a prepared copy once, directly from you to the buyer, after signing.

Four kinds of buyer are active right now. Frontier labs building general-purpose models want records that show how real organizations work, because the public web shows what companies say rather than what they do. Data labs assemble and curate training sets on behalf of labs, and often pay faster because they have a standing brief. Model builders working on a specific product, such as a support agent or a coding assistant, want records from the exact workflow their product performs. Enterprise AI teams at large companies buy to fine-tune internal models and tend to want narrow, high-quality sets.

They pay because the records are scarce. A model can read every public help center on the internet and still never see a ticket where an agent diagnosed a problem over three replies and closed it with a fix. That sequence, with its outcome, only exists inside companies that have been doing the work for years. Published buyer programs now quote ranges that run from the low six figures into the millions for a single company's operational records, and no serious buyer quotes a number before reviewing a manifest.

The records that carry value share three properties: they show a process in motion, they end in an outcome, and they exist in volume. In practice that means support tickets, call transcripts, team chat, email threads, CRM histories, engineering records, procedures, and operational systems.

What does not get licensed: anything you do not have the rights to, anything privileged, anything regulated in a way that blocks it, and anything your counsel excludes. Personal information is removed or replaced before training, not licensed. The data rights checklist walks through the first three; the de-identification guide covers the last.

A deal has six stages, and you approve each one.

1. Qualify. A few questions about headcount, years of records, and systems. Current buyer thresholds are 20 or more desk staff, 3 or more years of records, and data in 3 or more systems.
2. Manifest. A one-page inventory: which systems, which record types, how many years, roughly how much. No records leave your company. This is the document buyers price from.
3. Price. The manifest is benchmarked against what vetted buyers are currently paying for comparable sets. You see the indicative range before any buyer sees your name.
4. First offer. An indicative, buyer-benchmarked range comes back within 24 to 48 hours of a complete manifest. It is a range, not a quote; binding terms follow sample review.
5. Compete. Interested buyers review a small de-identified sample you approve, run diligence, and put terms on the table. Competing offers are the only reliable way to find out what records are worth.
6. Close. You and your counsel pick the terms. You sign directly with the buyer. The prepared copy moves from your systems to theirs. You are paid under the contract.

Elapsed time from manifest to signature is typically a few weeks, driven by buyer diligence and your own legal review. The first offer is fast because the pricing work is done before the manifest arrives.

Price gets the attention, but six other terms decide what a license is really worth.

• Scope. Which systems, which record types, which date range. Everything else is excluded by default.
• Field of use. What the buyer may do with the copy. Model training is the usual grant. Building a product that competes with you, or reselling the records, should be excluded explicitly.
• Exclusivity. Non-exclusive by default. If a buyer wants exclusivity, it should be priced separately and limited in time and field.
• Term and deletion. How long the buyer may hold the supplied copy and when it must be deleted, with written confirmation.
• De-identification. The standard, who performs it, who verifies it, and what happens if the standard is missed.
• Payment. How much, when, and in what installments. Net 30 to net 60 from signing is common in published buyer programs.

The value guide explains how each term moves the number.

You keep your records, your systems, and your ownership. You keep the right to license to other buyers unless you grant exclusivity. You keep the right to walk away at any point before signing. You keep control of scope: a buyer who wants more than the manifest offers has to negotiate for it.

What you give up is a copy, for a period, for a purpose. A model trained on that copy keeps what it learned after the copy is deleted, so the field-of-use and competition terms are the ones to read twice.

You do not need to export anything, build a dataset, or hire anyone to find out whether your records are worth licensing. The starting point is a manifest, and the starting point for a manifest is seven questions. If your company qualifies, the manifest takes a day and the first indicative range follows within 24 to 48 hours.

Before any of that, run the red-flags checklist against anyone who approaches you, including us. A real opportunity should survive it.