How to sell your company’s data to AI companies

Why AI developers want company data

The public web is finite. Epoch AI estimates the stock of usable public human text at about 300 trillion tokens, and that models are likely to have used all of it at some point between 2026 and 2032 (Epoch AI).

What the public web lacks is the record of how work is actually done: long conversations, decisions, drafts and the reasons behind them. OpenAI states that it seeks “large-scale datasets that reflect human society and that are not already easily accessible online,” and prefers “long-form writing or conversations rather than disconnected snippets” (OpenAI).

MIT Technology Review reported in October 2026 that demand for company archives “emerged suddenly last year, as AI training labs hit the limits of public and synthetic datasets” (MIT Technology Review).

Who buys it

Two kinds of buyer license company data. The first is the developers of artificial intelligence themselves. The second is the firms that supply those developers with training data, some of which now acquire company archives directly. One such supplier describes its work as partnering “with leading companies to license their operational data to the labs” (Fast Company).

Buyers differ in what they want, how they use it and the terms they accept. Matching an archive to the right buyer matters as much as the archive itself.

What kinds of company data sell

Operational data is the record a company creates in the course of running its business. It includes:

• Email and internal messaging
• Documents and shared drives
• Customer support records and call recordings
• CRM, quotation and invoicing records
• Project and task management histories
• Work orders and service records
• Accounting and ERP records
• Engineering drawings and design files

Buyers pay most for archives that show how decisions were made. A buyer quoted by MIT Technology Review looks for records of “disagreements to tricky judgment calls.” Records that connect several systems, such as tickets, design files and the email around them, are worth more than any one system alone, and as one wind-down platform put it, “scarcity is definitely more valuable” (Upstarts Media).

What it is worth

Most transactions are confidential, but a growing number have been disclosed in filings or reported by the press. Those involving company records give a sense of the range.

Value depends on the volume of records, the period they cover, the rights the company holds, how much personal information must be removed and how many useful training tasks a buyer believes the archive will yield. Buyer demand sets the final price. Spirit Airlines attracted a competing $12.5 million offer after its auction closed (Fortune).

Published unit prices exist for media, not for business records. Freepik confirmed two to four cents per image, and unpublished video has sold for $1 to $4 per minute (Reuters, Bloomberg via The Seattle Times). Company archives are priced one at a time.

What must be true before you sell

A company can only license what it has the right to license. Four questions decide whether it can:

1. Do your contracts allow it? Customer agreements, supplier terms and platform terms can restrict how records may be used. Slack’s API terms, for example, bar third-party applications from using data obtained through its API to train large language models (Slack).
2. Is personal information handled lawfully? The GDPR requires a lawful basis for each use of personal data (EUR-Lex). The CCPA gives Californians the right to opt out of the sale or sharing of their personal information (California Attorney General). The UAE’s Federal Decree-Law No. 45 of 2021 generally requires consent unless an exception applies, and does not apply in free zones with their own data protection laws, such as the DIFC (UAE legislation, DIFC). Saudi Arabia’s Personal Data Protection Law has been enforceable since September 2024 (SDAIA).
3. Were people told before the new use? The US Federal Trade Commission has warned that adopting AI training uses through a “surreptitious, retroactive amendment” to terms may be unfair or deceptive (Crowell & Moring).
4. Are employee records protected? Employee data draws the most scrutiny. In the Spirit Airlines sale, the unions secured the exclusion of medical and payroll records and an independent verification agent (Investing.com).

European regulators have also said that models trained on personal data cannot in all cases be considered anonymous (EDPB), which is why personal information is removed before delivery. This guide is general information, not legal advice. Counsel should review any license before it is signed.

Licensed, not sold

In most cases the buyer receives a copy of the agreed data with defined rights of use, and the company keeps its original records. The license sets the permitted use, the term and whether the rights are exclusive. Published deals show what such terms can look like: Wiley’s agreements are non-exclusive and limited in duration, and HarperCollins capped how much of a licensed work a model may reproduce (The New Publishing Standard, Australian Society of Authors).

Outright sales happen too, most often when a company is in bankruptcy and its assets are being sold. Spirit Airlines agreed to sell its business records outright, to be delivered with personal information removed, subject to the approval of the bankruptcy court.

How the process works

1. Discovery. An initial consultation on your company’s systems and a preliminary view of what its data is worth. It works from descriptions, so no data needs to leave your systems.
2. Organization. Establishing which systems, periods and records can be licensed under your existing agreements, and preparing a listing that describes them without personal information.
3. Representation. Introducing your company to the buyers best suited to its data. Your company decides which buyers, if any, receive information about it.
4. License and delivery. The license sets the permitted use, the term and how personal information is removed. The data moves directly from your company to the buyer, or to a specialist de-identification firm.

Riverwell works through the first three stages with the companies it represents and never takes possession of client data. Read how Riverwell works with sellers.

If the company is winding down

A company that is closing may still hold valuable records, provided they exist and the company retains the right to license them. SimpleClosure, which handles company shutdowns, reports that 47% of its transactions now include the sale of code or data (MIT Technology Review).

Timing matters most. Once accounts are closed and systems decommissioned, the records are usually gone. We recommend assessing the data before any system is switched off.

Questions

How much do AI companies pay for company data?

It depends on the volume, the period covered, the rights held and buyer demand. Disclosed examples range from about $10,000 for a small startup’s archive to $10 million for the business records of Spirit Airlines. Most transactions are confidential.

Is it legal to sell company data to AI companies?

It can be, where the company holds the rights under its contracts and personal information is handled lawfully under the laws that apply, such as the GDPR, the CCPA or the UAE’s data protection law. Counsel should review any license before it is signed.

Does the company give up its data?

In most cases, no. The buyer licenses a copy of the agreed data with defined rights of use, and the company keeps its original records.

Does our data have to leave our systems to be valued?

Not for a preliminary view. Riverwell works from descriptions of your systems, the periods they cover and their approximate volumes. Book a data strategy session.

Can a company that is closing sell its data?

Yes, provided the records still exist and the company retains the right to license them. Assess the data before any systems are decommissioned.