How to sell data to AI companies

By Cody Houser, Founder · Last updated October 28, 2026

The short answer: AI companies pay established businesses to license their internal records, such as email, chat, documents, CRM, and accounting history, because those records show how real work gets done. To sell yours, inventory what you hold, confirm you have the right to share it, exclude what you shouldn’t, then work with a buyer or broker on an agreement that limits how the data is used.

This guide covers who buys company data, what sells and what doesn’t, what deals have paid publicly, the legal checks, and the process step by step. It’s written for owners and executives of operating businesses as well as companies that are closing.

Do AI companies buy data?

Yes. In 2026 startups that were shutting down began selling their email and Slack archives for as much as hundreds of thousands of dollars, and in August Google won a bankruptcy auction for Spirit Airlines’ internal business data with a $10 million bid. Operating companies are now getting the same calls. AI developers have trained on the public web; what they lack is the private record of skilled work, with the reasoning attached.

Who buys company data

AI labs directly: large developers license data themselves, usually in big, visible deals, such as Google’s winning bid for Spirit’s records.
Data companies that supply labs: firms that source, clean, and package data for AI developers. Mercor, for example, placed the second-highest bid for Spirit’s data, at $7.5 million.
Closure and bankruptcy channels: wind-down services and court-supervised sales for companies that have stopped operating.
Brokers: intermediaries like Hyperdrive Data that assess a company’s records against current buyer requirements and coordinate introductions. See marketplaces vs. brokers for how the routes compare.

Which AI companies buy data

Companies with publicly reported data purchases or licenses include:

Google: won the auction for Spirit Airlines’ internal records and licenses Reddit’s posts
OpenAI: licenses news and magazine content from News Corp and Dotdash Meredith
Microsoft: licenses academic content from Taylor & Francis
Data companies that supply AI labs: Mercor and micro1 both bid for Spirit’s archive, and micro1 has committed $1 billion to buying company operational data
Foundation labs, simulated-workspace builders, and mature tech companies are all buying company archives, according to MIT Technology Review

Every priced deal is listed, with sources, in the AI data licensing deals tracker.

What AI companies have paid: published deals

Most deals are private. These were reported publicly:

Deal	Reported value	What was sold or licensed
Spirit Airlines to Google (2026)	$10 million (Mercor bid $7.5 million)	Years of internal communications and company records, with identifying details removed
Closed startups (2026)	Up to hundreds of thousands of dollars per company	Email, Slack messages, documents, and source code
Reddit to Google (2024)	About $60 million a year	Access to Reddit’s user posts
News Corp to OpenAI (2024)	Reported at more than $250 million over five years	Current and archived news content

Platform deals like Reddit’s and News Corp’s sit far above what an ordinary company receives. For most businesses, the closed-company deals are the fairer comparison, and an operating company with years of complete records is in a stronger position than a startup that ran for a year or two. Buyers that publish their ranges quote about $50,000 to $2 million or more for business data; see how much company data is worth for the full breakdown.

What data sells

Email and internal chat: threads where problems were worked through and decisions made
Documents and files: procedures, proposals, reports, and the drafts between versions
CRM records: deals from first contact to close, with the notes
Accounting and operations systems: how the business actually runs, transaction by transaction
Technical project files: code with its tickets and reviews, and CAD and drawing sets with their revisions

What makes records valuable is depth and context: many years, many people, complete systems, and the reasoning visible alongside the result. Specialist work such as legal, accounting, engineering, and finance is in particular demand.

Not sure where your company stands? Run through the data value checklist.

What rarely sells

Public data: published content and anything already on the open web
Personal data on its own: customer lists and contact databases. Buyers want records of work, and personal information is what they strip out.
Thin history: a few months of records or a very small team
Data you don’t control: client material held under contract, or records someone else has a claim on

The legal checks before you sell

A company that holds data doesn’t automatically have the right to sell it. Frankfurt Kurnit’s guidance for companies approached by buyers is a useful checklist. The main issues:

Rights in the data: email and chat contain other companies’ confidential information under NDAs; vendor-generated data and licensed content may belong to someone else. In the Spirit case, a software vendor objected that records its platform generated were its own.
Platform terms: how you extract the data matters. Slack’s API terms, for example, prohibit third-party apps from using API data to train large language models.
Privacy promises: companies are bound by the privacy notices and handbooks in effect when the data was collected. The FTC has warned that quietly changing those terms to allow AI training may be unfair or deceptive.
Privacy law: the CCPA covers California employees and business contacts, and transferring their personal information for money is a sale unless the data meets the law’s de-identification standard. The GDPR applies to EU personal data. See employee emails and Slack messages.
Employees: communications about pay and working conditions can be protected under the National Labor Relations Act, and union contracts may restrict a sale.
Privilege and trade secrets: a sale can waive privilege over legal communications and weaken trade secret protection.
Disclosure: California’s AB 2013 requires generative AI developers to disclose high-level information about their training data sources, so a sale may not stay confidential.

How to sell data to AI companies, step by step

1. Inventory your records. List each system, the record types it holds, the date range, and the approximate volume. Mark each as a candidate or excluded. Don’t export anything yet.

System	Record types	Years covered	Include?
Email	Internal and client correspondence	e.g. 2014–2026	Candidate, with exclusions
Chat	Team channels	e.g. 2018–2026	Candidate; exclude HR and private channels
Files	Procedures, proposals, project files	e.g. 2012–2026	Candidate
HR and payroll	Personnel files, pay, benefits	—	Exclude

2. Assemble the right people. Legal, privacy, security, HR, and communications should all be involved. Don’t share samples before this review; a sample is itself a disclosure.

3. Confirm your rights. Review customer, partner, and vendor contracts, the privacy notices in effect when data was collected, and platform terms for the systems involved.

4. Define exclusions. Leave out high-risk material entirely: privileged communications, HR and personnel files, payroll and tax records, health and benefits information, union-related communications, and confidential client material.

5. Choose your route. Go direct, list with a marketplace, or work with a broker. A broker suits most operating businesses, because internal records need assessment before they can be offered.

6. Describe, don’t send. The first conversation should be about your industry, team size, history, and systems, not the records themselves. Buyers price from that description and, later, from controlled samples.

7. Negotiate the agreement. Scope, permitted uses, whether trained models survive the agreement, deletion, warranties, exclusivity, and payment. Use the agreement checklist.

8. Control preparation and transfer. Confirm in the agreement who de-identifies the data (usually the buyer’s team or a third party it brings in) and to what standard, keep buyer tools out of your live systems, use a secure transfer, and document every decision. If the company is ever acquired, the sale will come up in diligence.

Operating company or closing company?

Operating businesses can choose their timing, scope, and buyer, and keep running throughout. The main costs are review time and the goodwill of current employees, customers, and partners, so a careful, documented process matters. Closing businesses have complete records and fewer ongoing relationships, but need to confirm who has authority to act and keep accounts alive until a buyer has evaluated them. See licensing data when your business is closing.

Mistakes to avoid

Sending sample data before legal and privacy review
Relying on redaction alone for sensitive categories that should be excluded
Quietly changing a privacy policy to make a sale possible
Giving a buyer’s tools direct access to company systems
Sharing proprietary records before you’ve accepted an offer and the agreement is in place
Purging old systems or letting accounts lapse before anyone has evaluated them
Assuming you can claw the data back: once a model is trained on it, it can’t practically be removed
Comparing offers on price alone, without scope, exclusivity, and permitted uses

Questions people ask

Can a business sell its data to AI companies?

Yes. AI companies license internal business records such as email, chat, documents, CRM, and accounting history to train and evaluate their systems. A company can license records it has the right to share, within privacy law and its own contracts.

How much do AI companies pay for company data?

There’s no standard price. Buyers’ published ranges run from about $50,000 to $2 million or more, closed startups have sold archives for $10,000 to $100,000, and Google bid $10 million for Spirit Airlines’ internal data. Depth of history, completeness, the kind of work the records show, and your rights in them drive the number.

Do I sell the data or license it?

Most deals are licenses: the buyer gets defined rights to use specified records for AI training, on terms set in the agreement. Some buyers ask for an outright purchase. Read what each proposal actually grants.

Is it legal to sell company data for AI training?

It can be, but most companies can’t sell everything a buyer asks for. Contracts, privacy promises made when the data was collected, privacy laws such as the CCPA and GDPR, and platform terms all limit what can be shared. Have counsel review before any data or samples leave the company.

Does my company need to be closing to sell its data?

No. Closed companies and bankruptcies make the news, but operating businesses can license data too, on their own terms, while they keep running.

What size company do AI buyers want?

Buyers count years of history and headcount. For Hyperdrive Data’s current opportunities, the strongest candidates have 50 or more desk and office staff, three or more years of history, internal communication mainly in English, and five core systems: email, file storage, chat, CRM, and accounting.

Do I have to anonymize the data myself?

No. Buyers don’t want sellers redacting records themselves. Personal information is removed by the buyer’s own team or, most often, an independent third party the buyer brings in, to a named standard. Your part is to flag categories that can’t be shared at all, such as privileged or client-confidential files, so they’re kept out.

Can an employee sell their employer’s data?

No. Only the company that owns the records, acting through someone with authority, can license them. Buyers check rights before anything moves.

Sources

This guide is general information, not legal, tax, or financial advice. Review any proposed agreement with your own advisers.

Related guides

Check whether your company is a fit. Start with a description of your records; no files needed.