Yes. In most cases a US company can sell, or more precisely license, its own operational records to AI companies, and no federal law bans it, Emory law professor Ifeoma Ajunwa told MIT Technology Review on Oct 13, 2026. As of Oct 23, 2026, micro1, Handshake AI and Mode all publish prices starting at about $100,000 per license, and most programs that publish a size floor set it at 20 to 30 employees. What you can include depends on your contracts, privacy law and what the records contain.
Each section below links to a longer guide. If you would rather start with a number, get a confidential estimate.
What does selling company data to AI actually mean?
You give an AI lab, or a data company that supplies AI labs, a copy of records your business already has, such as email and chat, support tickets, documents, CRM records and code. The buyer removes personal details and uses the copy to train and test AI systems.
Owners usually say “sell.” In practice it is a license. You keep the original and the ownership, and the contract sets what the buyer may do with its copy, for how long, and whether you can license the same records to anyone else. That difference matters later, when you compare offers. Our guide to what’s in a data licensing agreement goes through the clauses.
Why are AI companies paying for company records?
To teach AI agents how real work gets done. micro1 says it uses de-identified company data to build reinforcement-learning environments that mirror real business operations (micro1, Oct 9, 2026), and it committed $1B over 12 months to licensing this kind of data in the same announcement. MIT Technology Review reported that Surge AI buys company data to build “reinforcement-learning gyms” where AI agents practice tasks such as drafting appellate briefs (Oct 13, 2026).
Company records show the email, ticket and revision behind a decision, which public web text rarely does. Our guide on what AI labs want from company records explains this in more detail.
What kind of company data sells?
Records of work that connect across several tools and go back years. As of Oct 23, 2026, micro1 asks for “SOPs, knowledge bases, internal documentation, CRM data, project histories, and QA processes” (micro1). Handshake AI names records from tools such as Gmail, Slack, Dropbox, AutoCAD and NetSuite (Handshake). Collinear pays more for “pull requests with the tickets behind them,” “human-written code” and “the same work visible across several tools” (Collinear).
What does not sell, or drops out of a deal:
- Data you cannot show you own. Collinear lists “anything you do not hold the rights to.”
- Sensitive employee and health records. SimpleClosure does not sell them (SimpleClosure, Sep 3, 2026).
- AI-generated code and “loose documents with no company around them” (Collinear).
- Other companies’ confidential information held under NDAs or client contracts.
Some sectors have their own rules. A law firm, for example, starts from Rule 1.6; see can a law firm license its data and the pages for law firms, accounting firms and software companies.
Is your company big enough?
Six of the eight programs below set a floor of 20 to 30 employees. These are the terms on the buyers’ own pages, read on Oct 23, 2026:
| Buyer | Size floor | History |
|---|---|---|
| Handshake AI | 20+ full-time employees | 3+ years in operation |
| Nyne AI | 20+ full-time US W-2 employees | 3+ years of workflows |
| Pewter | 20+ full-time staff, now or at peak | 3+ years of records on hand |
| micro1 | 30+ employees | Not published |
| Mode | 30+ for most; 10+ accounting, finance, medical; 6+ law | Several years (strongest fit) |
| Replay | Lowest band starts at 20 employees | Not published |
| Collinear | No floor; prices rise with 15+ people at peak | No floor; several years raises the price |
| SuperAnnotate | None: “Any size, any industry, anywhere” | Records built up over time |
Many programs prefer US companies. micro1 says it prioritizes “US companies, followed by other Western markets,” and Mode says US-based teams are the strongest fit. Our comparison of AI data partnership programs covers every published program.
How much is company data worth?
As of Oct 23, 2026, the published bands run from about $10,000 to $5M per license. micro1 publishes tiers of $100K+, $500K+ and $1M+. Handshake AI shows “$100K to $4M+ payouts,” and Mode says company data “could be worth $100K–$5M,” with its calculator labelled indicative. Replay ties its band to headcount: $10K–$100K for 20 to 50 employees, $100K–$1M for 50 to 250, and $1M+ above that (Replay).
Disclosed deals are fewer. A shut-down crypto startup netted about $25,000 for its codebase (MIT Technology Review, Oct 13, 2026). Google agreed to pay $10 million for Spirit Airlines’ internal data in a bankruptcy auction, and as of Oct 15, 2026 the court hearing on it had moved to Nov 6 (The State of AI, Oct 15, 2026). Every figure, with its source, is in our market index, and our guide to what company data is worth explains what moves a price.
Is it legal to sell company data to AI?
Generally yes, with limits set by contracts, privacy law, labor law and privilege. Frankfurt Kurnit Klein & Selz (FKKS) set them out in its Oct 1, 2026 guide:
- Contracts. NDAs, customer contracts and software vendors’ terms. FKKS notes that Slack’s API terms “prohibit third-party application providers from using API data to train a large language model.”
- Privacy law. Under the CCPA, FKKS writes, “a transfer of workplace data for money or other valuable consideration is a sale of personal information,” which brings notice and opt-out duties unless the data meets the law’s de-identification standard.
- Labor law. Workplace messages “may be protected concerted activity under Section 7 of the National Labor Relations Act.”
- Privilege and trade secrets. FKKS warns “a sale may waive privilege over legal communications and jeopardize trade secret protection.”
Employees can object. At Spirit, the flight attendants’ union said “customer data got a privacy screen while worker records did not,” and the court’s privacy ombudsman recommended approval on Oct 5, 2026 only after passenger and booking data were cut (The State of AI, Oct 15, 2026). Our guide to whether it is legal has a ten-step checklist. Marin Data does not give legal advice; talk to your own counsel before you sign.
How does it work, step by step?
A typical order, put together from the buyer pages cited above as of Oct 23, 2026:
- Decide which systems and date ranges you would include, and leave out HR, health and legal files.
- Check that you own each set of records and that vendor and customer contracts allow it.
- Approach buyers, directly or through an intermediary. Buyers ask about headcount, years and tools first, and nobody needs raw records to make a first offer.
- Sign an NDA, then let the buyer review a sample and make an offer.
- Negotiate the license: scope, field of use, exclusivity, deletion, payment.
- The buyer, or a firm it names, de-identifies the records. Nyne removes identifying details before the data is used; Collinear replaces names with “consistent pseudonyms.”
- Deliver and get paid. Nyne pays once the dataset is “delivered and accepted”; Mode expects “about three months from the first conversation through payment.”
Payment can be one sum, an upfront fee plus a revenue share, or a payment on each relicense. SuperAnnotate pays “on delivery, and on each relicense.” Our guide to exclusive and non-exclusive licenses compares these models.
What are the risks?
Once your records have been used for training, you cannot take them back. FKKS writes: “Once data is used to train a model, it cannot practically be pulled back out.” A model trained on your records could help a competitor, so field-of-use limits and deletion of the raw copy belong in the contract.
De-identification has limits too. FKKS notes that linked records can still reveal information about individuals or small groups, and the Spirit ombudsman’s report “did not assess risks to employees or third parties” (The State of AI, Oct 15, 2026). Ask each buyer where and how it strips personal data, and see a sample before you sign.
Some offers come with pressure to decide fast. Our guide on unsolicited offers for your data lists the warning signs, such as 48-hour deadlines and requests for data before an NDA.
Should you go direct or use someone?
You can apply to a buyer program yourself. Mode, for one, says it “charges no broker fee,” and you get one buyer’s offer on that buyer’s terms. Marketplaces take a share of each sale, and intermediaries and advisors are paid in different ways that shape their advice. Our guide to how seller-side data advisors work sets out the four routes and their fees.
Marin Data takes a company to several buyers and negotiates for it. It costs the company nothing up front, and nothing at all if no deal closes; Marin Data is paid by the AI lab.
When is the answer no?
The answer is usually no if most of what you hold belongs to someone else, such as client files you cannot get consent to license, or if the records are mainly personal or health data. A short history kept in one tool will struggle to find a buyer, and a union contract or customer agreement can rule a deal out. A company that is closing has a separate set of options, covered in company shutting down? what to do with the data.
If none of those apply, you likely hold something buyers are paying for. Get a confidential estimate. It takes two minutes and asks for no documents.
Sources
- MIT Technology Review, AI labs are buying data from dead startups, Oct 13, 2026, read Oct 23, 2026
- Frankfurt Kurnit Klein & Selz, selling company data for AI training, Oct 1, 2026, read Oct 23, 2026
- The State of AI, AI labs are buying bankrupt companies’ data, Oct 15, 2026, read Oct 23, 2026
- micro1, $1B enterprise data commitment, Oct 9, 2026
- micro1, data partnerships, Handshake AI, license page, Mode, company data, Nyne AI, Collinear, SuperAnnotate, Replay, Pewter, all observed Oct 23, 2026
- SimpleClosure release via Webull, Sep 3, 2026