Guides · Published
Is it legal to sell company data for AI training?
Usually yes, for the data you own and have the right to share. But holding data isn’t the same as being allowed to license it: customer contracts, NDAs, privacy law, employee agreements, open source licenses and your software tools’ terms can each take some of it off the table. Here’s what to check, where removing names falls short, and what the license should say. This is not legal advice, so have a lawyer review your deal.
Owning data vs. having the right to license it
Most companies assume that if data sits in their systems, it’s theirs to sell. A law firm’s review of these deals found that most companies can’t sell everything a buyer asks for, and that the work to get ready is usually more than the offer suggests (Frankfurt Kurnit).
Think of each record as having layers. Your company may own the system and the file. But the content inside it can belong to, or be promised to, someone else: a customer, a partner, an employee, a software vendor or an open source author. A license only works for the layers you control.
Buyers care about this as much as you do. In the case over Anthropic’s training data, the judge found that training on books Anthropic had bought was fair use, but downloading and keeping pirated copies was not; Anthropic settled for $1.5 billion (NBC News). The lesson labs took is that how data was obtained is the legal risk, so expect them to ask you to prove your rights in writing.
Check 1: customer contracts and NDAs
This is where most data gets carved out. Email and chat hold customers’ and partners’ confidential information, often covered by NDAs and commercial agreements (Frankfurt Kurnit). Removing names doesn’t fix a contract that forbids sharing the information in the first place.
Pull your standard contracts and any negotiated ones, then look for:
- Confidentiality clauses that define what counts as the customer’s confidential information and who you may share it with.
- Data use limits that say you may only use their data to provide the service.
- No-training clauses. Customers increasingly negotiate terms that ban training on their data, require opt-in first or exclude confidential material (Morgan Lewis).
- Ownership terms where deliverables, tickets or files belong to the client.
- Deletion duties that required you to destroy data when the contract ended.
Where a contract forbids it, leave that customer’s data out. It’s usually a small slice, and excluding it is far cheaper than a breach claim.
Check 2: privacy law
Support tickets, CRM notes, chat and email are full of personal data about customers and staff. Which laws apply depends on where those people live, not just where you’re based, so this is the area where a privacy lawyer earns their fee. At a high level:
- GDPR (EU). Truly anonymous data falls outside it. Pseudonymized data, where a person could still be identified using extra information, is still treated as personal data (GDPR Recital 26).
- CCPA (California). For data to count as deidentified, a business must, among other steps, publicly commit not to re-identify it and contractually bind recipients to the same limits (National Law Review). That second point belongs in your license.
- Your own privacy promises (US). The FTC has warned that adopting more permissive practices, like sharing data with third parties or using it for AI training, and disclosing that only through a quiet, retroactive change to a privacy policy may be unfair or deceptive (FTC).
The practical answer is the same everywhere: remove identities before any buyer sees a sample, and leave out data that can’t be cleaned reliably.
Check 3: employee data, unions and works councils
Your team’s Slack threads and email are the most valuable records you have, and the most personal. The clearest example so far is Spirit Airlines. Google won a bankruptcy auction for Spirit’s internal data with a $10 million bid (Frankfurt Kurnit). The Association of Flight Attendants objected, worried about disciplinary records, payroll history and internal communications, and asked the court to exclude all flight attendant information unless the deal changed (Bloomberg Law).
The union’s point was that the deal’s privacy terms protected passengers better than staff. Spirit planned to de-identify the records, but the agreement kept links between datasets, which the union argued could let information about individuals or small groups be pieced back together (View from the Wing). The approval hearing was postponed (ch-aviation), and in late September the union said Spirit’s revised proposal still hadn’t fixed the problem (Bloomberg Law).
In early October, the court-appointed consumer privacy ombudsman reported that Spirit and Google had taken adequate steps to protect customers’ privacy, with added measures to remove or de-identify their personal information (Bloomberg Law). The sale leaves out Spirit’s 97.5 million passenger profiles and its loyalty-program records (Bloomberg Law). The ombudsman’s review covered consumers, not staff, and the judge still has to approve the sale. For your own deal, the lesson is that scope is negotiable: cutting whole datasets is often how a sale gets past privacy objections.
Privacy advocates are watching what the case means for other companies’ employee records. A lawyer at the Electronic Privacy Information Center told Bloomberg Law the unions’ concerns were valid, adding:
“[I]f this goes through, then so many other companies stand to profit from selling their employee data from decades ago where people didn’t even conceptualize what an AI model could do.”
Before including employee data, check:
- What your employee handbook, IT policy and employment contracts say about how work communications may be used.
- Any collective bargaining agreement, and whether a union or works council has to be consulted. Rules vary by country, so ask local counsel.
- Whether HR, payroll, medical, disciplinary and grievance records can be excluded entirely. In most deals they should be.
- How you’ll explain it to current and former staff if they ask.
Check 4: third-party code and open source
Code your engineers wrote is usually the cleanest asset you have. But repositories also hold licensed data, open source and vendor code, and files from third parties, which can carry their own restrictions or competing ownership claims (Frankfurt Kurnit). Spirit’s sale also drew an objection from a software vendor that says some of the data is its own (Frankfurt Kurnit).
Run a dependency and license scan, list vendored or copied code, check contractor and agency agreements for who owns their work, and exclude client-owned repositories. Strip keys, tokens and passwords from code and its history before anything leaves your hands.
Check 5: how the data leaves your SaaS tools
Your tools’ terms can restrict the export route even when the data is yours. In May 2025, Slack changed its API terms to bar bulk export through its APIs and long-term copies, and added that an app provider “may not use API Data to train a large language model” (Hunton Andrews Kurth). A buyer that wants to plug its own app into your workspace may run straight into that.
Prefer exports you run yourself through admin tools, and have your lawyer read the customer terms for your plan on each system in scope: chat, help desk, CRM, docs and code hosting. Avoid giving a buyer live access to your accounts.
Why removing names isn’t always enough
De-identification is necessary, but it isn’t a pass. Three limits to plan around:
- Context identifies people. A unique role, location and date can point to one person with no name attached. Review whole records, not just name fields.
- Links can rebuild identities. Buyers want records joined across systems, which is what makes the data valuable. The Spirit dispute shows those same links are what privacy objections target.
- It doesn’t cure a contract. If an NDA or customer agreement bans sharing, a redacted copy is still a breach.
The flight attendants’ court objection in the Spirit case shows what is at stake. Their lawyer argued that records with names replaced could still expose individuals and small groups of staff:
“A pseudonymized dataset can still disclose which crew bases generated grievances, how a small subset of flight attendants performed on recurrent training, which employees were subject to investigation …”
Agree the standard in writing, test it on a sample, and exclude anything you can’t clean with confidence.
What the license should say
A license, rather than an outright sale, lets you keep the original and set limits on use. These are the terms to get right, in plain English:
| Term | What it covers | What to look for |
|---|---|---|
| Grant and permitted use | What the buyer may do with the data | AI training named explicitly; anything else (resale, evaluation, RL environments) listed, not implied |
| Scope | Which systems, folders and date ranges are included | A written list, with carve-outs for client-owned, HR, health and legal material |
| Ownership and warranties | What you promise about your rights | Only what’s true: you own it and can sign. Not that the data is error-free |
| De-identification standard | What gets removed before delivery | A defined method, tested on a sample, plus a ban on re-identifying anyone |
| Onward transfer | Whether the buyer can pass it on | You approve each recipient; the same privacy and re-identification limits bind them |
| Raw data deletion | When the buyer destroys the copy you sent | A set number of days, backups included, with a written certificate |
| Trained models | What happens to models already trained | Usually the buyer keeps them, since a model can’t be untrained. Price that in |
| Indemnity and liability | Who pays if someone sues | A cap, mutual where possible. Push back on unlimited seller liability |
Have a lawyer with privacy and IP experience review the scope and the license before you sign, especially for EU staff or customers, regulated industries or anything cross-border. For how deals are found, priced and closed, see our guide on how to sell your company’s data to AI companies.
Where Smelted fits
Smelted licenses the data your company already has to AI labs. Identities are removed, and you approve every buyer and the terms before anything is shared. You still decide what’s in scope, and your own counsel should review the license. To see what your data might be worth first, try the 60-second estimate.
Questions founders ask
Is it legal to sell company data for AI training?
Often, for part of it. You can license data you own and have the right to share. Customer contracts, NDAs, privacy law, employee and union agreements, open source licenses and your software tools’ terms can each rule out some data or some ways of exporting it. Have a lawyer review scope and the license before you sign.
Do I need my employees’ consent to sell our Slack or email data?
It depends on where your staff are, what your policies and employment contracts say, and whether you have a union or works council. Personal details should come out before any buyer sees the data either way. Spirit Airlines’ data sale to Google drew a formal objection from its flight attendants’ union, so plan for how staff will react, not just what the law requires.
Is anonymized data outside GDPR?
Truly anonymous data is, under GDPR Recital 26. But pseudonymized data, where names are swapped for codes and a person could still be identified with extra information, is still personal data. Removing names alone is often pseudonymization, not anonymization.
Can I sell customer data if my privacy policy didn’t mention AI?
Be careful. The FTC has warned that quietly changing a privacy policy or terms of service to allow AI training or sharing with third parties may be unfair or deceptive. Ask a lawyer whether the data can be used at all, and plan to remove customer identities before anything is shared.
Can I include our source code in a data license?
Code your team wrote is usually the easiest asset to license, but vendored libraries, open source components and client-owned code may carry their own terms. Inventory what’s in the repository and exclude anything you don’t own outright, along with keys and secrets.