Is it legal to sell company data to AI companies?

01

Is it legal to sell company data? Usually, if you hold the rights

What "selling company data" means here

Every business leaves records behind as it works. Its support staff log each customer's problem and fix in tickets, its team discusses decisions in email and chat, and its salespeople track each customer in a CRM (customer relationship management software, such as HubSpot or Salesforce). AI labs (companies that build and train AI models, such as OpenAI, Anthropic and Google DeepMind) and data vendors (companies that buy records from businesses, anonymize them and package them as training material for the labs) now pay for those records. They pay $100K to $2M or more per dataset, because the records show step by step how employees handle real jobs, which AI models can't learn from the public internet.

Most of these deals are a license: a contract that lets the buyer use a copy of your records for set purposes, while your company keeps ownership. The legal questions are the same whether you license or sell, and data licensing vs. selling explains how the two structures differ.

Owning the records is not the same as owning everything in them

Your company owns its systems and the work its team recorded in them. But a single support ticket can also hold a customer's name and home address, a supplier's price list shared under a confidentiality agreement, and a technician's personal details. Each of those people and companies keeps rights over their part, and could complain to a regulator or sue if it is shared without permission.

The real question is which parts of your records you can include, in what form, and how they leave your systems. Four checks answer it.

• RightsIs each kind of record your company's to license, or does it belong in part to a customer, an employee, a contractor or another company?
• App termsDo the terms of Slack, HubSpot, Zendesk and your other tools allow the way you plan to export the records?
• Personal dataWhat do privacy laws, your privacy policy and your customer contracts allow, and does anonymization take the records outside them?
• DisclosureHas anything, including a "harmless" sample, left the company before a lawyer reviewed it?

02

Who owns what inside your company's records

Five kinds of information sit inside ordinary business records

Company records mix five kinds of information, and each has a different owner or a different set of rules.

Work product

Everything your team created while doing its job, such as written procedures, project plans, the fix recorded on each support ticket and the reasoning staff worked through in a Slack thread. This is the core of what AI labs and data vendors pay for, and it is generally yours.

Customer data

Names, emails, phone numbers, addresses and account history of the people and businesses you serve. You hold it, but privacy laws, your privacy policy and your customer contracts limit what you can do with it.

Third-party confidential information

Material another company shared with you under a non-disclosure agreement (NDA, a contract in which you promise to keep what it shows you secret and to use it only for an agreed purpose), such as a supplier's price list or a client's internal documents. The NDA, not you, decides what can be done with it.

Employee data

Staff names and contact details, HR files, performance reviews, pay and health information. Privacy laws treat it as personal data (information that identifies a person), and HR records are usually excluded from these deals outright.

Contractor work

Code, designs and documents made by freelancers and agencies. Who owns it depends on the contract they signed.

A single record often contains all of these at once, as the diagram below shows with one support ticket from an HVAC company (a business that installs and repairs heating and air conditioning).

Customer data: you hold it, but you don't own it outright

Your company collected its customer data and stores it, but the law treats a customer's personal details as information about that person, usable only in the ways the law and your own promises allow. Business customers can add a second limit, because the contract they signed with you may restrict how you use their data.

Anonymization (replacing names and contact details with consistent stand-ins such as "customer 203", so the record still reads naturally but nobody can be identified) handles most of the privacy side. Anonymization does not cancel a contract, though: if a client's contract says you may use its data only to serve that client, that promise still applies after the names are gone. The privacy section and the contracts section below cover both.

Other companies' confidential material travels in your inbox

Email and shared drives fill up with documents other companies sent you in confidence, such as proposals, pricing and client files. An NDA usually limits that material to the purpose it was shared for, and licensing it to an AI buyer is almost never that purpose. Anonymizing the names doesn't help, because the secret is the content itself, such as the supplier's prices. If that material reached an AI buyer, the other company could sue you for breaching the NDA, so the practical answer is to leave it out.

Employee data is personal data, and HR files stay home

Employees' names and contact details appear in nearly every record, and they are handled the same way as customers' details: replaced with stand-ins such as "employee 14". California's privacy law, the CCPA, covers employees' and job applicants' personal information as well as customers'. HR files, performance reviews, pay records and medical notes describe the employee rather than the work, and companies usually leave them out entirely.

Contractor work belongs to you only if the contract says so

Under U.S. copyright law, work an employee makes as part of their job generally belongs to the employer, but work by an independent contractor does not. Contractor work counts as a "work made for hire" (work the law treats as the hiring company's from the moment it is created) only in a short list of specific categories, and only if both sides signed a written agreement saying so. Otherwise the contractor keeps the copyright unless they signed an assignment, a document that transfers ownership to the company.

This matters most for code, design files and documents made by freelancers and agencies. A freelancer who still owns code your company licensed to an AI buyer could claim you licensed their work without permission. Ask your lawyer to check the contractor and agency agreements behind any large body of work you plan to include.

03

Slack, HubSpot and Zendesk: app terms that restrict AI use of API data

Two sets of terms cover the same data

Every software tool your company uses has two kinds of terms. The customer agreement is the contract between your company and the tool, and it usually says your company owns the content it puts in. The developer or API terms are the rules for software that connects to the tool through its API (application programming interface, the channel other programs use to read and write the tool's data).

Since 2025, several large tools have added clauses to their API terms that stop connected apps from using the data for AI training. Those clauses bind the app that pulls the data, so the same records can be fine to license through one export route and forbidden through another.

What each tool's terms say

These clauses are paraphrased from the terms in force when this guide was written. Terms change often, so your lawyer should read the current version and the version in your own contract.

The native admin export is the usual route

A native admin export is the export a workspace owner runs from the tool's own admin settings, producing a file of your company's content. It is your company taking a copy of its own data under its customer agreement, rather than a third-party app pulling it through the API. Data vendors generally ask for exports this way for exactly that reason.

Native exports have limits worth knowing before you start. Slack's export tool, for example, covers public channels on every plan, but private channels and direct messages only on the Business+ and Enterprise plans, after the owner applies for access.

Check the rest of your stack the same way

Slack, HubSpot and Zendesk are the best-known cases, not the only ones. For each system you plan to include, such as email, Google Drive, Jira or your CRM, list the plan you're on and how you would export from it, so your lawyer can check each route against that tool's terms.

04

Is selling customer data legal under GDPR, CCPA and state privacy laws?

Privacy laws regulate information about identifiable people

Personal data is any information that relates to a person who can be identified, directly or indirectly. A name and email are obvious examples, but so is a long email thread that never names someone yet makes clear who they are.

Anonymous data falls outside these laws. Europe's GDPR says so directly in Recital 26: data protection principles do not apply to information that no longer relates to an identifiable person. That is why AI labs and data vendors anonymize records before anyone uses them.

Pseudonymized is not the same as anonymous

Pseudonymization is the step most data vendors perform: each name is swapped for a consistent stand-in, and somewhere a key exists that links each stand-in back to the real name. Recital 26 says pseudonymized data that can be linked back with extra information is still personal data. The UK regulator, the ICO, explains that pseudonymized data stays personal data for whoever holds the key, though it may be anonymous in the hands of someone who doesn't.

In practice, the key should be destroyed or kept away from the buyer, and the contract should bar re-identification (working out who a stand-in really is). How company data is anonymized shows the technique step by step.

The transfer to the data vendor is the step privacy law cares about most

In most deals your company does not anonymize anything itself. The data vendor receives the raw records over an encrypted connection, anonymizes them with its own tools and then deletes the raw copy. Until that point the records are personal data, as the diagram below shows.

For a lawyer, that means the handover is a disclosure of personal data and needs the same paperwork as any other. The usual pieces are a legal basis for sharing (a lawful reason the privacy law recognizes, such as the customer's consent), a data processing agreement that limits what the vendor may do with the raw records, a written retention period, and, for records about people in the EU or UK, terms that cover moving data to another country. micro1, the data vendor we typically refer companies to, says on its data partnerships page that raw data is retained only for an agreed period and deleted on an agreed schedule; its contracts allow up to 180 days, usually much less.

GDPR and UK GDPR: Europe's privacy laws

The GDPR (General Data Protection Regulation) covers personal data about people in the European Union, and the UK GDPR is the near-identical British version. Under both, a company needs a lawful reason for each way it uses personal data. A new use must either fit the purpose the data was originally collected for or have its own lawful reason, such as fresh consent. Sharing customer records with an AI data vendor is a new use, so a company with customers in Europe should expect its lawyer to spend the most time here.

The GDPR also sets a high bar for anonymity: no one should be able to identify a person using means "reasonably likely to be used". A long email thread about one client rarely clears that bar even with the names removed, because the details of the job can still point to who the client is. Threads like that are often safer left out.

CCPA and CPRA: California's privacy law

The CCPA (California Consumer Privacy Act), as expanded by the CPRA (California Privacy Rights Act), applies to for-profit businesses that do business in California and meet one of three tests. According to the California Attorney General, those are gross annual revenue above $25 million, buying, selling or sharing the personal information of 100,000 or more California residents or households, or earning half or more of revenue from selling personal information. Businesses that "sell" personal information must offer a "Do Not Sell or Share My Personal Information" link and honor opt-outs.

Deidentified information is outside the CCPA, but the law's definition has three conditions, set out in Civil Code section 1798.140. The business must take reasonable measures so the data can't be linked to a person or household. It must publicly commit, usually in its privacy policy, to keep the data deidentified and not try to re-identify it. And it must contractually require anyone who receives the data to follow the same rules.

Other U.S. state privacy laws

A growing number of U.S. states have passed comprehensive privacy laws modeled loosely on California's, including Virginia, Colorado, Connecticut and Texas. The IAPP's state privacy legislation tracker keeps the current list. Most of these laws define selling broadly, give residents a right to opt out of sales, and exempt deidentified data that meets conditions similar to California's.

Many smaller companies fall below these laws' size thresholds, but their privacy policies and contracts still apply.

Do you need to register as a data broker?

A data broker, in U.S. state law, is a business that sells personal information about people it has no direct relationship with. Four states require data brokers to register: California, Vermont, Texas and Oregon, as Davis+Gilbert summarizes. California's registry, run by the California Privacy Protection Agency, charges a $6,000 annual fee for 2026 and requires registered brokers to process deletion requests from a statewide mechanism.

An operating company licensing records about its own customers usually isn't a data broker under the California, Vermont and Oregon definitions, because it has a direct relationship with those people, and anonymized records aren't personal information at all. Texas is different: its law turns on whether a company's main revenue comes from personal data rather than on the direct-relationship test. If a deal would include personal information that is not anonymized, ask your lawyer to check the registration question specifically.

05

Your privacy policy and customer contracts can limit a deal even when the law allows it

Your privacy policy is a promise regulators enforce

In the U.S., the Federal Trade Commission (FTC) treats a company's privacy policy as a commitment to customers. In a February 2024 statement, the FTC warned that quietly and retroactively changing a privacy policy to allow new uses of data, such as AI training, may be unfair or deceptive. A company that collected data under one set of promises cannot rewrite them afterwards to allow AI training without risking FTC action.

Three passages in your privacy policy matter most: what it says data is used for, whether it allows sharing with third parties, and whether it mentions de-identified data. Many policies already allow any use of de-identified data, which makes a deal easier. A policy that promises never to share customer information needs your lawyer's attention before any customer records are included.

Customer contracts and data processing agreements

Business-to-business companies usually sign contracts with their customers, and those often include a data processing agreement (DPA), the contract that sets out how a service provider may handle its customer's personal data. A DPA commonly says the provider may process data only on the customer's instructions and only to deliver the service. If you are a software company, agency or outsourcer, the records you hold about your clients' customers may be covered by those terms.

Some contracts allow the provider to use aggregated or de-identified data; others forbid any secondary use. Your lawyer will sort your customers into groups by contract type, and the records of customers whose contracts forbid secondary use are usually excluded.

NDAs follow the material, wherever it sits

A non-disclosure agreement protects the content another company shared, and it applies wherever that content ended up: in an email attachment, a Slack upload or a shared drive. NDA material is the clearest case for exclusion, because anonymizing names doesn't remove the confidential substance. A list of your active NDAs, showing which teams and folders received each company's material, lets the lawyer and the data vendor find and remove that material quickly.

06

Health data, privileged legal advice and trade secrets: what to leave out

Health information and HIPAA

HIPAA, the U.S. Health Insurance Portability and Accountability Act, protects health information held by healthcare providers, health plans and their service providers. HIPAA's Privacy Rule stops applying once data is de-identified by one of two methods the U.S. Department of Health and Human Services describes: Safe Harbor, which removes 18 specified kinds of identifiers, or Expert Determination, where a qualified expert documents that the risk of re-identification is very small.

Both methods take specialist work, and a mistake leaves the records covered by HIPAA, so data vendors mostly avoid medical records rather than clean them.

Health details also turn up in ordinary businesses, such as an employee's sick-leave email. The anonymization rules your company agrees with the vendor should say such details are deleted outright, not replaced with stand-ins.

Legal privilege can be lost by sharing it

Attorney-client privilege protects confidential communications between a company and its lawyers from being used against it in court. Sharing privileged material with an outside party can waive that protection, which means the other side in a future lawsuit could demand to see it. Emails with outside counsel, legal memos and litigation folders therefore stay out of any dataset, and your lawyer will help identify where they live.

Trade secrets: yours and other people's

A trade secret is business information that is valuable because it is kept secret, such as a formula, a pricing model or a proprietary process. Licensing your own trade secret is legally possible, but handing it to an AI buyer puts its secrecy at risk, and secrecy is what protects it in law. Other companies' trade secrets in your files are covered by their NDAs, as described above.

What is shared can't be taken back

A deletion clause (the part of the contract requiring the buyer to delete your data at the end) removes the buyer's copy of the dataset, not what an AI model has already learned from it. Regulators already treat what a model learned as something that can be ordered destroyed. In 2022 the FTC required WW International (formerly Weight Watchers) and its Kurbo app to destroy algorithms derived from data collected without proper consent, and pay a $1.5 million penalty. Sellers should therefore settle what to exclude before any transfer, because there is no reliable way to pull data back out of a trained model.

07

Can we include this data? A decision flow for each kind of record

Ask the questions in order, from hardest stop to easiest fix

The flow below starts with the questions that end in exclusion, because no amount of anonymization fixes them. It ends with the questions that have a routine fix. Run it once for each system or folder, such as the sales team's shared drive, the #support channel or the HR mailbox, not once for each record.

What the flow looks like for a typical company

For most operating companies, the bulk of their records land in "include after anonymization": support tickets, internal chat, project tools and CRM activity. A smaller set lands in "leave it out": HR mailboxes, legal folders, NDA deal rooms and anything medical. The middle bands, "lawyer decides" and "switch the export", are where a short legal review earns its fee.

Every exclusion removes records the buyer would pay for, so precise exclusions protect the price. Leaving out one HR mailbox costs little, while leaving out all email can remove much of what makes a dataset valuable. What your data is worth explains which records drive the price.

08

Samples are disclosures, and exploding offers are pressure

Why a small sample carries the full risk

Buyers sometimes ask for "a few example tickets" or "a week of Slack" to judge quality. A sample feels harmless because it's small, but it contains the same customer details, employee names and confidential material as the full dataset, and it leaves before any contract protects it. If a sample includes a document shared under an NDA or a customer's personal data, the company has already breached the NDA or privacy law, before any contract limits what the buyer may do with it.

A data manifest replaces the sample

A data manifest is a description of your data, not the data itself: which systems you use, how many years of history each holds, rough volumes, file types and the kinds of records involved. micro1 uses a manifest to price a deal, and its application and intro call require no files. Records move only after a signed agreement, through an encrypted transfer.

Exploding offers are a tactic, not a deadline

An exploding offer is a price that expires quickly, often within 48 hours. The pressure works because the sums are large and buyers say prices may fall. A licensing agreement signed in a hurry still binds the company for its full term, often years, so ask for the time your lawyer needs. A buyer who won't wait a few days for a legal review is unlikely to be more flexible later, when you ask about deletion or exclusions.

09

What a data licensing lawyer reviews before you share anything

What to ask the lawyer to cover

The contract terms themselves (exclusivity, term, permitted use, payment structure) are explained in data licensing vs. selling.

The review runs in a fixed order

A legal review of a data deal follows the order the data will travel. Each step depends on the one before it. A contract signed before the exclusions are settled, for example, may promise the buyer records the company can't legally share.

How to find a data licensing lawyer

You don't need a specialist firm. Look for a lawyer who regularly reviews technology contracts and privacy compliance, often listed as "technology transactions", "privacy and data protection" or "commercial contracts". If your company has outside counsel, start there; they already know your contracts.

Most of a lawyer's fee goes to reading your contracts, so gathering them in advance keeps the bill down. Bring the checklist at the end of this guide to the first meeting and ask for a fixed fee.

10

Only the company can license company data, never an employee

Employees hold access, not rights

An employee who can open the company's Slack or drive has access for doing their job, not the right to share it. Employment agreements and confidentiality policies typically forbid taking company files, and an employee who uploads them to a third party can be sued by the employer for breaching those agreements and trade secret law. The upload can also break privacy law, because the files hold customers' personal details. The same is true of contractors and former employees who kept copies.

The Mercor lawsuits show where the line is

Mercor is a large AI training-data company that hires contractors to produce data for AI labs. In April 2026, contractors filed at least seven class-action lawsuits against it after a data breach, as The Wall Street Journal reported.

The plaintiffs also allege Mercor trained client models on materials potentially owned by other companies. One complaint says workers were encouraged to use real company data in tasks if it was slightly altered or anonymized. Mercor says it strongly disputes the claims.

Whatever the outcome, the cases show why company data should come only from the company itself. The company decides to license it, runs its own exports, has a lawyer approve what is included and signs a contract with the buyer.

Closed and bankrupt companies

A company that has shut down can still license its records, but who signs changes. In a wind-down, the remaining officers or a liquidator may act for the company; in bankruptcy, a trustee or the court controls the assets, and proceeds often go to creditors. The same rights, privacy and app-terms questions apply, and the privacy promises made to customers survive the shutdown.

11

Tax questions to take to your CPA

What to ask before the money arrives

• License or sale?A license and an outright sale of an asset can be taxed differently. Ask how this contract's structure will be treated.
• One payment or several?A flat fee, an annual license and payments spread over milestones can fall in different tax years.
• Which entity receives it?If you run more than one company, the one that owns the systems should be the one that signs and gets paid.
• State and cross-border taxesAsk whether any state taxes digital goods or data licenses, and whether a buyer in another country triggers withholding or reporting.
• Effect on a future sale of the businessAn exclusive license can affect what a buyer of your company later pays. Ask your CPA and lawyer together.

12

Legal checklist before you apply to sell company data

Gather these before the first call

• A system inventoryEvery tool, the plan you're on, how many years it covers and how you would export from it.
• Your current and past privacy policiesThe promises in force when the data was collected matter, not only today's version.
• Customer contract templates and DPAsPlus any large customer with custom terms on data use.
• A list of active NDAsWith the teams, folders and channels each one touched.
• Contractor and agency agreementsFor any code, designs or documents you plan to include.
• A draft exclusions listHR, medical, legal and privileged folders, NDA material and any restricted customers.
• Where your customers and staff areCountries and U.S. states, so the lawyer knows which privacy laws apply.
• Who can signThe officer or owner with authority, and any board or partner approval you need.
• A rule that nothing leaves yetNo samples, screenshots or exports until the lawyer signs off.

If you aren't sure your company is the right fit, does my company qualify covers the buyer's criteria, and the estimate tool gives a value range. Where to sell data compares AI licensing with marketplaces and brokers, which carry different legal risks.