How to sell your company's data to AI companies

To sell data to an AI company, you license records your company created and owns, such as SOPs, support tickets, CRM history and project files. The buyer pays a fee and you keep your copy. Who removes personal data, and when, is settled in the agreement. Data brokers work differently: they resell personal information about people.

This guide follows a sale from inventory to payment: what buyers look for, who owns the records, whether to license or sell, who handles personal data, how the deal and payment work, and what to ask before you sign. To see where your records stand first, take the four-question eligibility check. It gives a preliminary answer, asks for no files and is not an offer.

What buyers look for in company records

Buyers look for records that show how your company works: the procedures, the problems your staff solve and the way work gets done. AI labs cannot scrape that, so data companies license it from operating businesses, then prepare and package the records for the labs. The records that fit are the ones a business produces while it runs:

• Standard operating procedures (SOPs) and QA processes
• Knowledge bases and internal wikis
• Support tickets and chat transcripts
• CRM (customer relationship management) history
• Project histories and files in cloud storage
• Internal chat and email
• Accounting exports

The records do not have to come from a software company. In July 2026 The Information reported that data labeling startups are paying HVAC companies for their data.

Records are easiest to license when they are your own work, carry little personal data and sit under no contract that blocks their use. SOPs and project histories often meet those tests. Tickets, chats and email carry the most personal data and need the most care. Be ready to describe each set in one sentence: what it is, which years it covers and which system it came from.

Data you hold for your own clients usually isn’t yours to license: if it sits under a client contract or NDA, leave it out. Third-party content such as licensed reports and vendor manuals stays out too, along with employee files. Ask any buyer which record types it accepts and which it excludes before you spend time preparing one. Our page on what company data AI labs buy goes through each data type.

Who buys company data

Data companies license records from operating businesses and supply AI labs, which train and evaluate models with them. Each buyer publishes its own size requirement, terms and payment, and says how it handles personal data. Our overview of AI training data buyers names them and lists what each one takes.

Selling to AI buyers and selling to data brokers

The Federal Trade Commission’s 2014 report on the industry describes data brokers as companies that obtain and share large amounts of consumer information, typically without consumers’ knowledge. A broker’s product is information about people. An AI data company licenses something different: your company’s own records.

If you searched for how to sell data to data brokers, you may be thinking of customer lists, email lists or contact data. Privacy laws treat that kind of sale differently from licensing records after personal data has been removed. Our guide to whether selling company data is legal covers where the line sits.

How to sell data to AI companies, step by step

Seven steps take a sale from inventory to payment. Steps 1 to 3 are homework your own team can do first. Steps 4 to 7 involve the buyer, and each point is written into the agreement.

1. Take inventory

List every system that holds business records. For each one, write down:

1. What it contains, for example tickets, deals, project files or procedures
2. How many years of history it holds
3. Roughly how much data there is and how it can be exported
4. Whether it holds client data or personal data
5. Who administers it

Most business records live in a handful of tools. This table shows where to look and what to watch for in each:

Example: A 60-person landscaping company has a field-service app with eight years of job records, a shared drive of bid documents and a helpdesk inbox. The inventory shows the job records and bids are mostly the company’s own work. The helpdesk mixes in customer personal data, so it needs the most care.

2. Confirm you own the records

A company can license only records it owns or has the right to license, so settle ownership before you approach a buyer. Ownership has four parts:

• Your own work. Under US copyright law, a work an employee creates within the scope of employment is a work made for hire, and the employer is its author (US Copyright Office, Circular 30). SOPs, wikis and project documents written by your staff are usually yours on that basis.
• Contractor work. A contractor’s work is a work made for hire only if it falls into one of nine listed categories and both sides signed a written agreement saying so. Otherwise a signed transfer of rights gives you ownership, so read each contractor agreement.
• Client and third-party material. Records you hold for a client under a contract or NDA (non-disclosure agreement) may not be yours to license, so leave them out unless counsel confirms you may. Ticket histories and project files can mix your work with client material, so check them collection by collection. Leave out vendor manuals, licensed reports and other third-party content, and check the license on any open-source code and your software vendors’ terms on exported data.
• Authority to sign. Confirm who can sign a license for the company, and whether your governing documents or a lender’s security interest may limit what you can license. Get approval from whoever your governing documents require.

3. Decide between a license and a sale

A license gives the buyer defined rights to use your data, for a stated purpose and period, while you keep ownership and your copy. A sale, also called an assignment, transfers ownership itself. In this market the deal is usually a license: you license to a data company, which packages the records and supplies AI labs.

Two terms decide how close a license comes to a sale:

• Exclusivity. A non-exclusive license lets you license the same data to others. An exclusive license stops that. US law treats an exclusive license as a transfer of copyright ownership (17 U.S.C. § 101), and a transfer is not valid unless it is in writing and signed by the owner (§ 204(a)). The statute excludes a non-exclusive license from that definition, so an exclusive license can work like a sale.
• Duration. A perpetual license has no end date, and a fixed-term license ends or renews on a schedule. For either, ask what the buyer must do with the data at the end and what happens to models already trained on it.

The U.S. Copyright Office states that copyright does not protect facts, ideas, systems or methods of operation, although it may protect the way they are expressed. Much operational data, such as ticket outcomes and CRM stages, reads more like facts than like creative expression, so a data license relies heavily on its contract terms. Our guide to data licensing agreements explains each clause.

4. Agree who handles personal data

Privacy law covers the personal data in your records whatever the contract says, and it defines a sale widely. California’s CCPA counts disclosing personal information to a third party for money or other valuable consideration (California Attorney General), so a paid license of records that still contain personal data can count as a sale. GDPR requires a lawful basis for sharing the personal data of people in the EU. As of November 2026 the United States has no single federal privacy law that covers all personal data, so state laws and sector rules do the work.

Records with personal data removed can fall outside those laws, but only if the removal meets the legal standard. California’s definition of deidentified information requires that it cannot reasonably be used to infer information about, or be linked to, a particular consumer (Civil Code § 1798.140). GDPR treats pseudonymized data, where a name is replaced by a code that can be traced back, as personal data still (Recital 26). Free text is the weak spot: tickets, chats and email carry names, phone numbers and order numbers in places no field list will find.

The agreement decides who removes personal data, and the law applies whichever way you split the work. If you remove it before transfer, less personal data leaves your company and you control the method. If the buyer or its partner removes it afterward, you save time, but your raw records reach that company first. That can count as a disclosure of personal data to another company and needs a contract that fits the law, such as a processor contract under GDPR Article 28. Ask counsel which role the buyer plays.

Whichever route you take, get four things in writing: who removes the data, by what method, whether you can review a sample first, and who is liable if something is missed. Employee information is personal data too, so decide whether internal chat and email belong in the package at all, and leave HR and payroll files out. Health, financial and children’s data have their own rules, so get counsel’s sign-off before any of those records go in. Your own privacy policy and customer terms are promises as well: the FTC has warned that quietly changing a policy to allow new uses such as AI training may be unfair or deceptive (FTC, February 2024). Our guide to preparing company data for licensing covers the removal method, formats and secure transfer.

5. Choose a buyer and agree the scope

Three parties sit in the chain: your company holds the records, a data company licenses and prepares them, and an AI lab receives the prepared data. Buyers publish their own requirements and terms, and the terms differ. Our comparison of AI data buyers sets them side by side, and our review of a buyer’s program shows how to read one program in detail.

The buyer reviews the company and its records and decides whether to make an offer, so meeting its requirements does not guarantee one. Apply with your inventory in hand: which systems, how many years, and whether client data is involved. If more than one buyer fits, compare them on whether you meet the stated requirements, who removes personal data and who pays for it, whether the license is exclusive and for how long, when and how you are paid, and what you may still do with your own data afterward.

Before data moves, agree these points in writing:

1. Scope. Which systems, date ranges and record types are included.
2. Personal data. Who removes it, by what method, and who checks the result.
3. Approvals. Whether you review a sample first and who can reject a batch.
4. Permitted use. What the buyer and its customers may do with the data. Ask the buyer to confirm in writing which labs may receive it, how it is kept separate from other data and how long the buyer retains it.

Example: A 90-person engineering firm agrees to license its project records from 2016 onward and its knowledge base. It leaves out the helpdesk, which is mixed with client material, and asks to review a sample batch before the full transfer.

6. Ask your questions, then sign

The license sets what the buyer may do with the data, for how long and whether the grant is exclusive. Read the draft with these questions in hand:

1. Who signs on the other side? Confirm the legal entity, who the end recipients will be and who stands behind the buyer’s promises if a recipient misuses the data. Ask whether the buyer may transfer the license to another company, for example after an acquisition.
2. What exactly is licensed? Look for a precise description of the systems and date ranges and a clear list of permitted uses, including whether the buyer may train, fine-tune and test models on the data and sell those models commercially.
3. Is it exclusive, and for how long? Ask the buyer to price exclusive and non-exclusive versions. If the gap is small, you are giving up options for little.
4. What do you promise? By warranting that you own the data and may license it, you take responsibility if a client claims the records were theirs or personal data turns up in what you delivered. Ask whether your warranties can be limited to what you know and whether your liability is capped.
5. Who checks the personal-data removal? Write the standard into the contract, with a duty to notify you if personal data is found later and a ban on re-identification.
6. What happens at the end? Ask whether the buyer must delete or return the data and certify it, and check that the clause covers models trained on the data as well as the files.

Be wary of warranties that cover data you did not create, uncapped indemnity, exclusivity across every field when the buyer needs only one, and audit rights with no limit on what the buyer can inspect. Have a lawyer review the draft, and ask your CPA how the fee will be recorded and taxed.

7. Transfer the data and get paid

Use the transfer method the buyer approves and agree who confirms receipt. Don’t email archives. Encrypt files in transit and at rest, send a small sample first, and keep a manifest that lists every file with its size and a checksum so both sides can confirm what arrived. Limit who can run the export, log each transfer and strip credentials from anything you send.

The agreement sets payment, and the usual structures are:

• One-time fee. A single payment, often paired with a perpetual license.
• Installments. Payments tied to delivery, review or approval milestones.
• Recurring fees. Annual or periodic payments under a term license.
• Per-scope fees. Separate fees for each workflow or dataset added.
• Usage-based payments. A share of revenue or a fee tied to use, which needs a clear way to measure and audit that use.

Ask when payment is due and what triggers it: signature, transfer or approval. Ask whether the buyer can accept part of your data and reject the rest, what happens to the fee if a batch is rejected, what happens to payment if the license ends early, and who bears taxes and transfer costs. What a buyer offers depends on the records and on terms such as exclusivity and duration. Our guide to valuing company data takes each factor in turn.

Example: A 70-person company agrees to a non-exclusive, three-year license to its knowledge base and project records. It reviews a sample before transfer, and payment comes in two installments: one on transfer and one on approval of the data. The agreement requires deletion at the end of the term and addresses models already trained on the data.

When licensing may not fit

Licensing does not suit every company. Pause if any of these apply:

• Most of your records are client data covered by contracts or NDAs, which leaves little you can license.
• You are preparing to sell the company. An exclusive, perpetual license is a commitment an acquirer’s lawyers may read closely, so talk to counsel first.
• Your records are thin or inconsistent, so the preparation work may not be worth the effort.
• You cannot spare the staff time for exporting, sample review and buyer questions.

Next steps

Start with the inventory in step 1 and the ownership check in step 2. Then use the eligibility check below for a preliminary answer on whether your records could qualify.