How to anonymize data for AI and prepare it for licensing

To anonymize data for AI licensing, take stock of your records, confirm you own them, remove personal data to a documented standard and agree a secure transfer with the buyer. Some buyers do part of this for you, so settle first who removes personal data and who is liable if something is missed.

Start with an inventory of your data

A data inventory is a list of the systems that hold your records, so you can decide what to offer and what to leave out. Build it before you talk to a buyer, because every later step depends on it. Record these fields for each system.

Our guide to what company data you can sell explains which record types buyers name.

Confirm you own the records and can license them

Ownership has three 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, customer and third-party material. 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 is what gives you ownership, so read each contractor agreement. Text that customers write in tickets and chats is a separate case, so check what your customer terms let you do with it. Leave out vendor manuals, licensed reports and other third-party content, and check the license on any open-source code.

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. Treat a perpetual or exclusive license as a major decision and get approval from whoever your governing documents require.

How to anonymize data for AI buyers

Anonymizing means removing personal data until the records can’t reasonably be linked to a person. California’s law defines personal information broadly. Its list starts with identifiers such as a real name, alias, postal address, online identifier, IP address, email address, account name, and Social Security, driver’s license and passport numbers. It goes on to cover purchase histories, geolocation, audio and visual recordings, and professional or employment information (Civil Code § 1798.140). GDPR defines personal data as any information relating to an identified or identifiable person (GDPR Article 4). Our guide to whether selling company data is legal explains when removal takes records outside those laws.

PII redaction, the removal or masking of personally identifiable information, is the main technique. Work through your records in this order:

1. Remove direct identifiers. Names, email addresses, phone numbers, postal addresses, account and ID numbers, IP addresses and usernames.
2. Replace what you need to keep readable. Swap each identifier for a consistent placeholder, such as CUSTOMER_1, so a ticket thread still reads in order. If anyone can use a key to trace placeholders back to people, GDPR still treats the data as personal data, so destroy the key if you don’t need it, or keep it where the buyer can’t reach it (Recital 26).
3. Generalize indirect identifiers. Exact dates, small locations, job titles and rare combinations of facts can identify a person without a name. Coarsen them where the record can stand it, such as an exact date to a month.
4. Search free text, attachments and metadata. Tickets, chats and emails carry names, phone numbers and order numbers inside sentences. Screenshots, PDFs, file properties, email headers, call audio and URLs with tokens can hold identifiers too.
5. Remove credentials and secrets. Passwords, API keys, access tokens and internal hostnames turn up in tickets, chats and code.
6. Audit a sample. Have someone who did not do the cleanup read a random sample from every system for anything left over. Repeat until the sample comes back clean, and record the method and the result.

Automated redaction tools can find many structured identifiers and are worth using at volume. They can miss identifiers in free text, so the sample audit is the check you can rely on.

Health data follows its own rules. If you are a HIPAA covered entity or business associate, HHS describes two ways to de-identify protected health information: Expert Determination, and Safe Harbor, which removes 18 types of identifiers (HHS guidance).

Decide who does the removal

Buyers differ. Some de-identify the records themselves, and some use a specialist partner to strip personal data before AI labs see the records.

Doing the removal yourself means less personal data leaves your company and you control the method. Letting the buyer do it saves your time, but raw records reach the buyer first, which can count as a disclosure of personal data to another company. 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. Our guide to data licensing covers the related clauses.

Choose formats and structure

Ask the buyer which formats it accepts before you convert anything. Whatever the format, keep the structure that makes records useful: titles, dates, the order of messages in a thread, links between a ticket and its resolution, and version history where you have it. Where a person was named, use a role or a placeholder so the record still reads.

Add a short data dictionary that lists what each system is, what each field means, the date range, the record counts and any known gaps. A buyer who can read your dictionary should be able to review your data faster, and you will have the answers ready for the contract.

Update your notices

Your privacy policy and customer terms are promises. If they say you never sell or share customer data, licensing records that still contain it conflicts with them. 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).

If you license data as deidentified under California law, the statute expects you to commit publicly to keep it deidentified and not to re-identify it, and to bind every recipient to the same terms by contract. Tell employees if internal messages are part of the package, and keep a short record of what you decided and why.

Plan the secure transfer

Use the transfer method the buyer approves. 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 inside your company can run the export, log each transfer, and strip credentials from anything you send. Agree in writing what the buyer deletes or returns when the license ends.

What the timeline looks like

1. Inventory and ownership check. Your team does this before contacting a buyer.
2. Application and review. The buyer decides whether to proceed.
3. Terms and method. You agree price, scope, anonymization and deletion.
4. Sample review. Ask to review a representative sample before the prepared data is used.
5. Transfer and acceptance. The buyer reviews what arrived.
6. Payment. It follows acceptance.

We expect the whole path from application to approved data to take months. Steps 1 and 3 are the part you control, and starting the inventory early shortens it.

Next steps

Finish the inventory and the ownership check first. Then compare buyers in our guide to AI training data buyers, or read how to sell your company’s data to AI companies for the whole process. To see whether your records could qualify, the eligibility check below asks four questions.