How long does an AI buyer keep licensed data, and is it deleted?

The short answer: the agreement sets the clock

How long an AI buyer keeps licensed data is decided by the license agreement, not by a general industry rule. The term, the exit provisions and any deletion or return obligations in the signed contract determine what the buyer may hold, for how long, and what it must do when the license ends.

Nothing is binding until the company agrees price and terms and signs, so retention is a negotiable term, not a surprise. The company keeps ownership of its data; it is licensed, not sold. Deals are typically exclusive for AI training for an agreed term, and the end of that term is the natural point to settle what happens to every copy.

This page is general information about how such terms are usually structured. It is not a promise about any specific buyer, and the actual terms are set in each agreement.

What copies of the data can exist?

Retention is rarely about one file. Ask the question for each layer, because an agreement that covers only the original delivery can leave the others open.

The practical lesson is to list the layers in the contract itself. A clause that says "delete the data" without defining "data" invites argument later.

Can trained model weights be deleted?

Generally not in a clean way. Once data has been used in training, its influence is spread across the model's parameters, and a buyer cannot simply remove one company's contribution the way it deletes a file. Anyone who promises otherwise should be asked how it would be verified.

What an owner can negotiate instead is control over the inputs and the use:

• A defined purpose and a defined term for training use.
• A ban on onward transfer or resale of the dataset.
• Deletion of every retained copy of the source data at term end.
• Limits on using the data to build competing products for the company's market.

The honest framing for a company is that deletion applies to the data the buyer holds, while learned weights are governed by use restrictions. Read the fair use and licensing explainer for why buyers still pay for permissioned data.

What should an owner ask for at term end?

Ask for outcomes you can check. A company does not need to be a lawyer to request these, and its counsel can turn them into clause language.

1. Return or deletion of the delivered dataset and all working copies by a stated date.
2. Written certification, signed by an officer of the buyer, that deletion is complete.
3. Coverage of derived artifacts such as caches and embeddings, named in the clause.
4. A reasonable audit or verification right, or a periodic compliance statement if a full audit is unrealistic.
5. A defined treatment of backups and logs, with a purge schedule.
6. A statement of what survives termination, such as confidentiality duties.

Who may see the data while the license runs is a separate question, covered in who has access to data in a licensing deal.

Post-delivery obligations checklist

Use this list when reviewing a draft agreement with counsel.

• The term and any renewal or extension mechanism are written down.
• Retention periods are stated per layer, not just for the main dataset.
• Deletion or return is triggered automatically at term end and on early termination.
• Certification of deletion is required and names who signs it.
• Derived artifacts, logs and backups are addressed.
• Onward transfer, resale and sub-licensing are restricted.
• Confidentiality survives the term.
• The company knows who at the buyer is accountable for compliance.

Identity of the buyer is also negotiable in some structures; see whether a company can choose or veto AI buyers and why buyers are often kept confidential.

What this means for a referral partner

Partners do not negotiate retention terms and should not describe what any buyer will do. Your role is to introduce the company; SourceX and the company's own counsel work through the agreement, and de-identification and redaction requirements are agreed with the company before any work begins. Data is delivered only after an executed agreement and the company's authorization.

When an owner asks about retention, say what is true: it is set in the contract, the company can ask for deletion or return and certification, and it should have its lawyer read those clauses before signing. Do not promise that data can be erased from a trained model, and do not quote terms you have not seen.

If the owner's concern is reputational rather than contractual, point to the sponsor-oriented risk guide on portfolio data licensing.

When retention concerns are a real stop sign

Pause the introduction if the owner insists on a guarantee that data will be removed from trained models, or if the company cannot accept any term at all in which a buyer holds a licensed copy. Those positions are inconsistent with how a training license works, and the conversation is better ended politely than stretched.

Retention is also a poor fit for a company whose records are mostly someone else's data, or mostly consumer personal data with no licensing basis. See onboarding and training records for an example of a record type where scope needs careful definition, and the referral earnings calculator if the reader wants to understand how the program formula works.

Next step

If you know a US company with 50+ full-time employees at peak (contractors excluded) and years of operational records, register as a partner and make the introduction. Companies can also apply directly at sourcex.si/apply. Questions about the program itself are answered in the FAQ.