Licensing records for training means a buyer uses them to change a model’s parameters, which generally cannot be undone. Licensing for retrieval, often called RAG, means a system looks up your records at answer time and can stop using them when access ends. The two uses carry different risks and are often priced and drafted differently, so define the permitted use precisely.
- Training is hard to reverse; retrieval access can be revoked.
- Define permitted use explicitly: evaluation, training, fine-tuning, retrieval or all.
- Deletion clauses mean different things for each use.
How the two uses differ
| Training or fine-tuning | Retrieval (RAG) | |
|---|---|---|
| What happens | Records shape model parameters | Records are searched at answer time |
| Reversible? | Generally no | Yes, by ending access |
| Output exposure | Indirect | Passages can appear in answers |
| Typical structure | Defined copy, one-time or periodic | Ongoing access, often recurring |
| Key clause | Use of trained models after term | Attribution, quoting and access controls |
Defining use in the agreement
Vague phrases such as ‘AI purposes’ can cover both uses. Spell out:
- Which uses are permitted: evaluation only, training, fine-tuning, retrieval, or a combination.
- Whether models trained on the records may be used, sold or shared after the term.
- Whether retrieved passages may be shown to end users verbatim.
- Deletion of copies at the end of term, and what deletion cannot achieve.
- Restrictions on use by or for your competitors.
Which fits your records
Archives that will not change suit a defined training license. Records that are updated constantly, such as documentation or knowledge bases, can suit retrieval access. See pricing models and the agreement checklist.
Copyright context
The U.S. Copyright Office has published reports on copyright and AI, including training. A license is the clearest way to set terms for your own content regardless of how those questions develop.
Next step
Record your preferred use before a draft arrives and compare proposals with the proposal review checklist.
Common questions
What is RAG?
Retrieval-augmented generation is a technique where an AI system searches a document collection and uses the retrieved passages to answer a question, rather than relying only on what the model learned during training.
Can a trained model forget my data?
Deleting the source copy does not remove what a model has already learned. Ask how the buyer handles models trained on your records if the license ends.
Which pays more?
There is no general rule. Training licenses are often one-time grants; retrieval licenses can involve ongoing access and recurring payments.
Sources and further reading
Origin Data Partners is independent and not affiliated with any buyer named on this page.