Training license vs RAG license: how the two AI data licenses differ

The short verdict: training licenses buy learning, retrieval licenses buy access

A training license lets an AI developer use a dataset to adjust a model's weights, so the model learns patterns from the records while the records themselves are never shown to end users. A retrieval license, often called a RAG or grounding license, lets the developer store content in a search index and pull passages into answers at the moment a user asks a question.

That single difference drives everything else. Training licenses usually cover a fixed snapshot delivered once and paid once, because the value is in what the model learns. Retrieval licenses usually need ongoing, current access and recurring payment, because the value is in fresh content shown with attribution.

For most established companies the choice is simple. Internal operational records, such as support tickets, project files, CRM histories and approval chains, are confidential and should never be quoted to the public, so a training license, sometimes alongside an evaluation license, is the relevant structure. Retrieval and display licenses suit publishers and platforms whose content is already public-facing.

Training license vs RAG license, side by side

The end-of-term row is the one contracts spend the most time on. A model cannot easily forget what it learned, so a training license should say plainly whether models trained during the term can keep being used after it ends.

Where evaluation and display licenses fit

Training and retrieval are not the only rights an AI developer may ask for. Two others come up often, and one agreement can combine several of them.

Evaluation rights matter because a test set loses its value once a model has seen it; the guide to benchmark contamination explains why never-published records make clean tests.

How public deals have combined these rights

Published announcements show the categories often travel together. These examples involve publishers and platforms, not SourceX deals, and are listed with their dates.

Two lessons carry over to any company. The Reddit figure is a multi-year contract total across several agreements, not annual revenue, so read headline numbers with care. And deals built on continuous access look very different from a one-time training delivery; the overview of enterprise data licensing deals covers the quieter side of the market, where operating companies rather than media brands license records.

Why the rights can be licensed separately

US copyright law treats ownership as divisible. Under 17 U.S.C. 201, ownership may be transferred in whole or in part, and any of the exclusive rights may be transferred and owned separately. That is the legal footing for licensing training use while keeping every other right, or for granting retrieval rights to one party and training rights to another.

Business records also carry contract and confidentiality obligations that copyright does not cover, so scope is set in the license agreement itself. Structure can affect accounting too: Deloitte's guidance on ASC 606 licenses of intellectual property distinguishes a right to use IP as it exists when granted from a right to access IP over the license period, which can change when the licensor recognizes revenue. A company should ask its auditors how a specific license will be treated.

This is general information, not legal, tax or financial advice. Confirm with your own counsel, tax adviser or professional body before acting.

When a training license is the better fit

Choose a training structure when most of these are true:

• The records are internal and confidential: tickets, email threads, CRM stages, approvals.
• The value lies in patterns across years of work, not in last week's content.
• The company prefers a one-time payment to an ongoing technical integration.
• Nobody wants the records quoted, summarized or linked in someone else's product.
• The company can deliver an agreed snapshot once, under redaction rules it sets.

When a retrieval or display license is the better fit

Retrieval and display rights make sense when content is meant to be read by the public and loses value as it ages: news, reference material, product documentation, reviews or forums. The owner usually wants attribution and traffic, can support a live feed or API, and accepts revenue that rises and falls with usage.

Internal business records almost never fit that description. A company selling retrieval access to confidential records would be agreeing to have them surfaced in another firm's product, which few owners, or their clients, would accept.

How SourceX deals are structured

SourceX deals typically center on AI training. They are usually exclusive for AI training for an agreed term, priced as one all-in figure with SourceX's fee included, and paid as a one-time payment, typically within about 60 days of invoicing once the buyer selects the data. The company keeps ownership because the data is licensed, not sold, as the page on licensing versus selling data explains. Nothing is binding until the company agrees price and terms and signs.

Before signing, the company and its counsel should settle these scope questions:

• Which uses are permitted: training, evaluation, retrieval or display.
• How long exclusivity lasts and exactly which use it covers.
• Whether models trained during the term may be used after it ends.
• What limits apply to reproducing records verbatim in model outputs.
• Which redaction and de-identification rules apply before delivery.
• Whether privacy notices and client contracts permit the use; the page on ethically sourced AI training data covers consent and provenance.

Referral partners do not negotiate any of this. A partner makes the introduction; SourceX and the company handle qualification, the data inventory, terms, buyer review and delivery, as laid out in how it works. Partners earn 25% of the eligible platform fees SourceX actually collects from the referred company's licensing deals, capped at $100,000 per referred company; the reward becomes payable only after the buyer pays and SourceX receives its fee, and no reward is guaranteed.

Next step

If a client holds years of internal records and asks how an AI license would work, start with the scope questions above. Run a quick, non-binding screen with the company fit checker, then register as a partner to make the introduction. A company can also apply directly at sourcex.si/apply.