Training license vs RAG license: what changes for you?
The short answer
A training buyer's model usually keeps what it learned from your records, which Troveo says almost every buyer insists on. Retrieval platforms such as Sphere charge per use while access lasts. Company-data programs describe training.
Most of what you negotiate in an AI data license follows from one choice: will the buyer train on your records, or look them up each time someone asks?
That one difference changes how you get paid, what you can take back at the end and whether your name ever appears.
What is the difference between a training and a RAG license?
A training license lets a model learn from your records, and the buyer usually keeps that model after the license ends. A retrieval license, often called a RAG license, lets an AI system look your library up each time someone asks, so the library has to stay available.
RAG is short for retrieval-augmented generation. The system searches the licensed library at the moment of the question, then writes its answer from what it found.
Troveo's guide to licensing deals (1 Sep 2026), from a data licensing marketplace, lists training, fine-tuning, evaluation and retrieval at answer time as separate rights with different values. One contract can grant any mix of them.
Training license and retrieval (RAG) license, term by term, from the public sources on this page.
| License term | Training license | Retrieval (RAG) license |
|---|---|---|
| How the buyer uses it | A model learns from it | Looked up for each answer |
| Pricing basis | Flat, per-unit or recurring fee | Per query, token or request |
| Term | The model usually outlives the term | Use runs while access lasts |
| At the end | Files may be deleted, the model usually stays | Access can be switched off |
| Your name | No attribution term found | Results carry the source |
| Who offers it | Company-data programs and labs | Publisher licensing platforms |
| Usually fits when | You hold a finished archive of internal records | You publish material people look things up in |
What does a training license look like?
A training license names training as the permitted use, and the trained model usually stays with the buyer afterwards. OpenAI's Data Partnerships page (9 Nov 2023) offers to prepare private datasets "for training proprietary AI models," including its foundation models.
Wiley's annual report for the year to 30 Apr 2026 (filed 24 Jun 2026) says the science and education publisher licenses content for AI, including training large language models.
For its learning content, those licenses carry a fixed price, and Wiley books the revenue at one point in time instead of spreading it over a term.
micro1's Enterprise Data Partnership page says a company's operational data and workflows help train and evaluate frontier AI models. That is micro1's own description of its program.
What does a retrieval license look like?
A retrieval license names lookup at answer time as the use. Wiley's 3 Mar 2026 release describes a deal close to one: OpenEvidence will cite Wiley content in its answers. The release doesn't say whether training is excluded.
Cochrane reviews and more than 400 journals and books will be referenced inside OpenEvidence's evidence layer. OpenEvidence is a medical answer engine for physicians, and the release describes its models as trained on peer-reviewed literature.
Two licensing platforms describe the retrieval model on their own pages:
- Sphere's FAQ says no training use is permitted and access is for inference and retrieval only, though the same page also mentions a "fine-tuning stack." Buyers pay by query, token or request volume, and every result carries source metadata.
- TollBit offers publishers "Licensed RAG access," and its bot and agent paywall uses usage-based pricing on terms the publisher sets.
Some deals cover both. Google's 22 Feb 2024 post on its Reddit partnership says the Data API gives it real-time content to "display, train on, and otherwise use." Read the permitted-use clause itself before you trust the headline.
How does the money differ?
Troveo's guide says a finished archive is licensed once for a flat price, and that the largest disclosed deals are mostly recurring, for sources that keep producing. Wiley's learning-content AI licenses are fixed-price deals too.
On the two retrieval platforms above, pay follows use. Sphere prices by query, token or request, and TollBit describes usage-based pricing, so the income tracks how often AI systems access your content and ends when that stops.
Think of training as letting a chef memorize your recipe binder for an agreed fee, and retrieval as charging every time they open it. Neither platform publishes a rate card, so no public number tells you which pays more for your material.
What survives when a training or retrieval license ends?
A trained model usually survives. Troveo's guide calls it the clause almost every buyer insists on, because, in its words, "nothing can be untrained." A retrieval license ends with access, and Sphere says licensed content is never fully exported.
Whether the buyer must delete your files at expiry is a separate, negotiated term. For retrieval, write deletion of any copies or search indexes into the contract anyway.
Model deletion has happened, through a regulator. The FTC's final order against Everalbum, a photo-app company, required it to delete the models and algorithms it built from users' photos and videos (FTC, 7 May 2021).
That followed FTC allegations that it deceived users, not a license expiring. Can you delete your data after training covers what regulators and contracts can and can't reach.
Does your company's name show up?
In the Wiley release and on Sphere, the source travels with the content. The Wiley release says its content will be cited in OpenEvidence's answers the way the platform already cites its sources, and Sphere attaches source metadata to every result.
For training, Troveo's guide names output restrictions, which limit a model reproducing licensed material word for word, and lists no attribution term.
For company records, staying unnamed is usually the point. Troveo's business data page says companies are never named publicly, and Handshake AI's company-data program says partner companies are never identified.
Which one fits company records?
The retrieval platforms above serve publishers, and the buyer profiles for TollBit and Dappier say the same. Internal records such as Slack, email and tickets reach buyers through programs that describe training.
If you publish a help center, product documentation or a research library, retrieval is worth asking about. Whichever you are offered, put these to the buyer in writing:
- Which uses does the license name: training, fine-tuning, evaluation, retrieval or a mix?
- Does the trained model survive the term? Assume it does unless the contract says otherwise.
- Must the buyer delete your files, copies and search indexes at the end, and confirm it in writing?
- Is the fee fixed or tied to use, and who reports the usage?
- Will your company be named, cited or kept anonymous?
Exclusivity is its own question, covered in what an exclusivity term means. The promises a buyer will ask of you are in what reps and warranties a buyer asks for, and the full contract walk-through is the data licensing agreement guide.
The free rights check walks through seven rights questions to take to your lawyer. Optional, paid managed readiness builds the systems inventory, questions for your counsel and buyer shortlist from descriptions and counts, quoted in writing first.
Apply to micro1's data partnership (paid link)micro1's own data-partnerships pagePaid link: micro1 may pay us a referral fee if a company we introduce goes ahead with it. You pay nothing extra.
Vaultfish helps you run the process. We are not lawyers and this is not legal advice. If you need counsel who knows this area, we can point you to one. Ask for a referral
