Sell RL environments to AI labs
An RL environment is a practice exercise for an AI model: a task to do, a world to do it in and a checker that scores the result. If you have built environments, or your company built them to test its own AI, AI labs may license them. We take a summary to a short list of labs we choose for the best quote. Only after you sign does a lab see more than a sample.
Tell us about your RL environmentsWhat's my data worth?
New to the term? Read what RL data is, or look up a word in the glossary. Selling records instead? See how to sell or license your data.
What an RL environment is
| Part | What it is |
|---|---|
| A task | What the AI must do, such as find a way into a web app or close a support ticket correctly |
| A world to act in | A copy of a real app, tool or workflow with realistic data, which can be reset after every try |
| A checker | A program that looks at the result and scores it, for example by confirming a change really happened in the system |
RL stands for reinforcement learning: the model tries, gets a score and learns from it. The checker is the valuable part, because it is what turns a task into a score.
What AI labs look for
| Quality | What it means |
|---|---|
| Graded by a program | The checker decides, not another AI. A score that depends on an AI judge is hard for a lab to trust |
| Measured difficulty | You can say how hard it is and how you measured it, for example against a strong model |
| Hard to memorize | Each task comes in several seeded variants, with different data and secrets, so a model cannot learn one answer |
| No personal information | All people, accounts and records inside it are made up |
| Clean rights | You wrote it, or you hold the right to license every part of it |
| Runs on the lab's own systems | Delivered as software the lab runs itself, so nothing of theirs leaves |
| A short datasheet | What is in the set, how many tasks, what languages and tools, and how it is graded |
What cannot be licensed
- Code or content you do not own, including open-source parts with terms that do not allow this use
- Anything you built for a client under a contract that gives them the work
- Real customer data, even if only a little of it is inside
- Content built from real security flaws that were never reported to the vendor. Report them first, and wait until you can show they were handled
Security content can be used both to defend and to attack, so ownership and responsible disclosure come first. Whether and when you may license something built from a real flaw is a question for your counsel.
How a license usually works
- Non-exclusive, so the same set can go to more than one lab, or exclusive for a set term.
- Priced per environment or per set, agreed in the contract.
- A small sample under a confidentiality agreement, so a lab can try it first.
- A watermark per buyer, so a copy can be traced.
- No promise of a price or a deal. We say plainly what we see.
What to prepare
An enquiry that comes with numbers and a sample gets looked at first. Before we approach any lab we need these, and if something is missing we ask you for it.
| Item | What it means |
|---|---|
| Counts | Number of environments and tasks, by type |
| Difficulty | How you measured it, and the result |
| A datasheet | One or two pages on what is in the set and how it is graded |
| One sample task | Something a lab can try, with its checker |
| Who owns it | The company or person that holds the rights, and which one would sign the license |
| What is already promised | Any lab you have shown it to, and any exclusive terms |
How we help, and how a deal runs
We bring your environments to several AI labs and ask each for an offer, so you can compare price and terms and make the final call. The AI lab usually pays our fee after the deal closes.
Tell us what you have
A short message, then a 30-minute call.
We check rights and fit
Client-held material comes out. We match you to current demand.
You sign with us
You don't sign anything to start. When a buyer or AI lab is interested, we ask for an exclusive broker agreement for three months while we present it.
AI labs quote
Your summary, never your data, goes to a short list. You choose.
Personal details removed
Names become stand-ins like Employee 632, checked by people.
You get paid
The lab accepts the data and pays on the written terms.
Other kinds of data: which records can be licensed, industries we cover and company workflows.
Questions about selling RL environments
Do you need to see the environments first?
No. We start from a short summary and one sample task that you approve. The full set stays with you until you sign.
Do you build environments for me?
No. We bring yours to AI labs. If you hold records or documented processes but have not built environments, see how to sell your company workflows.
I have already shown it to a lab. Does that matter?
Tell us which one and on what terms. We leave that lab out, and we ask whether there are others you want left out too.
Can the same environment go to more than one lab?
Yes, when the license is non-exclusive, which is the usual case. An exclusive license for a set term is agreed in the contract and limits what you can do with the same set in the meantime.
What if part of it was written by an AI?
Say so, and say how much. Some labs ask for that up front, and a clear answer helps more than a surprise later.
What does a license pay?
It depends on the number of environments, how hard they are, how clean the rights are and what labs are asking for at the time. We prepare the package first and the amount is negotiated after, so we do not quote one. The AI lab that buys pays our broker fee, after the deal closes.
Can I keep using my environments?
Yes. You license them, you do not sell them away, so you keep them and can use them for your own work.
Didn't find your question? Ask us directly.