Guides
Licensing company data for AI training: the guide for European companies
By Francesco Wiedemann and Peter Meier · · 4 min read
In short
- AI data licensing means granting an AI developer the right to train or evaluate models on a dataset, with a defined scope and clear restrictions.
- AI labs pay most for complete records of real work: how a team moves a task across tools, with corrections and a visible result.
- Data you process on behalf of customers can't be licensed without their permission; your own operational records usually can, once the legal basis, works council and trade secrets are cleared.
- A licence typically starts with a one-off payment for historical data and can continue as recurring revenue for data you keep producing.
Why AI labs pay for company data
Large language models have been trained on most of the public internet. What is missing is a record of how professional work actually gets done: a pricing decision that moves from a chat thread through a finance check to an updated CRM record, an IT ticket that is diagnosed and fixed, a code change that passes review. AI labs need exactly this to train and evaluate AI agents that work across real business tools.
Companies produce these records every day as a by-product of running the business. We call it operational exhaust. Unlike public web text, it is scarce, it shows multi-step work with outcomes, and it can't be scraped. That is why it has become a licensable asset.
Which company data qualifies
The most valuable data shows a complete piece of work across several tools, including mistakes, corrections and the result. Isolated records, data without outcomes and content that is already public are worth much less.
| Data type | Typical tools | Why labs value it |
|---|---|---|
| Sales and pricing | CRM, email, team chat | Decisions with approvals and a measurable result |
| Finance and operations | ERP, spreadsheets | Rule-based processes with exceptions |
| Engineering and product | GitHub, GitLab, Jira, Linear | Issues, code changes, reviews and incidents |
| Internal IT and helpdesk | Ticket systems | Diagnosis and resolution, step by step |
| Knowledge and processes | Wikis, documents | How procedures are actually used |
| AI-agent traces | Copilots, own agents | Tool calls with human corrections |
What can't be licensed
- Data you process on behalf of customers. Anything covered by a data processing agreement belongs to your customers' purposes, not yours.
- Data you don't hold the rights to, for example third-party content under restrictive licences.
- Anything you choose to exclude. A licence is defined source by source and time period by time period.
What has to be cleared first
Whether and which data can be licensed is decided per data source, not across the board. Four questions come up in almost every case:
- Legal basis under GDPR. Personal data in company records needs a legal basis for the new purpose. Anonymisation is itself a processing step that needs one. See our GDPR guide for AI training data.
- Works council. Where employee data is involved, co-determination rights may apply, and a works agreement can be required.
- Customer and third-party rights. Contracts are reviewed so that only data you control is included.
- Trade secrets. Commercially sensitive details are removed or generalised before anything is licensed.
How licensing works, step by step
- Check eligibility. A short assessment shows whether your team and tools produce relevant data. No data is uploaded.
- Clear. Together, legal basis, works council, customer contracts and trade secrets are reviewed. You decide what is in scope.
- Anonymise and refine. In the EU, names are replaced, details generalised or removed, and the result is checked. The data is then structured into training tasks.
- License and earn. The refined dataset is licensed to AI developers. You receive a first payment for historical data and recurring revenue whenever data from your ongoing stream is licensed.
Your effort depends on the data sources. For structured data without communications it is typically a few hours. Where employee data or communications are involved, steps such as a works agreement, a balancing of interests and informing employees can be added.
What company data is worth
Value depends on how rare, complete and recent the data is, how many tools it connects, how much of it there is, and which rights you grant. Providers in this market publish ranges from tens of thousands to several million US dollars per deal. At IsarTrust, we only start a licence if it earns your company at least €20,000.
How to choose a licensing partner
- Where is raw data processed, and who has access to it?
- How is anonymisation done and checked, and can you review a sample before delivery?
- Is your company named to buyers, or is the dataset delivered under a code name?
- Is the partner the licensor (refine and license) or a marketplace that connects you directly with buyers?
- How are payments structured: one-off, recurring, or both?
For an overview of the providers, see our comparison of AI data licensing providers.
Want to know whether your company's data qualifies? The eligibility check takes about a minute and needs no data upload.
Frequently asked questions
Do we lose ownership of our data?
No. You license an agreed, anonymised copy for defined purposes. Your original data and its ownership stay with you.
Can we license data that contains customer information?
Data you process on behalf of customers under a data processing agreement is excluded. Your own operational records can usually be licensed once they have been cleared and anonymised.
Do we need a minimum company size?
There is no formal minimum, but from around ten people a team usually produces enough connected data to be relevant.
Is your company's data licensable?
Three questions, about a minute, no data upload.
Check your eligibility →