// wiki · guide · September 2026
Deals with AI products
Draft article. The deal forms are working reference points; the content of a specific contract is determined by the product and the parties.
A draft for discussion. The article answers four questions of an AI product deal: what to check, what to transfer, how to formalise and how to get paid.
The short answer
A deal with an AI product is a deal with layers: code, weights, datasets, documentation, people. The main choice is whether to transfer the model itself (a licence or an assignment) or access to it (SaaS, API, white-label). Most conflicts arise not from price terms but from defaults: the parties did not fix which layers pass and what happens when the model is updated.
The check
The buyer and the investor check six blocks: rights to the model and code, dataset provenance, third-party licences and open source, regulatory requirements, the team, client contracts — the full analysis in “Due diligence of an AI asset”. For a licensee a shortened checklist suffices: confirmation of rights (registers, contracts), absence of known claims, the composition of components.
Forms of transfer
| Form | What passes | When to choose it |
|---|---|---|
| Non-exclusive licence (by instances, users, inference volume) | The right to use the model within agreed limits | A typical B2B product sale; the vendor keeps control |
| Assignment of rights | Code, weights, datasets, documentation — in full, with warranties of title | Selling a product or a carved-out technology |
| SaaS / API subscription | Nothing passes; access under an SLA | A quick start, no infrastructure on the client's side |
| White-label / partner programme | The right to render the service under the partner's brand | Scaling through integrators |
| Joint development | Rights divided by agreement | Industry models on a customer's data |
Transferring the technology
- Completeness: weights without inference code and hyperparameters do not run — the contract lists the artefacts (checkpoints, configurations, pipelines, tests).
- Know-how: agree measures to preserve confidentiality when transferring and using information. Protection depends on Civil Code Article 1465; loss of confidentiality terminates the exclusive right under Article 1467.
- Datasets: only those with rights to them are transferred; the provenance log passes with them (data provenance).
- Updates: a licence of “weights 1.0” without fine-tuning terms loses value in a year — fix the versioning procedure.
Monetisation
- Subscription and payment for usage (tokens, inference): predictable for the client, requires metrics in the contract.
- Royalties from revenue or usage — a frequent compromise in white-label deals.
- A licence by domains and terms — the basic model of B2B sales.
- Taxes: a licence for register software and access to a service are taxed differently (the tax regime of an IT company); foreign clients add VAT questions (foreign services — the mirror situation).
Typical mistakes
- transfer “by acceptance act” without a list of artefacts and versions;
- a licence without usage limits — the client builds a product competing with the vendor's on the model;
- forgotten open source components in the stack — the product cannot be transferred (open source in the product);
- no terms on fine-tuning on the client's data: who owns the adapted version.
Drafting or reviewing an AI product contract is the work of technology rights practice and deals practice.
Sources for selected points
This selection concerns only the identified issues and does not constitute a full legal review of the article.