Sublicensing and resale rights in a data licence

Four permissions that get confused

Deals go wrong when a team negotiates for one of these and receives a clause written for another. "The licensee may share the data with its affiliates and partners" reads like the third item on this list and is almost always the first.

• Internal use across a corporate group. A parent, its subsidiaries and its foreign branches are separate legal entities, and a licence granted to one does not automatically extend to the others.
• Use by contractors and service providers. A vendor annotating, hosting or evaluating the data for you is a processor, not a sublicensee. That normally needs a flow-down obligation rather than a sublicence, but it still needs a clause.
• Distribution inside a product. If your customers receive the material, or something derived from it, that is redistribution and it is a different permission from training.
• Resale as a product. Selling the corpus, or a material part of it, as a dataset. This is the one almost never granted, because it competes directly with the vendor's own business.

Affiliates: define them, and list them

Affiliate definitions are usually written as a control test — an entity that controls, is controlled by, or is under common control with the licensee, above some ownership percentage. That covers the group as it exists at signature and can fail quietly as the company changes.

Two additions close the gap. A right to add new affiliates by notice, with the vendor's approval not unreasonably withheld, instead of a requirement to renegotiate the licence. And a rule for what happens when an affiliate leaves the group: its rights end, or it may apply for its own licence on the same terms. Without that rule, a spun-out entity keeps a licence it should not have.

Check the direction of the grant as well. A licence granted to the parent usually extends down to subsidiaries, but a licence granted to a subsidiary often does not extend up, and teams discover this after an acquisition rather than before it.

Inside a product, and the flow-down that controls it

Distributing a trained model is the case most teams actually need, and it is not the same as sublicensing the data. A model that has learned from a corpus does not contain the corpus, and a clause permitting sublicensing of the data may still say nothing about distributing a model trained on it.

Write the product permission separately and be specific about the form: a hosted API, an on-device build, a downloadable artefact, or a model made available to end users under their own terms. Each raises a different question about whether the end user receives anything derived from the licensed material.

If the data itself rather than a model is embedded in your product — a lexicon, a transcript set, an evaluation suite shipped to customers — that is redistribution of the data and needs its own grant, with its own limits on which customers and in what form.

Every permission to share should carry an obligation: the recipient is bound by the same restrictions as the licensee, in writing, before it receives anything. That one sentence keeps a sublicence from becoming an uncontrolled copy, and it is the first thing a vendor's counsel will look for.

Flow-down has limits worth naming. You cannot flow down a restriction you have not been told about, so ask for the vendor's list of prohibited uses as a schedule. And flow-down is hard to enforce across a large group, so pair it with a register: a written list of recipients, kept current, available to the vendor on request. A register is cheaper than an audit and it catches the real problem, which is material reaching an entity nobody intended.

Resale, and the honest position

Resale is refused for a simple reason: a buyer who can resell becomes a competitor. If you need it, the conversation is not about drafting but about structure, and three shapes exist. A reseller appointment, where you sell under the vendor's terms and they take a share. A value-added product, where you sell something materially different that incorporates the data, with the difference defined in the contract. Or a full assignment of a specific corpus, where you buy the material outright and the vendor exits that market.

If none of those is available, the practical route is to accept a no-resale clause and then check that it cannot reach your actual business. That means reading the definition carefully: a broad definition of resale can capture a product whose main feature is derived from the data, which is exactly what most buyers are building.

Terms to check

Work through these before the grant is finalised, because several of them are cheap to fix at signature and expensive to fix later.

• Whether affiliates are covered, how control is defined, and how the list is updated.
• Whether the licence runs down from a parent and up from a subsidiary.
• Whether a departing affiliate keeps its rights.
• Flow-down obligations, and the vendor's schedule of prohibited uses.
• A recipient register, and how often it is refreshed.
• Distribution of trained models, written as its own permission.
• Whether end users receive anything derived from the data, and on what terms.
• The definition of resale, tested against your actual product.
• What happens to sublicences if your own licence terminates.

One note on scope

Sublicensing clauses are usually drafted broadly because they are copied from software licences, where the asset is a single artefact rather than a corpus that will be trained on. Read every sharing clause against what your product actually does with the data, not against the category it was written for.

This is a practical checklist rather than legal advice. Sub-licensing and resale restrictions are enforced strictly, and the wording should be reviewed by counsel in the relevant jurisdiction.