Rights and contracts

Does 'internal business purposes' cover AI training?

By SourceX Editorial · Reviewed by Noah Loul · October 9, 2026

Short answer

Whether internal business purposes covers AI training depends on the contract's definitions and context, and the answer is often unclear. Training a model used only inside the licensee's own operations fits the phrase more easily than training a model sold or offered to others. Because ambiguous scope invites disputes, say expressly whether training, fine-tuning and evaluation are permitted.

Key takeaways

The phrase has no single legal meaning; it is read in the context of the whole contract.
Training a model that serves only the licensee's own staff fits internal use more easily than training a commercial model.
The phrase appears on both sides of a company's contracts: in licenses it grants and in vendor and customer contracts it signs.
A licensor that wants to exclude training should name training, fine-tuning, embeddings and evaluation expressly.
A licensee that needs training rights should name them, along with what happens to models when the license ends.

What does 'internal business purposes' usually mean?#

Internal business purposes usually means use by the licensee for its own operations, as opposed to distributing, sublicensing or using the material to serve third parties. The phrase is a scope limit: it tells the licensee what it may do and implies that other uses are not granted.

The phrase was common in software, data and content licenses long before generative AI. Contracts drafted then were aimed at reporting, analysis and running the licensee's business, and their authors rarely pictured a model that absorbs the licensed material and then answers questions for thousands of outside users. That gap is why the same words now produce opposite readings.

Factors that decide whether training is covered#

The factors that decide coverage come from the rest of the contract and the circumstances of signing, not from the phrase alone. A court or counterparty will look at what the model does, who uses it and what the other clauses permit or prohibit.

Factors that decide whether training is covered
Factor	Points toward training being covered	Points toward training not being covered
Who uses the model	Only the licensee's employees	Customers, partners or the public
What the model produces	Internal routing, search or reports	A product or service sold to others
Definition of use	Broad, including analysis and derivative works	Limited to viewing, reporting or named purposes
Derivative works clause	Licensee may create derivatives	Derivatives prohibited or owned by the licensor
Competing product restriction	None	Licensee may not build competing products or benchmarks
Return and deletion duty	Silent or limited to raw copies	Requires deletion a trained model cannot honor
Context of signing	AI use discussed or contemplated	Signed before AI use was considered, with no mention

Where the phrase shows up in your own contracts#

The phrase shows up in three places that matter to a company considering a data license, and each one cuts a different way. Review all three before signing anything, because a restriction in one can undo a permission in another.

Vendor terms show how much the wording varies. Linear's Terms of Service license customer submissions for the sole purpose of providing the service. Section 8.4 of one version of the GitLab Subscription Agreement gives GitLab a limited license necessary to provide the software and for its development and improvement, which is broader. Read the version your company actually signed, because terms change between versions.

Licenses you grant: if your data license limits the buyer to internal business purposes without more, the buyer may read it as permitting training and you may read it as forbidding it.
Vendor contracts you accept: SaaS terms often let the vendor use customer data for its internal business purposes, such as improving its services, which may be argued to include training the vendor's own models on your records.
Customer contracts you rely on: where customers let you use their data only for your internal business purposes, licensing that data to an AI developer is unlikely to fit, so those records usually need consent or exclusion.

Why the model's audience matters more than the training step#

The model's audience matters more than the training step because internal business purposes is a limit on who benefits from the licensed material. A support-routing model used only by the licensee's own agents keeps the benefit inside the business. A general assistant sold to the public carries the licensed material's influence to every user, which looks much more like external commercial use.

Between those poles sit harder cases. A model used internally to draft responses that customers then read, or a model trained by one subsidiary and shared across a group, can be argued either way. Map where the model's outputs go before deciding which side of the line a use falls on, and record that map with the contract file so the reasoning survives staff changes.

The same logic applies in reverse when a company asks whether a vendor's internal-use right covers training. A vendor model that improves only your own instance is closer to internal use than a shared model offered to every customer of the platform.

Drafting fixes for licensors#

A licensor's drafting fix is to stop relying on the phrase and to name the AI uses it allows or forbids. Spell out each technique, because buyers may argue that a ban on training does not cover fine-tuning an existing model, building embeddings for retrieval or running evaluations.

Restrictive form, for discussion with counsel: Licensee shall not use the Licensed Data to train, fine-tune, evaluate or create embeddings for any machine learning model, or to provide retrieval to any model, except as expressly permitted in an Order. Permissive form: Licensee may use the Licensed Data to train, fine-tune and evaluate machine learning models, including models made available to third parties, provided no model is designed or prompted to reproduce Licensed Data and all other restrictions in this Agreement continue to apply.

Drafting fixes for licensees#

A licensee's drafting fix is the mirror image: name the training rights it needs and settle what happens to the resulting models. A company receiving data, whether a model developer or an operating business licensing third-party data for its own AI tools, should not rely on internal business purposes to carry training rights.

Name the techniques, say whether models may be offered to customers or only used internally, and state that trained models survive termination if that is the deal. If affiliates will use the models, include them in the grant. If the licensor insists on deletion at term end, agree in writing that the duty covers copies and embeddings but not trained weights, or price the risk accordingly.

Illustrative example: a 3PL reads its contracts both ways#

Illustrative: a fictional regional third-party logistics company runs its warehouses on a WMS whose terms let the vendor use customer data for its internal business purposes, including improving its services. The 3PL's client contracts let it use client order data only for its internal business purposes. It wants to stop the vendor from training on its data and to license its own exception-handling records to an AI developer.

At renewal, the 3PL negotiates an express no-training clause with the WMS vendor. For the license, counsel concludes that the client contracts do not stretch to third-party licensing, so client order data is excluded. The 3PL licenses its own exception notes, root-cause categories and resolution workflows with client names and identifiers removed, and the license names exactly which AI uses the developer may make.

How SourceX sets permitted use#

SourceX states permitted use expressly in every license rather than relying on general phrases, and the supplier approves that scope in the Approval step of the SourceX five-step transaction. During the Rights step, customer and vendor contracts are read for internal-use language that could limit what the supplier may license.

The permitted use is recorded in the SourceX Evidence Packet together with provenance, licensing rights, the privacy record and release authorization, so neither side has to reconstruct later what the words were meant to cover.

Frequently asked questions

Is 'internal use only' the same as 'internal business purposes'?

They are close, but internal use only is sometimes read more narrowly, as use inside the organization without any output reaching others. Both depend on the surrounding definitions. Neither phrase reliably answers whether AI training is allowed, so name the uses expressly.

Is fine-tuning treated differently from training?

Contracts rarely distinguish them unless someone drafts the difference in. A licensee may argue that fine-tuning an existing model is not training a new one, and a licensor may argue the opposite. List training, fine-tuning, evaluation, embeddings and retrieval separately to avoid that debate.

Does building a retrieval index count as use for training?

Not necessarily. A retrieval system stores and serves the records to a model at query time rather than training on them, but it may still exceed an internal-use limit if outside users see the results. Address retrieval as its own permitted or prohibited use.

What if the contract was signed before AI use was common?

Interpretation then turns heavily on context and the other clauses, which makes the outcome harder to predict. The practical fix is an amendment that states the AI position in plain terms, ideally negotiated before any training begins rather than after a dispute arises.

Do affiliates fall within internal business purposes?

Only if the contract says so. Some licenses define the licensee to include affiliates; others limit use to the signing entity. If a group company will run the model or use its outputs, include affiliates expressly and define them as of the signing date.

Sources

Linear's Terms of Service grant Linear a license to use, process and display User Submissions for the sole purpose of providing the Service to Customer. Source
Section 8.4 of the GitLab Subscription Agreement (version GLSA_01.12.26 v8) states the customer retains all right, title and interest in Customer Content, subject to a limited license to GitLab necessary for its provision of the Software and its development and improvement. Source

Related resources

See if your company qualifies

A short company assessment. No data uploads are needed.

Build your deal terms checklist Talk to SourceX