Can a SaaS Vendor Train AI on My Data?
Short answer: It depends on the agreement, the product settings and the law that applies to the data. A no-training promise should identify the protected data and prohibited uses, cover downstream providers, and match how the service actually works. Data ownership alone does not answer the training question.
A customer asks your startup to remove every right to use its data for AI. Or your company is buying software whose terms permit broad use of information to improve the service. Before either side signs, turn that general language into a clear description of permitted processing, restricted uses and operational controls.
Start with the actual contract and product
Read the signed order form, master agreement, data processing addendum, AI terms and incorporated policies together. Check which document controls if they conflict and how online terms can change. Then verify the account tier, enabled features, settings and providers used for this particular deployment.
Provider terms illustrate why the details matter. Anthropic's Commercial Terms, effective June 17, 2025 and checked for this update, prohibit model training on Customer Content from the covered Services. That is a commitment under those terms, not a rule governing every product that uses Claude or every SaaS vendor. A reseller's agreement and its own handling of data still require review.
Separate training from other uses of data
Training or fine-tuning changes a model using a dataset. Retrieval augmented generation supplies information from a separate source when answering a question. Running the service can also involve logs, safety checks, support access, storage and evaluation. Describe these activities separately rather than treating all processing as training.
A no-training commitment does not, by itself, promise zero retention or no human access. Ask where prompts, uploaded files, outputs, embeddings and logs go; how long they remain; who can access them; and whether any are reused across customers. Customer-specific fine-tuning also needs a defined purpose and access boundary.
What should a no-training clause cover?
Define the protected information. Depending on the product, that may include customer files, prompts, outputs, personal information, confidential information and derived datasets. Review exceptions for feedback, telemetry, aggregated information and de-identified data so that they do not silently defeat the agreed restriction.
Define the restricted activities and any permitted exceptions. Address training, fine-tuning, evaluation and product improvement expressly, with a separate written approval process for any allowed use. Specify the dataset, purpose, model, authorized users, duration and consequences of withdrawal.
Carry the commitment through the service chain. Identify model providers and other subcontractors, confirm what protections the vendor can obtain, and address provider changes. Add appropriate retention, deletion, security and incident obligations. Ask separately what happens to a customer-specific model after termination; deleting source files is not a promise that prior model changes can be reversed.
If an enterprise customer sends your startup a ban
Begin with a data-flow review by the people who operate the product. Identify which processing is necessary to deliver the contracted feature and which reuse is optional. Compare that answer with the customer's proposed wording before promising that no data is ever stored, reviewed or used to improve anything.
A workable negotiation may preserve narrowly defined processing needed to serve that customer while prohibiting reuse to develop a shared model. Another deployment may require disabling a feature or changing a provider. Those are decisions to resolve with engineering and the customer, not exceptions to hide in a definition.
Check permission as well as ownership
A customer can own information while remaining subject to confidentiality obligations, privacy requirements or another party's rights. Review whether the proposed processing is consistent with those obligations. There is no single consent rule or contract sentence that resolves every dataset and jurisdiction.
The FTC has warned AI companies that privacy and confidentiality commitments matter, including promises about whether customer data is used to train models. Marketing statements, product settings and the signed documents should describe the same practices.
What to bring to an AI contract review
Bring the agreement and incorporated terms, the customer's proposed changes, a short description of the data involved, the providers and features used, and the business deadline. Avoid sending sensitive datasets just to explain the issue; a high-level data map is usually the better starting point.
Jacobs Counsel can help scope review and negotiation of customer-data rights, AI restrictions and the surrounding commercial terms. The next step is to identify the deal, the operational facts and the work needed before agreeing on an engagement.
Related deep-dives
Sources
Updated September 16, 2026. Content updated on that date; this is general information only and not legal advice for your specific situation. A free fit-and-intake call is for intake and scoping, not legal advice. Representation requires a written engagement agreement.