Can a SaaS company license customer data for AI training?

What counts as customer data in a SaaS contract?#

Customer data is whatever your contract defines it to be, so the definition is the first thing counsel reads. Most master subscription agreements define it as content that customers and their users submit to the service, and many extend it to outputs the product generates from that content.

The same agreement usually defines neighboring terms such as usage data, service data, telemetry or aggregated data, and gives the provider broader rights over those. Pull the click-through terms, order forms, the DPA, the privacy policy and every negotiated enterprise redline, because your largest accounts may have signed definitions that differ from the public version.

Watch the gap between aggregated data and records. An aggregated-data clause typically lets you keep statistics that identify no customer or person. AI training usually needs record-level content such as ticket text or workflow steps, which those clauses were rarely written to cover.

Which contract clauses decide the answer?#

The seven clauses below decide whether a SaaS company can license customer data, and they have to be read together. A permissive aggregated-data clause can be cancelled out by a strict confidentiality clause, and a DPA can override both for personal data.

The decisive words are often small. A license to use customer data solely to provide the service is, on its face, enough to rule out third-party training; Linear's published terms, for example, license user submissions for the sole purpose of providing the service. A clause that lets the provider use usage and aggregated data to improve its products reaches statistics, not record content. Some enterprise customers also add an explicit ban on using their data to train AI or machine learning models.

Read each clause against the actual use: an outside developer receiving copies of records to train its own model. That is a different purpose from improving your own product, and many clauses written before generative AI did not anticipate it.

• 1Ownership: who owns customer data and outputs, and whether the provider owns anything derived from them.
• 2License to the provider: whether the customer grants rights only to operate the service or something broader.
• 3Permitted use or service improvement: whether the provider may use data for its own products, and whether third parties are mentioned.
• 4Aggregated and de-identified data: how the contract defines them and what the provider may do with each.
• 5Confidentiality: whether customer data is confidential information and whether disclosure to third parties is barred.
• 6DPA processing terms: processing only on documented instructions, subprocessor notice and deletion at termination.
• 7Amendment and precedence: whether the provider may change terms unilaterally, and whether negotiated paper overrides online terms.

Why does the DPA carry so much weight?#

The data processing agreement usually decides the question for personal data, because it casts the SaaS provider as a processor or service provider acting on the customer's behalf. Licensing that data for a buyer's training would be the provider's own purpose, not the customer's.

The GDPR makes the consequence explicit. Article 28 says that a processor which infringes the regulation by determining the purposes and means of processing is treated as a controller for that processing, with the duties that follow. US state privacy laws such as California's CCPA take a similar line, requiring service provider contracts to bar use of personal information outside the business purposes the contract names.

Which laws apply depends on where customers and their end users are, and sector rules may add more. These questions are assessed deal by deal with counsel, not settled by a single policy.

Can you update your terms to allow AI training?#

Updating terms to permit AI training is possible in principle, but a quiet or retroactive change is the riskiest route. In February 2024, FTC staff warned that a company which adopts more permissive data practices, such as using consumers' data for AI training, and tells people only through a surreptitious, retroactive change to its terms or privacy policy may be engaging in unfair or deceptive practices. That was staff guidance focused on consumers, not a rule, but it shows how regulators may read a quiet change.

Customers react as well as regulators. In August 2023, after backlash over changes to its terms earlier that year, Zoom added a sentence saying it would not use audio, video or chat customer content to train its AI models without consent. Enterprise customers add a further constraint: negotiated agreements usually override online terms, and security questionnaires may ask directly whether customer data trains any model. An opt-in amendment that describes the use, the recipient, how personal details are removed and how to withdraw is cleaner, but it takes time and can unsettle renewals.

Even an accepted change may reach only data collected after customers agree, not the archive behind it. A common mistake is treating a new clause as permission over years of older records gathered under narrower promises, so counsel should map each record family to the terms in force when it was collected.

A decision path for each record family#

A record-by-record decision path turns the contract review into scope decisions the CEO and counsel can sign. Apply the rules below to each record family and write down the outcome, including what was excluded and why.

• 1If the contract limits customer data to providing the service, exclude that category unless customers opt in to the specific use.
• 2If a customer's negotiated paper bars AI or machine learning use, exclude every record that holds that customer's material, including tickets and chat.
• 3If records are company-written but quote customer content, include them only after the quoted content and identifiers are removed.
• 4If a usage-data clause covers aggregated statistics, do not stretch it to raw records, text or files.
• 5If records came from churned customers, check the termination clause first, because deletion duties may already apply.
• 6If personal data is involved, confirm your processor or service provider role and the applicable privacy laws before anything else.

What can a SaaS company usually license instead?#

Company-owned operational records are what most SaaS providers can realistically license. Engineering history, product decisions, incident postmortems, internal documentation and resolved support cases describe how the company builds and runs software, which is what developers of coding and workflow agents look for.

Internal records still contain customer material, so they need cleaning before they qualify. Bug reports quote customer screenshots, support macros include account names and test fixtures sometimes hold copies of production rows.

Illustrative: a property management software vendor sorts its records#

Illustrative: a fictional vertical SaaS company sells property management software to landlords. Counsel finds three kinds of paper: click-through terms for small landlords, a standard enterprise MSA, and negotiated riders for the largest accounts. All three limit customer data, including tenant records, to providing the service, and two riders expressly ban AI training.

The CEO and counsel leave the product database out entirely. They scope Jira issues, GitHub pull requests and review threads, Confluence design notes and Zendesk ticket resolutions. Tickets from the two rider accounts are excluded, and the rest are cleaned of tenant names, addresses and landlord account details. The package is company-owned engineering and support history, with every carve-out and its contract reason written down.

How SourceX handles customer data questions#

SourceX treats customer data as a rights question answered before any file moves. In the Rights step of the SourceX five-step transaction, contracts, DPAs and notices are reviewed for each record family, and customer content the contracts do not clearly cover is carved out rather than argued around.

The initial fit check collects only metadata, such as systems, years of history and known restrictions. For whatever proceeds, the SourceX Evidence Packet sets out the licensing rights and permitted use behind the package, and the company approves the scope at every step.