A training-data clause is the contractual provision that governs whether a vendor can use customer data, prompts, or outputs to train its AI models. It became one of the highest-profile clauses in 2023 when several major SaaS vendors quietly updated their terms to permit broader training use, and customers (and procurement teams) responded with detailed restrictions.
The clean modern pattern carves out three categories. Customer-supplied data fed into the system (typically prohibited from being used for training without explicit, separate consent). Customer prompts (typically prohibited or default-prohibited with an opt-in for non-confidential cases). Customer outputs (typically prohibited from being used for training, with the same caveats). Each category gets a separate provision because the privacy and confidentiality stakes differ.
Enterprise customers have largely won this fight. The default in 2025 enterprise AI vendor agreements is that customer data does not flow to model training unless the customer specifically opts in. Consumer-tier and free-tier products often default the other way, which is why employees using a personal ChatGPT account for client work create a real risk for their employer that an enterprise license would have prevented.
The drafting trap on the customer side is accepting "we may use de-identified or aggregated data for model improvement." That phrase is much broader than it sounds. De-identification can be reversed at scale; aggregation can leak information about specific customers. Customers in sensitive industries (healthcare, finance, legal) typically negotiate that out and replace it with a hard prohibition on any training use of their data, full stop.