How do you find out whether an AI vendor trains on your data?

The question people ask is whether the vendor trains on their data, and it is the least useful of the three questions bundled inside it. Training is the one vendors most readily disclaim because it is the cheapest to give up. Retention — how long inputs and outputs are stored and where — is a different commitment with different consequences, and it is what determines whether your material appears in a discovery request against the vendor or in a breach they suffer. Human review — whether staff or contractors can read your inputs, for abuse monitoring, quality or support — is a third, and it is the one most often left unmentioned.

The tier is the variable that decides the answer more than the vendor is. Nearly every provider offers different terms on consumer, business and enterprise plans, and the difference is usually exactly this: consumer tiers may use inputs for improvement by default, business tiers do not, and the enterprise tier adds contractual commitments about retention and jurisdiction. A business whose staff signed up individually is on the consumer terms whatever the business believes about the vendor, and this is the single most common gap between a policy and what is happening.

The document to read is the one that governs, and it is rarely the page that comes up first. In descending order of authority: the signed agreement or order form, the data processing terms it incorporates by reference, the product-specific terms for the feature in question, and only then the published policy pages. Marketing claims about privacy sit below all of these and bind nobody. A vendor that will not tell you which document governs has answered the question.

Features change the answer within a single plan, which is why a blanket answer about a vendor is usually wrong. Connectors that read a mailbox or a document store, an agent that can browse, a beta feature, a model from a third party offered inside the product, and anything labelled research preview commonly carry different terms from the core product. The specific question to ask is whether the commitment applies to every feature or to the base product, and the answer is frequently the second.

Subprocessors are the part that decides how many organisations the answer actually concerns. A vendor's commitment not to train binds the vendor; the list of subprocessors tells you who else receives the data and under what onward terms. That list is usually published because enterprise customers require it, and reading it is the fastest way to discover that a tool you assessed as one company's product routes through three.

Finally, verify rather than believe, because verification is cheap here and belief is not. The settings your account is actually running under are visible in the admin console, the retention period is usually stated per workspace, and whether staff are signed into a business account is answerable in an afternoon. A vendor's terms describe what is permitted; the console describes what you configured, and the gap between them is where the exposure sits.