Your vendor's subprocessor list already names the AI they use. Go read it.
Procurement keeps asking vendors which AI they run on. The answer is already published: the subprocessor list. What Atlassian's page discloses, and the quarterly habit that makes it governance.
A client's procurement team spent most of July chasing vendors with a new questionnaire item: which AI providers do your features use. Three vendors answered vaguely. One answered wrong. And for at least one of them, the correct, legally maintained answer had been sitting on a public page the whole time.
Every SaaS vendor that processes personal data under a DPA publishes a subprocessor list. It is a legal commitment, not marketing, and when the vendor added AI features, the AI went on the list.
Take Atlassian's page, effective May 2026. Under AI and analytics providers it names OpenAI, "generative AI services provider for intelligence product features," for every cloud product with Atlassian Intelligence, Loom AI or Rovo enabled. Beside it: AWS Bedrock and Google Vertex AI for the same feature set, and Databricks as "infrastructure provider for machine learning development, processing and training." One vendor, four AI subprocessors, each with a stated purpose and a location of processing.
That paragraph answers most of what the questionnaire was trying to ask, with a page procurement never opened.
(The vague vendor answers were not evasion, mostly. The sales engineer genuinely did not know. The legal team that maintains the subprocessor list did.)
So the habit, which costs an hour a quarter.
Pull the subprocessor page for each of your top ten vendors and file the AI entries next to the vendor's file. You now know whose models sit behind which features, which is the inventory half of the shadow AI problem solved from the supply side.
Diff it quarterly. New AI names appear on these lists before they appear in release notes, because the DPA obliges the vendor to disclose before processing starts. A calendar entry and a saved copy is the whole tooling.
And match it against your own approvals. If your AI policy blesses one model provider and your project tool quietly runs on three others, that is not necessarily a problem, but it is a fact, and it belongs in the risk register rather than in nobody's head.
The questionnaire still has its place; ask vendors what the list means, not what is on it. Which features route to which subprocessor, whether the AI entries can be disabled per product, and what the notice period for list changes is. Those answers are not on the page. Everything else was.