Every security team that has looked has found AI use nobody registered: a browser tab on a personal account, an editor with a bundled model, a tool a developer wired into an agent last Thursday. The interesting question is not how much of it there is. It is which part of the problem is worth paying for.
Showing which AI services your users touched will be a checkbox inside every endpoint suite within a year. Several tools already do it for a few dollars per user. Treat a vendor selling you discovery alone as selling you next year's free feature.
A number that only ever goes up is a marketing metric. A coverage report worth having names its own blind spots: traffic that never reached you, tools that cannot be redirected at all, editors whose bundled models go to their vendor regardless of configuration, a browser tab on a personal account.
Ours does that, and says in as many words that nothing in it is evidence your traffic is all of it — only that it is what reached us.
Point your fleet at a proxy and you learn where AI traffic goes. Route an application through a gateway and you can act on it and prove what happened. Both are worth doing; only one of them survives a dispute.