Most enterprise security stacks were built to inspect packets, flows, files and sessions. The material that now leaves the business does not look like any of those. It is a paragraph of contract language, a customer list, a set of forecast numbers or a block of source code, pasted into a browser window and sent as encrypted traffic to a domain that looks entirely legitimate.

Ask most organizations how many AI tools their staff use and the honest answer is an estimate. Ask what left the business in a prompt last week and there is usually no answer at all. The tooling that would catch the same data leaving as an email attachment does not see it.

This roundtable brought together security and infrastructure leaders from across UK enterprise to compare what they can actually see, what they have chosen to block, and where blocking has cost more than it protected.

The discussion covered the visibility gap between sanctioned and unsanctioned AI use, who holds the authority to say no when a business unit has already bought the tool, and what a workable policy looks like when the alternative is people using personal accounts on personal devices.