Not Won't. Can't.: Why Your AI Strategy Needs a Real Kill Switch

35% of companies can't shut off a rogue AI agent once it's turned on. Not "won't." Can't.

A separate study of 225 security leaders found the same problem, described a different way: 60% have no way to stop an agent that starts misbehaving. 63% can't even limit what it's allowed to touch in the first place.

When you sit in steering committee meetings, almost the whole meeting goes to deciding whether to approve something. Almost none of it goes to asking whether we could undo it once it's running.

Approval Power vs. Execution Control

PMOs have had this same problem for years, just slower. A checkpoint that recommends canceling a project isn't the same as one with the real power—and the access—to pull the funding and shut it down that same afternoon.

Every factory line has a cord any worker can pull to stop the belt. Nobody argues about whether that's necessary. It gets built in before the line ever runs.

If your riskiest AI agent started acting up at 2am, does someone know how to stop it—or just a name to call who might?