93 percent: the number that devalues approval gates on their own
According to vendor telemetry, around 93% of all approval prompts for AI agents get confirmed without a real check. That matches a general pattern from security research: the more often a warning or confirmation prompt appears, the more it becomes a routine gesture instead of a deliberate decision. An approval gate that theoretically protects but gets reflexively clicked away in practice offers no real safety – only the feeling of it.
The second problem: over-privileged agents from the start
Approval fatigue is made worse by a structural problem: around 90% of deployed agents are over-privileged relative to their actual task – they're technically capable of more than their function requires. Combine over-privileged setup with an approval that's only formally confirmed anyway, and there's practically no effective control left. Around 88% of organisations reported at least one AI agent security incident in the same period – direct evidence that this combination actually goes wrong in practice.
Why step-by-step checking doesn't solve the problem
A single step an agent takes almost always looks harmless on its own – reading a file, setting a parameter, calling a tool. Only the chaining of several harmless individual steps adds up to a risky overall action. Anyone who only evaluates each step in isolation rarely sees the actual risk – which is exactly why leading providers now evaluate entire action chains instead of individual steps, to close precisely that gap.