The ratio is the headline and the least useful part of the finding. Published figures for the same measure range from 45 to 1 up to 144 to 1 depending on who is counting, which tells you the industry has not agreed what a machine identity is, not that some enterprises have three times more of them than others.

The new part is the composition. Seventy-nine of the 109 machine identities per human are AI agents, and the same survey expects agents to grow faster than any other identity class over the next twelve months. Last year's edition, published by CyberArk, put the overall ratio at 82 to 1.

The number worth acting on is the pair underneath it. On average, 40% of AI agents and 40% of machine identities already reach organizational data. Yet only 15% of respondents define a privileged identity as any machine identity with access to sensitive data. The 2025 edition found the same gap: 88% of respondents then applied the word privileged only to humans.

That is a governance failure before it is a tooling failure, and it explains why buying a machine identity product first usually disappoints. The product will enumerate thousands of identities and hand back a list nobody owns. Assigning ownership is the prerequisite, and it is unglamorous policy work that no vendor can do for you.