Data classification used to be slow work with a slow payoff. A human had to go looking for a file before a missing label cost anyone anything. An AI agent doesn't go looking. It reads everything it's allowed to read, in seconds, and answers questions from whatever it found.
Why it counts more with agents
Labels are what a permission system reads to decide who sees what
An unlabeled file usually defaults to the widest setting, not the tightest
Obscurity used to protect files nobody could find, and agents end that
One answer can pull from many files at once, so one bad label leaks into many replies
Where teams get it wrong
The classification program is treated as a project that finishes, so the pilot ships first and the labeling catches up later. It doesn't catch up. The agent is live against unlabeled data the whole time, and every question a user asks it is a test of labels nobody has written yet.
Here's a concrete one. A company turns on an assistant across its file shares. Someone asks it what the new comp bands are. It finds an HR planning sheet in a folder that was shared with the whole company in 2019, because back then nobody could find it and nobody ever did. The assistant found it in under a second and quoted it.
Classify the places the agent can reach before you let it reach them. That's a smaller job than classifying everything, and it's the only part that's urgent.