Key insight

You cannot protect data appropriately until you know how sensitive it is. Classification sorts data into a few sensitivity levels; a sensitivity label attaches that decision to the data itself so it travels wherever the data goes — and an AI agent needs exactly that travelling signal to avoid treating a customer record like a public webpage.

A restaurant’s lunch menu and a patient’s medical history both live on computers, but they plainly deserve wildly different levels of care. You cannot decide how carefully to protect a piece of data until you first know how sensitive it actually is — and that plain observation is what data classification and sensitivity labels are built to solve. This article explains both, and why AI agents need them more than any system before.

1 · The simple truth underneath: know before you protect

Protecting all data at the maximum level is wasteful and slows everyone down; protecting it all at the minimum level is dangerous. The only sensible path is to match the protection to the sensitivity — which is impossible until sensitivity is actually known and written down. Everything else in this article follows from that one requirement.

2 · Classification: sorting data into sensitivity levels

Data classification is simply the act of sorting data into a small number of sensitivity levels, most commonly something like public, internal, confidential, and restricted, so that everyone knows how much care each piece needs. The number and names vary between organizations, but the idea is always the same: a handful of clearly defined buckets, each with its own handling rules.

Four sensitivity levels from public to restricted, each with increasing protection Four boxes from left to right labelled public, internal, confidential, and restricted, shaded from calm green through amber to red, illustrating increasing sensitivity and increasing required care. Public Internal Confidential Restricted least care most care
Figure 1. A few clearly defined buckets, each with its own handling rules — the exact number matters less than that everyone agrees which is which.

3 · Labels: the decision that travels with the data

Classification is a decision; a sensitivity label is what makes that decision stick to the data itself. Think of the ordinary handling labels on a shipping box — “fragile,” “this way up,” “refrigerate.” The label does not protect the contents by itself, but it tells everyone who handles the box how to treat it, and it stays on the box wherever the box travels. A sensitivity label does the same for a file: it travels with the data into email, into cloud storage, onto a laptop, rather than living only in a separate list that the data quickly outruns. Once the decision travels with the data, a rule can follow it automatically — for example, a file labelled “confidential” can simply refuse to be emailed outside the company, no matter where it ends up.

A labelled file carrying its confidential label unchanged into email, cloud storage, and a laptop A file box marked confidential connects with three arrows to email, cloud storage, and a laptop, each still showing the same confidential label, illustrating that the label travels with the data wherever it goes. Fileconfidential Emailconfidential Cloud storageconfidential Laptopconfidential
Figure 2. The classification decision rides along with the file — so a “confidential” rule can be enforced everywhere the file lands, not just where it started.

4 · A worked example: one labelled file, three destinations

A financial summary is classified “confidential” and given a matching sensitivity label the moment it is created. An employee later tries to attach it to an email addressed to an outside partner. Because the label travels with the file, the email system reads it, recognizes “confidential,” and blocks the send automatically — no human had to remember the rule in that moment. The same file, saved to internal cloud storage, is allowed, because the label’s rules permit internal storage. The protection followed the data, applied itself correctly in both situations, and required nobody to look anything up.

5 · Why AI agents need labels more than ever

An AI agent, software that reads, combines, and produces data at high speed, has no dependable way to know how sensitive something is unless it is told. Without a reliable label, an agent has no way to know that the customer records it just summarized are far more sensitive than the public webpage it summarized a moment earlier — both were just “text it read.” A clear, travelling label is exactly the signal an agent needs to handle sensitive data with appropriate care rather than treating everything it touches identically. An agent that respects labels can refuse to send confidential content to an outside tool; an agent with no labels to read has nothing to base that judgment on.

6 · Applying classification on purpose

7 · A simple test you can run this week

Try this before an incident forces the question

1. Pick one genuinely sensitive file you rely on.
2. Check whether it carries a sensitivity label that travels with it, or whether its sensitivity lives only in someone’s head.
3. Copy it somewhere new and see whether the label — and its rules — came along.
4. If the sensitivity did not travel, that is exactly the gap a label closes.

8 · Glossary — every short-form term, spelled out

Data classification
Sorting data into a small number of sensitivity levels so everyone knows how much care each piece needs.
Sensitivity level
One of the defined buckets — commonly public, internal, confidential, and restricted — each with its own handling rules.
Sensitivity label
A marking attached to the data itself that carries its classification decision wherever the data travels.
AI agent
Software that decides, on its own, which tools to call and which actions to take, and which reads and produces data at high speed.
Key takeaways

You cannot protect data appropriately until you know how sensitive it is.
Classification sorts data into a few clear sensitivity levels; a label attaches that decision to the data itself.
A travelling label lets a rule follow the data automatically, wherever it goes.
AI agents need labels more than ever, since without them an agent cannot tell a customer record from a public webpage.

References

  1. NIST Federal Information Processing Standards Publication 199, Standards for Security Categorization of Federal Information and Information Systems, National Institute of Standards and Technology. csrc.nist.gov