In 30 seconds

Not all data belongs in an AI workflow. Some of it is too sensitive, some you are not allowed to share, and some simply is not needed for the task. The safe habit is to send the agent the least it needs, and to keep a short list of things that never go in. Decide this before you build, not after an incident.

The default rule

Keep these out unless you have a specific, cleared reason

Secrets and credentials

Passwords, API keys, access tokens, and anything that grants access to a system. These should never be in the text an agent reads. If an agent needs to act on a system, it is given scoped access through the surrounding setup, not by being handed a password.

Payment and financial identifiers

Full card numbers, bank details, and similar. These are governed by strict rules and have no place in a general AI prompt. Reference an order or account by its ID instead.

Special-category personal data

Health, ethnicity, religion, sexual orientation, biometric data, and similar. This carries the highest legal protection. Do not put it into an AI workflow without specific advice that you are allowed to.

Anything you promised to keep private

If you told a customer, an employee, or a partner that their information would be handled a certain way, sending it to a third-party AI may break that promise, even if the vendor is careful. Your commitments travel with the data.

The grey area: ordinary personal data

Names, emails, and addresses often do need to be in a workflow. That can be fine, but it depends on the vendor's data practices. Check first whether the vendor trains on your data, and prefer a setting where they do not retain it.

A useful technique: mask what you can

Often the task does not need the real identifier at all. An agent checking whether a message is a complaint does not need the customer's name. Where you can, replace sensitive values with a placeholder or an ID before the text reaches the AI.

What this does not solve

Keeping the wrong data out is one layer. You still need to protect the data you do send, in how the workflow is built and who can see the results. Pair this with how to keep customer data safe with AI.

Good enough to move on when

  • You have a short, explicit list of data types that never enter this workflow.
  • The agent receives only what the task genuinely needs.
  • Where the real identifier is not needed, it is masked or replaced with an ID.