In 30 seconds

An AI workflow that touches customer data is still your responsibility, even when a vendor runs the model. The good news is that the protections are mostly ordinary data-handling done deliberately: send less, choose a vendor who does not train on it, limit who sees the results, and be able to explain what happens to it. None of this requires deep technical skill, just intent.

Step 1: Send the least you can

The strongest protection is the data you never send. Give the agent only the fields the task needs, and keep the rest out. Where the real name or number is not needed, replace it with an ID or a placeholder.

Step 2: Choose the vendor deliberately

Where the data goes matters as much as what you send. Confirm that the vendor does not train on your data, that they store it in a place you are comfortable with, and that you can have it deleted. A business-tier setting usually covers this; the default consumer setting may not.

Step 3: Limit who sees the output

The agent's results can contain customer data too. Decide who can read them and where they are stored. A draft reply sitting in a shared inbox everyone can open is a data exposure, even if the model was careful.

Step 4: Keep a record you could explain

If a customer or a regulator asked "what does your AI do with my data?", you should be able to answer plainly: what it reads, where it goes, how long it is kept, and who can see it. Writing this into the workflow's specification means the answer exists before anyone asks.

Step 5: Plan for the mistake

Decide in advance what happens if the agent mishandles data: exposes it, sends it to the wrong person, or stores something it should not. Knowing who is told and what is done turns a scramble into a procedure.

What this does not solve

These steps reduce risk; they do not remove your legal duties, which depend on where you operate and the data involved. For anything involving special-category or high-volume personal data, get specific advice. This guide keeps the everyday cases safe, not the edge legal questions.

Good enough to move on when

  • The agent receives only the customer data the task needs.
  • The vendor's data terms are known, in writing, and acceptable.
  • You can explain in plain words what happens to the data, end to end.