Shadow AI is a privacy problem before it is a governance problem

The AI policy debate usually starts at the top: principles, committees, risk tiers. Meanwhile a support engineer is pasting ticket text into a model to draft a reply, and that is already a disclosure to a new recipient.

An experiment is still processing

A prototype that reads customer records has a purpose, a legal basis, a recipient and a retention period, whether or not anyone has written them down. The paperwork does not create the obligation; the call does.

Four questions that declare an agent

Ask these before the feature ships, not after:

What personal data reaches the model, and can it be reduced?

Which provider receives it, and where does inference run?

Is the output used to make a decision about a person?

Who reviews it, and what happens when it is wrong?

Declare it and move on

Teams stall because declaring feels like asking permission. It is the opposite: an agent with a written purpose, a named provider and a human reviewer is one you can defend and keep shipping. The undeclared one is the one that gets switched off during a security review.

Try it on your own site

Bring your AI features into the same registry.

Komploop keeps agents beside vendors, cookies and documents, with purpose, basis, provider and reviewer on the record.

Run the free assessment See the platform

Takeaways

Every LLM call on personal data is processing, prototype or not.

Minimise the input before arguing about the model.

Name the provider, the region and the human reviewer.

Declared agents survive security reviews; shadow ones do not.