Personal data in a message
The model that answers your agent does not run on your machine. Whatever you write to it, leaves — and what you write is up to you.
What goes out, and what stays
Your message goes to the model as you wrote it. Your company's data does not: the agent reads a record when it needs one, and only the part it reads travels with the question.
The conversation itself is stored in your own database, like everything else in kyte.
The automatic marking is off
kyte used to look at your draft while you typed and mark what looked like personal data — names, addresses, account numbers — so you could replace each one before sending.
It is switched off, because it was not reliable enough: it missed things and marked things that were not personal data. A privacy mark that is sometimes right is worse than none, because it suggests a check that is not happening. It comes back when it can be believed.
Until then: what you type is what travels. If a name has no business leaving the building, write the customer instead, and let the agent look the record up for itself — the record stays here.
The bigger question
Where the model runs at all is a separate decision — a workspace can use a model on your own network, and then nothing leaves the building. See which model answers.