"Does the AI see my data?" deserves a more exact answer than yes or no. On a real organization, the first answer is no: nothing goes to a model until an owner turns it on in Settings. After that, it depends on the job.
Organizing an upload
To propose how your spreadsheets connect, the model sees column headings, counts and random stand-ins for values. It works out that a column looks like an order ID without seeing a real one.
Answering a question
It sees what each kind of record means, and the values of short categorical fields such as a shipment status: twelve or fewer distinct values of at most sixty characters, never names, IDs or free text. A field where every value is different counts as an identifier, whatever it is called. Questions that only need counting are answered with no model at all.
Running an agent
An agent has to read what it is working on, so here the model does see what your records and tools return. That is the honest cost of having it do the work, and it is why this needs the owner's decision first.
In every case
- Each answer must carry the AI service's proof of zero data retention, or it is thrown away.
- Requests ask that nothing is used for training.
- Each organization is charged for its own use, and a request over its plan is refused.
Search
What a run searches for goes to our search provider, so confidential details should stay out of those requests. Once a run has read your records or connected tools it can't search at all, and it can only open addresses it was already given. That stops a page from tricking an agent into sending your data somewhere.