Where does an in-app AI assistant's data go? Eight questions to ask any vendor
Published 13 October 2026
Ask where conversations are stored, what the model sees, whether it trains on your data, who else processes it, how long it is kept, who on the vendor's side can read it, how changes are confirmed, and what the vendor does not have yet. Rendel stores data in Frankfurt, masks what users type before the model sees it, and lists every sub-processor on its security page.
The eight questions, and Rendel's answers
Each answer below is what Rendel does today; the security page keeps the full list.
| Question | Rendel's answer |
|---|---|
| Where are conversations stored? | In the EU, in Frankfurt (eu-central-1). |
| What does the model see? | Your actions' names, descriptions and parameter types, the app state you pass, and the conversation, with emails, phone numbers and card numbers masked first. |
| Is our data used to train models? | No. |
| Who else processes it? | Every sub-processor is listed on the security page, with what it sees and where. |
| How long is it kept? | As long as you set, per app; you can also erase one of your users yourself. |
| Who at the vendor can read transcripts? | Only staff on an explicit allowlist, and every view is in your own audit log. |
| How are changes confirmed? | Every change waits for the user's tap on a card the server builds; the server records the answer. |
| What don't you have yet? | No SOC 2 report or ISO certificate, no penetration test report, no console SSO, and no data stored in Turkey today. |
Why the last question matters most
A vendor that answers every question with “yes” is hard to trust. Asking what is missing shows how the vendor talks about risk, and gives your review the gaps to plan around.
For teams in Turkey
Rendel stores its data in the EU today. After an assessment, a Turkey region or your own servers, by separate enterprise agreement.
Questions
Is the model provider in the EU too?
No. Conversations are stored in Frankfurt; the model runs at Anthropic in the United States and is not trained on your traffic. The sub-processor list says so.
Can we see what the assistant did for a given user?
Yes. Every action is on the record in the console with when, the parameters, who confirmed it and what came back, and owners can export it.