AI data handling - Casa Conect Docs

Using AI on client documents raises fair questions. Which companies see the content? Where is it processed? Is it kept, or used to train models? This page answers them as precisely as we can.

The short version

The agent’s Claude models run on Amazon Bedrock in EU regions.
The agent’s ChatGPT models run on the OpenAI Platform, and every request is sent with response storage switched off.
A few supporting services handle search, document parsing, dictation, web search and cloud browsers. Each is listed below with what it receives.
The default model, Auto, is a ChatGPT model. An admin who wants conversation inference to stay in the EU can set a model policy, with the caveats described there.
Casa Conect does not use your content to train models.

Conversation models

Models	Provider	Where it runs	Notes
Opus 5, Sonnet 5	Anthropic, via Amazon Bedrock	EU regions, through Bedrock’s EU inference profile	Requests are routed among EU regions only.
ChatGPT 6 Astra, 5.6 Sol, 5.6 Terra, 5.6 Luna	OpenAI Platform	OpenAI’s infrastructure. The region is not pinned by Casa Conect.	Every request is sent with `store: false`, so OpenAI does not retain it as a stored response.

On EU processing. Choosing a Claude model keeps the conversation model’s processing in the EU, which is why Claude usage carries a 10% regional premium from the provider. It does not by itself make every AI operation EU-only. See Supporting services, and Models that work behind the scenes.

What is sent to a conversation model

When the agent works, the model receives:

your message, and the conversation history that is still in view
the agent’s instructions, your guidance, and any skill in use
the results of the tools it calls: the passages a search returned, the section of a document it opened, the emails in a thread

It does not receive your whole vault. The agent retrieves what a task needs, step by step, and each retrieval is authorized as you. Casa Conect sends no account identifiers, payment details or credentials to model providers.

Supporting services

Purpose	Service	What it receives
Document parsing	LlamaParse, on its EU endpoint	The uploaded file, to extract text, tables and layout
Search embeddings	Azure OpenAI	Text passages from your documents, turned into numerical vectors. The vectors are stored in our own database.
Search query understanding	Azure OpenAI	Your search query, to pick out names, identifiers and key terms
Search reranking	Cohere	Your query and the candidate passages, to order them by relevance
Conversation memory	OpenAI	Short passages from your conversations, embedded so that they can be recalled later
Voice dictation	OpenAI	The audio you dictate, returned as text
Translation in Shared Spaces	DeepL	The short texts people enter in a Shared Space, such as names, descriptions and review comments, translated between English and Spanish. Not documents.
Web search	Tavily	The search query the agent writes. No documents.
Cloud browsers	Browserbase	The pages the agent visits. The provider’s own model interprets page layout to find buttons and fields.
Prompt management	Braintrust	The agent’s system instructions, which we author there. Not your content.

Each is listed, with its role, on the Subprocessors page.

Your controls

Model policies

Admins choose which conversation models the organization may use.

Permission levels

Decide how much the agent may do, and therefore read, without asking.

Budgets and limits

Cap AI spending for the organization and for each member.

Where you put documents

Documents are parsed and indexed when they are uploaded. Do not upload what must not be processed.

Retention

Data	Where it is kept	Removed when
Conversations and agent memory	Our database, in the EU	You delete the conversation, or the project or matter it belongs to
Search index: passages and vectors	Our database, in the EU	The document is deleted
Workflow run details	Short-lived storage	Automatically, after a few days
Agent run traces, used for debugging	Our database	Pruned automatically by age
Content sent to model providers	Subject to each provider’s terms	Casa Conect disables storage wherever the provider offers the control

What we have not yet formalized

We want this page to be relied on, so we are explicit about its limits. Casa Conect sets store: false on OpenAI requests and uses Bedrock’s EU profile for Claude, and both are enforced in code. We do not currently publish contractual zero-data-retention terms for every provider, and the processing region of the supporting services above is not pinned, except where stated. If your engagement requires specific contractual assurances, contact us before uploading the material in question.

​The short version

​Conversation models

​What is sent to a conversation model

​Supporting services

​Your controls

Model policies

Permission levels

Budgets and limits

Where you put documents

​Retention

​What we have not yet formalized

The short version

Conversation models

What is sent to a conversation model

Supporting services

Your controls

Retention

What we have not yet formalized