> For the complete documentation index, see [llms.txt](https://docs.toucanai.cloud/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.toucanai.cloud/govern/security-model/ai-assistant-data-handling.md).

# AI assistant data handling

{% hint style="info" %}
**Target Audience**: Developers building AI chat or embed experiences with Toucan AI.
{% endhint %}

### TL;DR

* The AI assistant **processes** your questions, database metadata, and query results to answer and build visualizations.
* **Conversation history is stored** on the platform so users can continue a thread.
* **Sub-agents** (query building, chart building, exploration) run in isolated steps; only the main assistant thread keeps long-lived history.
* **Third-party AI providers** may receive prompts and context when AI features run — see [Third-party subprocessors](/govern/security-model/third-party-subprocessors.md).

***

### When to use this

Use this page when assessing privacy impact of AI chat, what may be retained after a session, and what may be sent to external AI services.

***

### What the assistant processes

During a conversation, the assistant may use:

* **User messages** and assistant replies
* **Database metadata** (table and column names, descriptions) to understand your data model
* **Query results** from your connected database (for example previews or aggregated answers)
* **Optional context** from your embed integration (non-identifying clues you provide)

Query results used in AI workflows are fetched **on demand** from your database; they are not bulk-downloaded into a Toucan data store. See [Data storage & retention](/govern/security-model/data-storage-and-retention.md).

***

### What the assistant stores

| Item                                    | Stored? | Notes                                           |
| --------------------------------------- | ------- | ----------------------------------------------- |
| Conversation messages                   | Yes     | Per user/thread; can be cleared in product      |
| Structured “plans” for complex tasks    | Yes     | Linked to the conversation thread               |
| Full copy of your database              | No      | Queries run against your DB when needed         |
| Query result rows (platform-wide cache) | No      | Results may appear inside a stored conversation |

Clearing a conversation removes that thread’s stored assistant history.

***

### Third-party processing

Depending on configuration, AI requests may be processed by:

* An **LLM provider** (by default Anthropic, or another provider configured for your deployment)
* **Analytics or tracing tools** (for example product analytics or execution tracing)

***

### Best practices

* Do not include personal data in optional AI context fields.
* Use row-level security so the assistant only queries data the end user is allowed to see.
* Clear conversation history when your product requires it (for example on logout or account closure workflows you implement).
* Review subprocessors if your compliance program requires a register of external processors.