> ## Documentation Index
> Fetch the complete documentation index at: https://help.casaconect.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Model policies

> How organization admins control which AI models the team can use in the agent.

A model policy is an organization-wide switch for a single model. Admins use policies to keep spending predictable, to meet a client's requirements about where data is processed, or simply to give the team a shorter list to choose from.

## Set a policy

<Steps>
  <Step title="Open the Models card">
    Go to **Settings → Access & billing** and find the **Models** card. Every selectable model is listed with its maker: *Made by Anthropic* or *Made by OpenAI*.
  </Step>

  <Step title="Switch models on or off">
    Each model has a switch showing **Allowed** or **Blocked**. Every model is allowed until an admin blocks it. Changes apply immediately to the whole organization.
  </Step>
</Steps>

<Note>
  Any member with access to billing settings can see the list. Only organization admins can change it.
</Note>

## What a blocked model does

Policies are enforced on the server, when a message is sent, against the model that will actually run.

* A member who sends a message on a blocked model sees: *"This model is not enabled for your organization. Pick another model, or ask an admin to enable it under Settings → Billing."* Nothing is sent to the provider and no credits are used.
* **Auto** is checked the same way. Auto resolves to the default model, ChatGPT 5.6 Luna, so blocking Luna also blocks Auto. If you block Luna, tell your team to turn Auto off and pick an allowed model.
* The check happens after the subscription check and before the credit check, so a blocked model is refused even when the organization has plenty of credits.

## What policies do not cover

Policies apply to the model a member picks for an agent conversation. They do not change the fixed models Casa Conect uses for background work such as naming conversations, applying matter templates, shared space summaries, and the steps inside the [Populate](/workflows/populate) and [Anonymize](/workflows/anonymize) workflows. Search, document parsing and dictation also use their own dedicated models.

If your organization needs a guarantee about every model that touches your data, read [AI data handling](/security/ai-data-handling) and [contact us](mailto:support@casaconect.ai) to discuss your requirements.

## Common set-ups

<AccordionGroup>
  <Accordion title="Keep conversation inference in the EU" icon="map-pin">
    Allow **Opus 5** and **Sonnet 5** and block the four ChatGPT models. Because Auto resolves to a ChatGPT model, ask members to turn Auto off and select a Claude model. Their choice is remembered in their browser.
  </Accordion>

  <Accordion title="Cap the cost of a conversation" icon="piggy-bank">
    Block **ChatGPT 6 Astra** and **Opus 5**, leaving Sol as the strongest option. Combine this with an organization budget or per-member limits. See [AI credits](/admin/ai-credits).
  </Accordion>

  <Accordion title="Keep the choice simple" icon="list">
    Allow one fast model and one strong one, for example **ChatGPT 5.6 Luna** and **Opus 5**, and block the rest. Members then choose between Auto and a single alternative.
  </Accordion>
</AccordionGroup>

## Related

<CardGroup cols={2}>
  <Card title="Choosing a model" icon="scale" href="/models/choosing-a-model">
    Advice to share with your team.
  </Card>

  <Card title="Roles and permissions" icon="users" href="/admin/roles-and-permissions">
    Who counts as an organization admin.
  </Card>
</CardGroup>