> ## Documentation Index
> Fetch the complete documentation index at: https://docs.orbit.devotel.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Inbox AI privacy — keep message content off the LLM

> Control whether inbound message bodies and close-time summaries ship to a third-party LLM. Two default-on toggles for HIPAA and EU tenants, plus high-confidence spam handling.

# Inbox AI privacy

The inbox runs two automatic AI behaviours that ship conversation content to
a third-party LLM: **auto-categorize**, which classifies each inbound message
body into a conversation category (sales, support, spam), and
**auto-summarize**, which generates a summary when a conversation closes.
Both are ON by default on every workspace.

If your organization works with regulated content — PHI under HIPAA, or
EU-governance rules that restrict which processors touch message data — you
can turn either behaviour off at **Settings → Compliance → Inbox AI
privacy**. The gate is tenant-owned: it is your organization's choice, and no
platform mandate flips it for you.

<Note>
  These controls only govern the inbox categorize/summarize hops. Redaction
  of universal enforcement (opt-out phrasing, consent markers) and the
  outbound [DLP scanner](/compliance/dlp-scanner) stay active regardless.
</Note>

## What the gates control

Three settings are read by the inbound classify and close-summary paths:

| Setting           | Default | Behaviour when ON                                                                                                                                          |
| ----------------- | ------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `auto_categorize` | ON      | Each inbound message body is sent to the LLM to classify the conversation category.                                                                        |
| `auto_summarize`  | ON      | Closing a conversation sends its recent messages to the LLM to generate a summary.                                                                         |
| `spam_action`     | `none`  | When the classifier is confident a conversation is spam, applies the chosen terminal action: `none` (tag only, stays in the inbox), `archive`, or `close`. |

Disabling a gate stops its corresponding LLM call. Spam handling runs as part
of categorization, so when `auto_categorize` is off there is no
classification and `spam_action` has no input to act on — the setting still
saves, but nothing triggers it.

## The persisted organization settings

The page reads and writes two keys on your organization's settings blob, the
same keys the inbound classify and close-summary paths consult on every
event:

* `settings.inbox_auto_categorize.enabled`
* `settings.inbox_auto_summarize.enabled`

The spam terminal action lives at `settings.inbox_spam_policy.action`. The
keys persist on organization settings rather than per-user preference so the
decision is a posture commitment, not a per-operator habit.

## Endpoint surface

The same surface is available over the public API:

```http theme={null}
GET /api/v1/settings/compliance/inbox-ai-privacy
PATCH /api/v1/settings/compliance/inbox-ai-privacy
```

`PATCH` accepts any subset of the three fields and leaves the others
untouched:

```json theme={null}
{
  "auto_categorize": false,
  "auto_summarize": false,
  "spam_action": "none"
}
```

Writes are restricted to organization **owners**, mirroring the
[AI turn audit](/compliance/ai-turn-audit) and
[policy scanner](/compliance/policy-scanner) posture gates — shipping
message content to an external processor is a compliance posture decision,
not an agent preference.

## Compliance posture framing

Both behaviours ship message content to a third-party LLM, and several
obligations treat processors as a regulated lane:

* **HIPAA**: inbound message bodies can carry PHI. Until your BAA posture
  covers the LLM hop, keeping `auto_categorize` and `auto_summarize` off is
  the conservative read. See [HIPAA](/compliance/hipaa).
* **EU data governance**: message bodies may contain personal data subject
  to processor restrictions. Opting out keeps processing inside your agreed
  scope. See the [GDPR posture guide](/compliance/gdpr-posture-guide).

This control is always tenant-owned — the platform defaults it ON for
day-0 convenience, and opting out is a posture decision your organization
makes. There is no platform-level mandate in either direction; the
tenant-owned posture model applies to all compliance gates
([posture overview](/compliance/posture-overview)).

## Effect on features when off

Turning a gate off stops the corresponding LLM call and nothing else:

* **Auto-categorize off**: conversations still arrive; inbox tabs and saved
  views still render, with no category assigned. Spam handling stops.
* **Auto-summarize off**: closing a conversation skips the summary write.
  Conversation history, notes, and tags are unaffected.
* **Spam action**: with `auto_categorize` off, the classifier never runs, so
  `spam_action` has nothing to act on regardless of the saved value.

Disabling a gate does not backfill or remove data accumulated while it was
on — existing categories and summaries stay until you clear them.

## Verify the state landed

1. Apply your change in the dashboard or with `PATCH`.
2. Read back with `GET` and confirm the fields match what you set:

```json theme={null}
{
  "auto_categorize": false,
  "auto_summarize": false,
  "spam_action": "none"
}
```

3. The write records an audit log entry
   (`settings.inbox_ai_privacy_updated`), so the change is visible in your
   organization's audit trail.

The inbound path caches your organization's settings for up to 60 seconds,
so a flip can take up to a minute to take effect on new messages.

## Related

* [DLP scanner](/compliance/dlp-scanner) — outbound message body scanning
* [Agent identity governance](/compliance/agent-identity-governance) — AI
  agent posture
* [HIPAA](/compliance/hipaa) — PHI handling posture
* [GDPR posture guide](/compliance/gdpr-posture-guide) — EU data governance