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

# Architecture overview

> How data flows through OpenCX — channels in, AI inference, knowledge base, helpdesk handoff out, and what your team controls along the way.

OpenCX sits between the channels your customers use and the systems your team operates. This page is for admins and security reviewers — what data crosses what boundary, what stays inside OpenCX, and what your team can switch off.

## Data flow at a glance

```mermaid theme={"dark"}
flowchart LR
  A[Customer message] --> B[Channel]
  B --> C[OpenCX]
  C --> D[AI agent]
  D --> E[Knowledge base]
  D --> F[Actions / your APIs]
  C --> G[Your helpdesk]
  C --> H[Your team's inbox]
```

Every message lands in OpenCX first. The AI reads from your knowledge base, optionally calls your APIs, and replies on the same channel. Anything the AI hands off goes to your team — in the [OpenCX inbox](https://platform.open.cx/inbox), your helpdesk, or both.

## Where data lives

**Sessions** are the unit of every conversation — customer messages, AI replies, handoff events, and action calls all attach to one session. Sessions stay inside your OpenCX organization. See [Core concepts](/core-concepts) for the full vocabulary.

**Knowledge** is everything the AI reads from. Articles you author in the [Help Center](/help-center/overview) plus content synced from connected sources (Zendesk, Notion, Confluence, Shopify, websites). Drafts and segment-restricted content stay internal — the AI uses them in agent-facing surfaces only, never in customer replies. See [Agent Training](/knowledge/introduction).

**Identifiers shared with third parties** are the minimum needed to make handoff reversible. When a session hands off to a helpdesk like Zendesk or HubSpot, OpenCX writes the session ID and tags onto the ticket so your team can click back. The full transcript syncs as ticket comments.

## What touches the AI

* The customer's current message and the recent conversation history.
* The relevant knowledge base entries, ranked by similarity.
* Action specs (your API definitions). The AI does not see the data your APIs return until it calls one — and only the response of that call.

## What your team controls

<AccordionGroup>
  <Accordion title="Autopilot" icon="sliders">
    Per-channel switch plus topic gating. Turn the AI off entirely for a channel, or limit it to specific topics (reply to billing questions, hand off login issues). Configure from [Autopilot settings](https://platform.open.cx/settings/autopilot). See [Autopilot topics](/handoff/topics).
  </Accordion>

  <Accordion title="PII redaction" icon="user-shield">
    Automatic redaction of personal data before storage and inference. Emails, phone numbers, card numbers, and other patterns are masked in transcripts and never reach the AI in raw form. See [PII redaction](/safety/pii-redaction).
  </Accordion>

  <Accordion title="Inspect mode and Debug mode" icon="magnifying-glass">
    Open any session and see exactly what the AI saw — the knowledge entries it ranked, the actions it considered, the prompts it built. Useful for security review and for tuning answers. See [Inspect mode](/safety/inspect-mode).
  </Accordion>

  <Accordion title="SSO and access control" icon="key">
    Enforce SSO at the org level and assign roles per teammate. Restrict who can change Autopilot, view transcripts, or run actions. See [SSO](/security/sso).
  </Accordion>
</AccordionGroup>

## Limits and honest expectations

AI replies typically land within a second or two; under load this can stretch. Knowledge syncs poll on a schedule — typically every few hours, configurable per source. Handoffs to third-party helpdesks are best-effort; if a third-party API is down, OpenCX retries and surfaces the failure as a `conversation.handoff_to_third_party_failed` event.

***

## Related Documentation

<CardGroup cols={2}>
  <Card title="AI Safety" icon="shield" href="/safety/introduction">
    Inspect mode, debug mode, PII redaction, prohibited topics.
  </Card>

  <Card title="SSO" icon="key" href="/security/sso">
    Enforce SSO at the org level.
  </Card>

  <Card title="Bug bounty" icon="bug" href="/security/bug-bounty">
    Report security issues.
  </Card>

  <Card title="API authentication" icon="code" href="/api-reference/authentication">
    How API keys and scopes work.
  </Card>
</CardGroup>