> ## 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.

# How OpenCX Works

> The mental model behind OpenCX — channels ingest, AI answers from your knowledge, Autopilot gates replies, Actions do real work, Handoff routes the rest.

Every OpenCX agent is the same five pieces wired together. Once you understand the shape, every channel, feature, and integration fits the same mental model.

## The flow

```mermaid theme={"dark"}
flowchart LR
  A[Customer message] --> B[Channel]
  B --> C{Autopilot}
  C -->|on| D[AI agent]
  C -->|off| F[Human team]
  D --> E[Knowledge + Actions]
  E --> G[Reply]
  D -.handoff.-> F
```

1. **A message arrives** on one of your channels — widget, phone, WhatsApp, email, or SMS.
2. **Autopilot decides** whether the AI replies or the message goes straight to your team.
3. **If the AI replies**, it reads from your Knowledge, optionally calls Actions (your APIs), and sends a reply back on the same channel.
4. **Anything the AI cannot confidently handle** (low confidence, prohibited topic, frustrated user) hands off to a human.

## The five pieces

<AccordionGroup>
  <Accordion title="Channels — where messages come from" icon="satellite-dish">
    Each channel (widget, phone, email, SMS, WhatsApp) captures messages from your customers and normalizes them into **sessions**. A session is the OpenCX-wide unit of a single conversation regardless of channel. See [Build Agents](/build-agents/get-started).
  </Accordion>

  <Accordion title="Knowledge — what the AI reads from" icon="book">
    The AI answers from two sources: **Help Center content** (articles you write) and **Knowledge Sources** (connected systems like Zendesk, Notion, Confluence, your website). It also uses **Training Scenarios** you write by hand for exact answers. See [Agent Training](/knowledge/introduction).
  </Accordion>

  <Accordion title="Autopilot — when the AI replies" icon="sliders">
    Autopilot is the on/off switch per channel. When on, the AI responds automatically. When off, every message routes to your team's inbox. You can also gate autopilot by **topic** — reply to billing questions, hand off login issues. See [Autopilot topics](/handoff/topics).
  </Accordion>

  <Accordion title="Actions — what the AI can do" icon="bolt">
    Actions let the AI call your APIs at conversation time — look up an order, cancel a subscription, update a field. You define them with an OpenAPI-style spec and the AI picks the right one based on the customer's intent. See [AI Actions](/actions/introduction).
  </Accordion>

  <Accordion title="Handoff — how humans take over" icon="user-group">
    When the AI hits its limits, the session hands off to a human. OpenCX syncs the full conversation to your helpdesk (Zendesk, Hubspot, Intercom, etc.) so your team sees the context. See [Handoff](/handoff/introduction).
  </Accordion>
</AccordionGroup>

## Why this shape

Most customer support tools either automate everything (and answer the easy questions but fail on the hard ones) or route everything to humans (and burn payroll on FAQs).

OpenCX separates the decision from the answer. **Autopilot decides**, **Knowledge + Actions answer**, **Handoff catches the rest**. You tune any piece without rebuilding the others — swap the helpdesk without retraining, add a channel without redoing autopilot rules.

***

## Related Documentation

<CardGroup cols={2}>
  <Card title="Core Concepts" icon="book" href="/core-concepts">
    The vocabulary: session, ticket, autopilot, handoff, action, insight.
  </Card>

  <Card title="Quickstart" icon="rocket" href="/quickstart">
    Skip the theory — stand up your first agent in 10 minutes.
  </Card>

  <Card title="Build Agents" icon="robot" href="/build-agents/get-started">
    Pick a channel and go live.
  </Card>

  <Card title="AI Safety" icon="shield" href="/safety/introduction">
    What the AI won't do — PII redaction, prohibited topics, inspect mode.
  </Card>
</CardGroup>