# AI agent and knowledge

> An AI agent that answers first from your docs and hands off to you the moment someone asks for a human.

Source: https://conciergesdk.com/docs/ai-agent/

Turn on the AI agent and it answers new messages in seconds, from the knowledge you give it. Every AI reply is labeled as AI in the chat, in all SDKs, and the chat shows a short notice that replies may come from an AI assistant (the EU AI Act requires this, so the label cannot be turned off; you can rename the assistant). It steps aside the moment a user asks for a person.

## Turn it on

In the dashboard, open your app's **AI** settings:

| Setting | Default | What it does |
|---|---|---|
| Enabled | Off | The agent replies to customer messages. |
| Model | `claude-haiku-5-5` | The Claude model that writes replies. |
| Show as AI | On | AI replies carry an "AI" label in the chat. |
| Handoff | `on_request` | `on_request` hands off when the user asks for a human. `never` keeps the agent on. |
| System prompt | Empty | Extra instructions: tone, what not to promise, when to escalate. |

AI usage is included in every plan. No per resolution fees. There is no free plan or trial, so the AI (replies, drafts, the knowledge autopilot) starts once your workspace subscribes; before that, messages from your test devices go to your inbox.

## Give it knowledge

Add knowledge docs in the dashboard: FAQs, pricing, known issues, how to restore a purchase. Each doc is a title plus text. The agent reads all of them on every reply (prompt cached, so it stays fast and cheap) along with your system prompt.

Small knowledge bases don't need search. Indie app docs fit in the prompt, and that gives better answers than retrieval.

Only **live** articles reach the agent. Drafts wait for your review on the Knowledge page.

Don't want to write them yourself? The [knowledge autopilot](https://conciergesdk.com/docs/knowledge.md) drafts articles from your help pages, product notes, codebase (through MCP) and resolved conversations, shows you the questions the agent couldn't answer, and adds buttons that open the right screen in your app.

```http
POST /v1/apps/:appId/knowledge
{ "title": "Restoring purchases", "content": "Go to Settings, tap Restore. If Pro still looks locked..." }
```

## Languages

Users write in whatever language they like, and the agent replies in the same one. Your knowledge docs can stay in one language (yours): the agent reads them and answers in the user's. Store review reply drafts work the same way, in the reviewer's language.

The chat's own interface (buttons, labels, the composer placeholder, the AI label) comes in 34 languages and follows the device language, right to left included. See [Localization](https://conciergesdk.com/docs/localization.md).

## Handoff

When a user asks for a human, the agent replies with your handoff text, pauses itself on that conversation and marks it **high priority** in your inbox. When you reply yourself, the agent pauses on that conversation too, so it never talks over you. Resume it any time from the conversation.

## How it shows up in the app

- Replies arrive like any other message, with the "AI" label when **Show as AI** is on.
- The typing indicator shows while the agent writes.
- Push notifications work the same as for human replies.

> **Coming soon:** - **Actions, not just answers.** Your app registers tools (restore purchases, resync library) and the agent can run them after the user confirms. Server tools look up subscriptions and send offers your policy allows.
> - **More knowledge sources:** file imports and your App Store description, alongside the links, notes and resolved chats the [knowledge autopilot](https://conciergesdk.com/docs/knowledge.md) reads today.
> - **Draft mode:** the agent drafts, you approve with one tap, until you trust it.
> - **Limits as numbers:** replies used, cost per reply, a visible cap.