AI
AI agents
An AI agent answers the conversations it handles, in the persona and language you set, from the knowledge you give it, using the tools you allow. When it can't help, it hands the conversation over to your team. AI agents need a plan with AI.
The workspace's Brand doesn't apply to AI agents: they follow their own persona.
Create an agent
- Open AI agents → New Agent (Start fresh, or Import backup).
- Name: what you call it in the dashboard (
Support bot). - Persona: the Persona Name customers see (
Ayu), a Communication Style (Professional, Friendly, Candid, Quirky or Custom) and Additional Preferences (Always answer in two sentences or less). - Response Language: Auto (match user's language) or a fixed language (English, Indonesian, Japanese, Chinese, Korean, Spanish, French, German, Portuguese, Arabic, Hindi, Thai, Vietnamese).
- Press Create Agent, then add knowledge and tools.
To put it to work, set a channel's Default handler to the agent (Channels), or assign it to a conversation from the inbox.
The agent page
| Tab | What's there |
|---|---|
| Settings | Name, Status (active or inactive), persona, response language, Delete Agent |
| Knowledge | Files and markdown it reads. See Knowledge |
| Tools | Built-in tools, plugin tools and your API connections. See Tools and API connections |
| Behaviour | Handoff and memory (below) |
| Backup | Export and import (below) |
Behaviour
Handoff to people
When the AI steps aside and your team takes the conversation:
- Keywords that hand over immediately (
human, agent, complaint). - Hand over after this many "can't help" answers in a row.
- Also hand over when the customer is upset (sentiment).
- Message to the customer when it hands over (a template; empty sends nothing).
The AI also hands over when the customer asks for a person (the Hand over to the team tool), when the conversation reaches its daily AI spend cap, or when the workspace runs out of AI tokens.
Memory
- Refresh the conversation summary every N messages: keeps long chats short for the model.
- Remember customers across conversations: long-term memory per contact (needs a plan with AI memory). See Contacts › AI memory.
Attachments
When customers send files, the AI can read images, PDF and text documents and voice notes (transcribed), within your plan, the channel's AI reads attachments setting and the platform limits. Videos and stickers stay placeholders.
Backup
- Export downloads name, persona, tools, API connections and behaviour as a JSON file. API secrets are not included.
- Import replaces this agent's settings, tools, API connections and behaviour with a backup file. Status and knowledge are kept; secrets of new API connections must be entered again.
AI tokens and usage
Your plan includes AI tokens every month, and unused ones expire. After they run out, the AI uses gifted tokens, then tokens you buy in Settings › Subscription, which never expire. Settings › AI shows your tokens and an estimate of the replies they buy, and has:
- Sentiment & intent: analyse every inbound text (uses AI tokens); powers the mood in the inbox and Top intents in Analytics.
- Memory retention: unpinned contact memories nobody used are forgotten after this many days (empty uses the platform default).
The owner gets an email when tokens run low. With none left the AI stops, even in the middle of a turn, memory and summaries pause, and the thread says AI stopped: no tokens left. See Billing.
FAQ
Which AI model does the agent use?
The platform's default reply model, picked by the administrator. You don't need your own OpenAI key.
Can the AI make mistakes?
Yes. Give it clear knowledge, keep high-risk tools set to ask first, and set handoff keywords so customers can always reach a person.
Does the AI answer conversations my team is handling?
No. It only answers conversations it handles; taking over a conversation stops it.
How do I stop the AI for one conversation?
Press Take over in the conversation, or use an auto reply with Pause the AI.