Enable AI Chatbot
The AI chatbot automatically answers visitor questions before a human agent needs to intervene. It uses your Knowledge Base as a knowledge source (RAG — Retrieval Augmented Generation).
Prerequisites
- AI feature must be included in your plan (
ai_chatbotFeature-Flag) - The AI provider must be set up by the Super-Admin (IONOS)
- Ideally: Knowledge Base with published articles and/or snippets
Activation
- Go to Settings → AI Chatbot
- Enable "AI Chatbot activate"
- The basic settings are created with sensible defaults
What happens after activation?
When a visitor sends a message and no agent is assigned:
- The bot first checks the Bot Rules (keyword matching)
- If no bot rule applies, the AI chatbot takes over
- The AI bot searches your KB for relevant articles/snippets
- It generates an answer based on the found context
- If the confidence is too low or handoff keywords are detected, it is handed over to an agent
When the bot is allowed to respond (activation mode)
Under Settings → AI Chatbot → Activation, you define in which situations the bot responds at all:
| Mode | The bot responds … |
|---|---|
| Always active (default) | to every message |
| Agent fallback | only if no agent is available (offline or absent) |
| No agent online | only if no agent is connected |
| Outside business hours | only outside the widget business hours |
| According to schedule | only in the defined time windows |
| Only if an agent is online | only while at least one agent is online |
The last mode is the supervision mode: It reverses the logic. The bot only speaks as long as someone from the team can read along and intervene; if no one is there, it remains silent and the message waits for an agent. Intended for the introduction phase, when you don't yet trust the answers to be unsupervised.
This mode is deliberately fail-closed: If the presence cannot be determined, the bot remains silent. All other modes respond rather in doubt (fail-open).
💡 Even stricter is the AI Copilot: Widget AI off, Copilot on — then the AI only suggests answers internally and an agent sends them with a click. See AI Copilot.
Attention when combined with the widget: If the Offline Behavior of a widget is set to "AI takes over", but the activation mode is set to "Only if an agent is online", the two cancel each other out — the widget then shows the offline form.
Provider & Model
The provider is IONOS with data centres in Germany; the model is configured globally by the Super-Admin. As a tenant admin, you see the active provider under AI settings, but you cannot change it yourself.
| Provider | Models | Hosting |
|---|---|---|
| IONOS AI Model Hub | Mistral Small 24B (standard), Qwen 3.5 9B (budget), GPT-OSS 120B, Llama 3.3 70B (premium) | Germany (EU) |
💡 GDPR note: All AI requests run through the IONOS AI Model Hub — inference and data processing remain in Germany. Providers based in the US are technically blocked.
Status display
Under AI Chatbot → Settings you see:
- Provider: Which provider is active
- Model: Which model is used
- Indexed chunks: How many KB entries are available for the bot
- Token limit: Your monthly token budget
The bot has not responded — what's the reason?
The bot intentionally remains silent if one of these conditions is met:
| Reason | Check under |
|---|---|
| Chatbot disabled for the tenant | Settings → AI Chatbot |
| Chatbot overridden for this widget | Settings → Widgets → [Widget] → AI |
| Activation mode does not currently fit | see section When the bot is allowed to respond |
| Conversation is assigned to an agent | in the inbox; after an agent response, the bot deliberately remains silent |
| Token/cost budget exhausted | Settings → AI Chatbot (Budget card) |
| Platform-wide emergency stop | only the operator can see this — contact support |
Each of these cases is logged server-side (ai_no_reply with reason). If you can't make sense of it, tell the support the time and the conversation — the reason can then be read out unambiguously.
Not among the reasons are two cases that existed earlier and have been fixed since 08/12/2026: a second message that came while the bot was still writing the first (now put in the queue and answered), and an empty answer from the model (now results in a substitute message or handover to an agent, instead of disappearing without a trace).