The AI pipeline
When a customer sends a message, Chatonio's AI pipeline processes it through several steps:
- Context building — the system prompt, conversation history, relevant FAQ entries, and knowledge base articles are assembled
- AI response — the configured AI model generates a response using this context
- Grounding check — before delivery, the drafted reply is automatically audited. Responses containing facts, IDs, amounts, or website URLs that aren't backed by a tool result, the customer's own messages, or your KB/FAQ are blocked and replaced with a safe fallback (and the conversation is escalated to a human). If the customer pasted a URL matching one of your configured integrations, the AI is forced to actually look it up via the tool instead of guessing
- Escalation check — a lightweight check determines if the conversation should be flagged for human review
- Delivery — the response is sent back through the originating channel
The grounding check is why the AI can't invent order details or links — if it can't back a claim up, the claim doesn't reach your customer.
Knowledge sources
The AI draws from two main knowledge sources to give accurate answers:
- FAQ entries — question/answer pairs matched by meaning (semantic search), so a customer's question matches even when it's phrased differently or asked in another language
- Knowledge base articles — longer-form content; the best-matching articles are found by meaning and injected as reference, with translations automatically used to match the conversation's language
For a FAQ entry or an article to reach the AI’s context at all, it must have AI Enabled (include in AI context) ticked. That is independent of visibility: an internal article with the box ticked is still used to answer customers, and a public one with it unticked is not.
Model configuration
Chatonio runs the AI models for you — there's nothing to install or connect. Each project can configure:
- Primary model — used for customer responses (higher quality)
- Utility model — used for summaries, escalation checks (lower cost)
- Vision model — optional; when set, the AI can read images the customer sends (screenshots, receipts, photos). Off by default
Models are configured at the project level and can be overridden per channel.
Understanding images
When you choose a vision model for a project (Settings → AI defaults (inherited by channels) → Vision model), the AI can look at images a customer sends — a screenshot of an error, a photo of a receipt, an order confirmation — and read the details out of them. It pulls out the visible text, numbers, and on-screen elements and answers the customer's question from what it sees, in the same reply. For privacy, it's instructed not to transcribe sensitive data such as full card numbers or security codes.
Image understanding is opt-in and stays off unless you set a vision model, and it's billed per image analyzed like the rest of the AI. It applies to images (JPEG, PNG, GIF, WebP); PDFs and other files are still delivered to your operators but aren't read by the AI. Only the most recent few images in a message are analyzed, to keep costs predictable.