AI Overview

Ticket0 uses a hybrid AI pipeline designed for low cost, good accuracy, and safe handoff.

High-level flow

When a new ticket or chat message arrives, the system runs this sequence:

1. Fast understanding (Gemini Flash / Haiku-class tasks)
2. Python context assembly (knowledge base, customer context, similar tickets, language)
3. Final response generation (Sonnet-quality drafting)
4. Confidence scoring + action decision (auto-send, draft for review, or suppress)

What happens on inbound ticket email

1. Ticket is created and classified (category, intent, urgency, confidence).
2. AI tool calls gather context from KB and prior conversations.
3. The agent drafts a response using workspace persona + learned team voice patterns.
4. A confidence score is computed from retrieval quality, classification confidence, sentiment, topic risk, and pattern matches.
5. Depending on the score and the workspace's auto-reply settings, the draft is auto-sent without review (autoReply on, high confidence), shown in the composer for one-click approval (mid confidence), or suppressed entirely so the operator writes from scratch (below the 60 floor). See Confidence thresholds for the bands.

What happens on chat widget messages

1. Customer message is stored in the conversation.
2. AI generates a response or triggers escalation when confidence is low.
3. If escalation is needed, the user is offered escalation in-chat and asked for a follow-up email if one isn't already verified.
4. Once the email is confirmed, a new email-style ticket is spawned for an operator to work. It contains an AI summary of the chat, the escalation reason, and a link back to the original chat. The two tickets are bidirectionally linked. The chat session stays open — the AI keeps responding to any further customer messages — and humans never reply through the chat widget. See Live chat AI → Escalation in widget for the full lifecycle.

Design goals

• Keep routine responses fast and low-cost
• Keep high-risk conversations in human hands
• Be explicit when confidence is low
• Use operator feedback as the learning signal over time

Continue with How the AI learns from your team and What the AI will and won't do.