Smart Escalation: When AI Should Hand Off to a Human

The mistake that destroys trust

The worst chatbot insists on answering when it doesn’t know. It makes up information, gives generic responses, or repeats the same unhelpful answer. This doesn’t just fail to solve the problem — it damages trust in the brand.

The solution: a smart escalation system that knows when to yield and transfers all context to the human team.

Three escalation mechanisms

1. Automatic by confidence score

Every Citai response has a confidence score (0 to 1) calculated from vector similarity, cross-encoder reranker score, and source quantity/relevance.

When the score is below threshold (configurable per KB, default 0.5), Citai auto-escalates instead of giving a low-quality response.

2. Manual by user

A “Talk to a human” button is always visible. The user never feels trapped in an AI loop.

3. By business hours

Outside configured hours:

• FAQs keep working 24/7
• Human-needed queries show a capture form (name, email, subject)
• Team receives everything next business day

Webhooks: full context without repeating

When escalated, Citai sends an HTTP POST with conversation history, confidence levels, KBs consulted, and pre-chat data.

{
  "event": "escalation.triggered",
  "reason": "low_confidence",
  "confidence": 0.28,
  "conversation": [
    {"role": "user", "content": "My order #45678 hasn't arrived after 10 days"},
    {"role": "assistant", "content": "Standard shipping time is...", "confidence": 0.28}
  ],
  "pre_chat_data": {"name": "Mary Smith", "email": "[email protected]"},
  "kbs_consulted": ["Logistics", "Policies"]
}

Common integrations

Negative feedback as trigger

When a user gives negative feedback (thumbs down), Citai can auto-escalate if confidence was already low, and feeds the knowledge gaps system for continuous improvement.

Escalation metrics

The analytics dashboard shows resolution rate, escalation reasons, resolution time comparison (AI vs human), and estimated savings.

Smart escalation included in all plans. Try Citai →