Human handoff — when and how the chat passes the case

The most important moment in any AI chat system is when it does not answer. When it recognizes this is a human case and hands off. We call this human handoff, and Nortinia AI Chat is tuned to three triggers.

The three triggers

1. Three turns without resolution

The chat always carries an intent tracker. It tries to categorize what the visitor wants (information, add to cart, complaint, booking, etc.). If for three user turns we see the same intent and it does not converge to a resolution (e.g. the chat repeats itself or the user re-asks), we auto-route.

Detecting "does not converge" is not semantic magic — it is a simple heuristic: we compare consecutive question-answer pairs by token-overlap. If similarity is 70%+ across two consecutive turns, the chat has not moved — bring a human in.

2. Sensitive topic

There are three topic classes we do not let AI handle:

• Refund > €500 — the financial risk is too high to take a model hallucination on.
• Legal matter — mention of litigation, GDPR request, consumer-protection complaint. The model never gives legal advice, and it should not be allowed to improvise.
• Formal complaint — "I want to file a complaint", "Are you a bot or a human?". A complaint is an elevated user expectation, and an AI reply to a complaint generates another complaint.

We detect these with regex + LLM-classifier combination. On a hit: immediate handoff, no third turn.

3. Explicit request

"I want to speak to a human", "Live customer service?", "Transfer me to an operator". Unambiguous. Immediate handoff, no explanations. (The chat is often tempted to reflex with "I can help too!" — that is forbidden. If asked, granted.)

The anatomy of a handoff

What happens when it switches to handoff:

1. Email notification — a structured notification goes to the tenant's configured handoff email (or list). It is not just "New case, come look!" — it carries the full transcript and an AI-generated summary.
2. Dashboard notification — on the operator dashboard the session gets a red badge and moves to a top-priority queue.
3. User feedback — the chat immediately signals: "I am connecting you to a colleague. They typically reply in 4 minutes." The concrete wait time is based on the last 30 days' median handoff-to-reply on that tenant.
4. AI goes silent — no more AI messages on this session. If the user writes on, the messages go to the operator side but there is no AI reply.

The quality of the handoff — the classified summary

The operator does not want to read 47 messages. The handoff email contains an AI summary that hands over the case in 5 fields:

Intent:       Refund request
Amount:       €780
Product:      Logistics Pro license, purchased 2025-12
User claim:   "I never received the December invoice and have not used it since December."
Risk:         Medium — purchase 6 months ago, usage trace in logs until December 15

The operator understands the situation in 30 seconds, and only needs to read the transcript if the summary missed something.

The hand-back

Frequently the operator resolves the main issue, but there is a small follow-up left ("where do I see the updated invoice?", "when does the confirmation email arrive?"). Operator decision: close the case in human mode, or hand it back to the chat for the remainder. If handed back, the chat receives the session via a structured "context resume" prompt: the operator-recorded resolution and the allowed follow-up types (e.g. "informational questions OK, further financial movement NO").

The data

Fleet-wide median handoff-to-reply time is 4 minutes. P90: 18 minutes. Outside business hours (8 PM – 8 AM) P90 = 11 hours, because most tenants do not staff support then. So at night the chat includes in the handoff message: "My colleague will be in touch at 8 AM. Until then I am here for information questions."

About 3.2% of handoffs return to chat mode (operator chose hand-back). The rest close as human cases.

Lesson

A good handoff is worth more in a chat system than a hundred perfect answers. A single "why didn't you transfer me to a human?" review destroys a hundred "what a great answer!" pieces of feedback. The chat is good when it knows what it does not know.