The third question determines whether an AI assistant is a useful front door or another support dead end:
What happens when the bot hits a question it can’t answer? Does it pass full transcript context to a human, or does the user have to re-explain everything?
The short answer
In FAQ Hub, an unsupported answer is not presented as a confident guess. The assistant says it could not find a definitive answer and offers a support ticket. The visitor can also choose Talk to a human directly.
When escalation or ticket creation happens from the same chat session, the conversation is stored and linked to the support workflow. Human staff can open the complete thread alongside useful journey context, so the customer generally does not need to start again.
There are important boundaries: external notifications carry recent conversation and context, not necessarily every message; a deliberately new or expired session is not the same continuous thread; and an unsupported answer offers escalation rather than silently assigning a human.
First, the assistant must know when to stop
FAQ Hub retrieves product-scoped support content for each substantive question and asks the support agent to answer from that evidence. A second judge pass reviews grounded answers before they are returned.
If the proposed answer cannot be verified, FAQ Hub replaces it with an explicit fallback: it could not find a definitive answer and can help create a support ticket. Provider failures and unsafe prompt-manipulation attempts also use controlled fallbacks rather than exposing internal errors as invented support advice.
This is more useful than a generic “confidence score”. A numerical threshold can look scientific while still allowing unsupported claims through. The operational rule is simpler: show a grounded answer with useful source links, or disclose the gap and offer the next step.
Two escalation paths, chosen by the visitor
Talk to a human
A visitor can request a person in natural language or use the visible Talk to a human control. FAQ Hub marks the chat as escalated, makes it available in the backoffice, and can notify configured channels such as Slack, Microsoft Teams, Telegram, n8n, or Zapier.
Support staff can open the conversation, reply with their identity, and continue the same thread with the visitor in real time.
Create a ticket
The visitor can create a ticket from the chat panel. FAQ Hub asks for contact details when they are missing, uses the latest visitor message as the initial description, and links the ticket back to the chat session.
The ticket also carries useful support context: the current page or article, search terms, sentiment, escalation reason, a latest-exchange summary, and retrieval evidence. The ticket dashboard can follow the session link and display the complete pre-ticket conversation.
Is the full transcript passed to a human?
For a human working inside FAQ Hub, the full stored thread is available oldest-first. User messages, AI replies, and staff messages are persisted.
For external notifications, FAQ Hub sends a bounded recent transcript plus journey and contact context. That is enough to alert a team without turning Slack or another webhook into the system of record. The authoritative conversation remains in FAQ Hub.
The AI itself uses the latest ten messages when generating the next answer. That keeps its working context bounded; it is separate from what a human can retrieve from the stored session.
So the precise answer is:
A customer escalating from the same active chat generally should not need to re-explain the issue. Human staff can retrieve the complete conversation and surrounding support context.
If the visitor intentionally starts a new chat, returns after the visitor token has expired, or continues after the previous session was closed, they may need to restate details. Prior journey summaries can help, but a new session is not automatically treated as the full old transcript.
Turn unanswered questions into documentation work
Escalation should not be the end of the loop. It is evidence that the support system needs one of three things:
- a missing article;
- a clearer article that uses the customer’s language; or
- a product or policy decision that documentation cannot resolve.
Review unsupported questions with the page, search, retrieval, and transcript context attached. Add the missing answer to the documentation backlog, assign an owner, publish the correction, and test the original question again.
FAQ Hub’s shared publication lifecycle matters here. The corrected article becomes available to the portal, search, and grounded chat instead of being copied into a separate bot-only training set.
A human handoff checklist
Before calling any support assistant production-ready, verify that it can:
- say when evidence is insufficient;
- show the customer an obvious human and ticket route;
- preserve the active conversation;
- attach page, article, and search context;
- identify whether the responder is AI or human;
- notify the right support channel without making that channel the only transcript store;
- turn repeated failures into owned documentation tasks;
- test the original customer wording after the article changes.
That is how documentation becomes active operational infrastructure. The bot is not a wall between the customer and support. It is the first participant in a traceable workflow that either answers from current knowledge or hands the problem forward with context intact.
Read the rest of the series: