How the Chatbot Handles Out-of-Scope Questions

A visitor types a question your knowledge base never covered. Maybe it is a competitor's pricing, a feature you do not offer yet, or a niche edge case nobody wrote a help page for. What the chatbot does in that single moment tells the visitor whether they can trust the rest of the conversation. We built BestChatBot to treat that moment as a feature, not a failure. When a question falls outside your content, the widget declines instead of inventing, and it can hand the visitor a real next step. Here is how that works, and why declining is the honest answer.

What "out of scope" actually means

Out-of-scope questions are the ones your knowledge base has no grounded answer for. The widget answers from the content you trained it on: your help center, uploaded docs, scraped pages, and manual entries. When a visitor asks something those sources cover, retrieval finds the matching passages and the model writes a reply from them.

An out-of-scope question is the opposite. There is no matching passage, because you never published anything about it. A general chatbot bolted onto a raw language model will often answer anyway, producing a fluent paragraph with nothing behind it. That is the exact moment a made-up answer slips through and a visitor walks away misinformed.

We think the right move is the boring one. No matching content means no grounded answer, so the widget says it does not have that information rather than dressing up a guess. The question is out of scope, and the bot tells the truth about it.

How the widget decides a question is out of scope

Every reply starts with a retrieval step. The visitor's question is matched against your knowledge base, and the closest passages come back ranked by how well they fit. The model only writes an answer from passages that actually relate to the question.

When nothing relevant comes back, there is no grounded material to write from. The widget does not fall back to the model's general training to fill the gap. Instead it recognizes the question sits outside your content and switches to a decline. This is the same grounding discipline behind our no-hallucination approach: the answer is tied to your real content or it does not get made.

The short version of the flow looks like this:

1. The visitor asks a question.
2. The widget retrieves passages from your knowledge base.
3. Strong match, the model writes a grounded answer.
4. No match, the widget declines and offers a next step.

That branch at step three is the whole game. A grounded answer and an honest decline come from the same retrieval, so the bot is never improvising about whether it knows something.

What a good decline looks like

Declining badly is its own problem. A blunt "I cannot help with that" leaves the visitor stuck and annoyed. A good decline does three things at once: it admits the gap plainly, it stays in your brand voice, and it points somewhere useful.

The widget can also keep showing sources when it does answer, so the visitor sees where real answers come from and understands why this one has nothing behind it. Citations are a toggle, and they pair naturally with scope handling. When an answer is grounded, the cited page proves it. When a question is out of scope, the absence of a source is the signal. We cover that side in detail in the post on source citations.

A decline is not a dead end. The widget can offer to capture the question and route it to your team, which is where escalation comes in.

When the chatbot hands the question to your team

Some out-of-scope questions are just noise. Others are real customers with a real problem your content does not yet address. For those, the widget escalates by creating a support ticket. It does not pass the chat to a live agent. It collects the question and routes it to your team's inbox by opening a ticket in Zendesk or Freshdesk, then tells the visitor their question is on its way to a person.

That keeps the experience honest from both sides. The visitor is not left guessing, and your team gets a clean record of what the bot could not answer. You pick one support tool, Zendesk or Freshdesk, and the widget files the ticket there. The visitor sees a clear confirmation instead of a vague apology.

Escalation also doubles as a map of your gaps. Every ticket born from an out-of-scope question is a question your knowledge base should probably cover. The widget plus your ticketing tool quietly builds a backlog of content worth writing. If you want the fuller picture of how the chatbot routes work that needs a person, the post on human handoff via tickets walks through it.

Why declining builds more trust than guessing

A bot that always has an answer sounds impressive until the first time a visitor catches it inventing one. After that, every answer is suspect, even the correct ones. Trust is fragile that way. One confident lie poisons a hundred true replies.

Declining flips that math. When the widget says "I do not have that, but I have passed it to the team," the visitor learns the bot only speaks when it actually knows. That makes the answers it does give more believable, not less. A boundary is a credibility signal. The bot that knows its limits is the one people start to rely on.

There is a practical upside too. Out-of-scope handling is human-in-the-loop by design. The widget does not quietly teach itself from a question it could not answer. You review the conversation in the dashboard, see the gap, and update the knowledge by adding a page, re-scraping a source, or editing an entry by hand. The next visitor who asks gets a grounded answer instead of a decline. Scope control is one of several settings on your website chatbot, sitting next to tone, citations, and escalation rules. If you are still wiring things up, the setup guide shows where these controls live.

FAQ

What does the chatbot do when it cannot answer a question?

It declines instead of inventing an answer. When a question falls outside your knowledge base, the widget says it does not have that information rather than producing an unbacked paragraph, and it can offer to route the question to your team.

Does the chatbot ever make up an answer to an out-of-scope question?

No. Answers are grounded in your content, so a question with no matching passage gets an honest decline. There is no fallback to generic model knowledge that would let a fabricated answer through.

What happens after the chatbot declines?

It can capture the question and create a support ticket in Zendesk or Freshdesk, routing it to your team's inbox. The visitor gets a clear confirmation that a person will follow up, instead of a dead end.

How do I fix recurring out-of-scope questions?

Review them in the dashboard, then update your knowledge base by adding a page, re-scraping a source, or editing an entry by hand. Out-of-scope handling is human-in-the-loop, so the bot does not learn on its own. Once the content exists, the next visitor gets a grounded answer.

Can the chatbot hand the conversation to a live agent instead of a ticket?

No. Escalation creates a ticket and routes it to your team's inbox, it does not transfer the live chat to a human. That keeps a clean record of every question the bot could not answer.

To see which plan fits the way your team wants to run grounded answers and ticket escalation, head to pricing.