What a Chatbot Should Say When It Does Not Know: Designing the Fallback

TL;DR: The fallback is the rarest thing a chatbot says and the one visitors remember. In Chatonbo's analysis of 5,470 real website chat conversations, only 1.0% included a request for a human, 2.3% of opening messages asked for a human or contact details, and assistants redirected visitors to a human channel in 3.7% of conversations. So the "I do not know" moment is uncommon, but every one of them is a trust test. A good fallback admits the gap in one sentence, gives the related fact it does have, offers one concrete next step, and asks for an email only when a follow-up is actually needed. Then someone reads the transcript and writes the missing answer so it never happens again.

How often does a chatbot actually need a fallback?

The numbers come from 246 live assistants between 13 June and 11 September 2026; the method and the rest of the data are on the What shoppers ask an AI chat assistant study page.

Read those together and a design principle falls out. The assistant redirected to a human (3.7%) more than three times as often as visitors asked for one (1.0%). Some of that is right: billing disputes and complaints belong with a person. Some of it is the assistant giving up on a question your website simply never answered. Either way, the visitor did not want a human. They wanted an answer, and 39% of the time there was no human on shift to hand them to.

Why does the fallback matter more than its frequency?

Three reasons.

It is where hallucination lives. An assistant that has the answer just quotes it. The risk of an invented return policy or a made-up delivery date sits entirely in the moments when the assistant does not have the answer and fills the silence anyway. A well-designed fallback is the cheapest hallucination control you can buy.

It is the only time the visitor judges the assistant instead of the answer. When the assistant is right, the visitor credits your store. When it is stuck, the visitor decides whether the whole thing is a gimmick. 82% of rated conversations in the study got a thumbs up. The down votes cluster around the stuck moments.

It is your gap report. Every fallback is a question a real visitor asked that your pages do not cover. That list is worth more than any keyword tool.

What should a good fallback say?

The structure that works, in order, with the whole thing under four sentences:

1. Admit it plainly. "I do not have information about wholesale pricing." Not "I'm sorry, I didn't quite catch that", which blames the visitor for a gap that is yours.
2. Give what it does know. "Retail pricing and the volume discount at 10 units are on the pricing page." Related facts keep the visitor on your site instead of a competitor's.
3. Offer one concrete next step. A link to the right page, an email address with a stated reply time, or a WhatsApp button. One, not a menu.
4. Ask for contact details only when there is a follow-up to deliver. "If you leave an email, the team will send wholesale terms within one working day." In the study 12.8% of conversations ended with the visitor leaving contact details; a specific promise is what earns them.

A complete example: "I do not have wholesale pricing here. Retail pricing and the 10-unit discount are on the pricing page. If you leave an email I will pass this to the team and they will send wholesale terms within one working day."

What does a bad fallback look like?

Most of the fallbacks we see in the wild fail in one of four ways:

• The loop. "Could you rephrase that?" followed by the same non-answer. Visitors already type fragments (the average opening message is 51 characters and only 35.7% contain a question mark); asking them to rephrase rarely produces new information.
• The invention. "Wholesale orders ship free and arrive in two days." Confident, specific, and false. This is the one that ends up in a complaint email with a screenshot.
• The forced handoff. "Let me connect you to an agent" at 23:00 on a Saturday, followed by silence. 24.7% of conversations happen on a weekend.
• The wall. "I can only answer questions about our products." True, unhelpful, and the visitor leaves.

How should the escalation path work?

The human path has to exist, but it should be sized for 1%, not designed as the main road.

• Default to asynchronous. An email reply with a stated time frame beats a fake live handoff. It works at 03:00 and it leaves a record.
• Attach the transcript. Whoever picks it up should see the question, the assistant's answer and the visitor's details without asking again. On Chatonbo the transcript and an AI summary travel with the lead to your inbox, CRM or webhook.
• Escalate immediately for the categories that need it. Complaints (0.9% of conversations), anything about a refund or a chargeback, and anything legal. Tell the assistant this explicitly in its instructions.
• Do not escalate for a missing fact. Add the fact instead. The support agent setup shows where handoff rules and the knowledge base sit side by side.

How do you turn fallbacks into knowledge?

This is the part most teams skip. Once a week:

1. List every conversation where the assistant said it did not know or redirected to a human.
2. Group them by topic. In the study the big topics were product questions (32.7%), support and account help (17.2%), pricing (10.3%) and shipping (5.0%); your gaps will cluster the same way.
3. Write the missing answer once, in plain language, on a page or in the knowledge base.
4. Ask the assistant the same question again and confirm it now answers.

Do this for a month and the fallback becomes what it should be: a rare, honest sentence that hands a genuinely unusual question to a person, instead of a daily apology for pages you have not written.

You can see how an assistant behaves on your own pages, including what it says when it is stuck, on the live demo. Ask it something your site does not answer. The reply tells you more about the assistant than any feature list.