Chatbot Fallback Message Examples: 20 Ways to Say "I Don't Know"

Every chatbot meets questions it cannot answer, and what it says at that moment decides whether the visitor rephrases, asks for a person or closes the tab. Many bots still reply with some version of "Sorry, I didn't understand that", repeated until the visitor gives up. The line names no problem and offers no route, and by the third time it reads as a wall.

The guide below is for whoever sets an assistant's wording, from founders to support leads. It defines the five causes of a fallback, gives 20 messages grouped by cause, sets out five rules that keep a fallback from becoming a dead end, shows where the wording lives in an AI support agent, and explains how to measure fallback rate.

Every source cited, from Nielsen Norman Group's usability research to two CHI conference papers, was read on 19 October 2026, and PepoChat's behaviour comes from its documentation and code. PepoChat is our product; the messages work in any tool that lets you shape what the bot says.

What is a chatbot fallback message?

A fallback message is what a chatbot says when it cannot give a grounded answer to the visitor's last message, because the content does not cover it, the question is unclear, or a rule stops the bot from answering. Older intent-based bots send one when no intent matches; Google's conversation design guidance on errors calls that a no-match error, raised when the system cannot interpret what the user said in context. An AI support agent falls back when its knowledge base search finds nothing relevant or an instruction tells it to decline.

A fallback is the honest alternative to a hallucination, a confident answer that your content does not support. The fallback says "I can't answer that" and offers a route; the hallucination says something false, and the visitor acts on it. Whenever you tune a bot to fall back less, check that it is not simply guessing more. The guide to stopping chatbot hallucinations covers the grounding that makes an honest fallback possible.

Why do fallback messages matter more than greetings?

Fallback messages matter more because they reach visitors at the moment the chat has already let them down, and the next sentence decides whether they stay. A greeting is read by people who have not asked anything yet. A fallback is read by people who asked, waited and did not get what they came for.

Real logs show how little patience visitors have. Chi-Hsun Li and colleagues analysed three months of a bank's chatbot on Facebook Messenger, 17,136 exchanges from 1,685 users, in a conversation analysis presented at CHI 2020. Counting a user who stayed away for more than 10 days as having abandoned the bot, they found that 45% of abandonments came after a single stalled exchange and 75% after no more than two in a row. One bad fallback is often the only second chance a chatbot gets.

Nielsen Norman Group saw the same pattern in moderated sessions. In The User Experience of Chatbots, Raluca Budiu reports that participants were annoyed when a bot repeated the same answer, that bots struggled to recover from unexpected input, and that bots which owned the failure and offered an escape hatch were received well. Her guideline is short: "Be honest about not understanding." She pairs it with a way out, such as a real person, a phone number or a link to another channel, and the post on website chatbot best practices summarises the rest of her guidelines.

The banking log also shows where stalls come from. Of the 9.4% of exchanges that made no progress, 97.4% started with something the user typed in their own words rather than a button they tapped. Free typing is exactly what an AI support agent invites, so the fallback deserves the same care as the greeting; the welcome message examples cover the other end of the conversation.

What are the five causes of a chatbot fallback?

A chatbot falls back for one of five reasons: the answer is not in its knowledge base, the question is ambiguous, the request is out of scope, the answer needs verification or account data, or a policy forbids it. The cause matters because each one has a different right next step, so a single generic fallback gets at least four of them wrong.

To tell the causes apart, ask what would have made an answer possible. More content fixes the first cause, and a clarifying question fixes the second. Nothing fixes the third, because the visitor came to the wrong place, so the fallback redirects without making them feel foolish. Identity or a live lookup fixes the fourth. The fifth is not a gap at all: you chose the limit, so the fallback should say who can help instead.

20 chatbot fallback message examples by cause

The 20 messages below come four per cause, each with its situation and the next step it offers. Brand names are made up, and details such as a 2-year warranty stand in for your own facts. Paste them into an intent-based bot as fixed responses, or use them as examples in an AI support agent's instructions.

Not in the knowledge base: messages 1 to 4

A grounded agent searches its content before every reply, and this fallback covers searches that come back empty or incomplete. Make clear the gap is in your content, not in the visitor's question. Message 2 shows the partial answer: give the part you can, then hand off only the rest.

Ambiguous question: messages 5 to 8

An ambiguous question has more than one reasonable answer, and a wrong guess costs more than a question. Offer the likely meanings as named choices: in the banking log, free-typed messages stalled 25% of the time and taps on buttons or quick replies 0.4% of the time, and a short named choice is the closest a free-text chat gets to a button. Ask once, and if the second reply is still unclear, offer a person.

Out-of-scope questions: messages 9 to 12

Out-of-scope requests are questions your assistant should not answer at all, from homework to another company's app. A flat "I can't help with that" leaves the visitor with nothing. Restate what the assistant covers and offer one or two tasks it does well, which also shows the visitor what to ask.

Needs verification or account data: messages 13 to 16

Some answers depend on who is asking, or live in a system the assistant cannot read. Say what the assistant cannot see, without implying the visitor did anything wrong, then offer the route that unlocks it: an identifier for a lookup, email verification, or a person with access. Never ask for a password or card number in chat, and never confirm whether an email address exists in your records.

In PepoChat, visitors verify with a six-digit code sent to their email from the shield icon in the chat header, which unlocks meeting memory where you have switched it on, bookings with their details and an emailed transcript. Integration lookups, such as an order found by email, are not gated by verification. Writing the requirement into the action's "when to use" description makes the agent ask, but the agent is not told whether the visitor completed it, so trim what the action returns, as the visitor verification docs advise. Message 16 is the wording for that rule, and the guide to chatbot identity verification covers which lookups need it.

Policy restriction: messages 17 to 20

A policy fallback is a limit you set on purpose: no medical or legal advice, no discounts, no refunds in chat, no talk about competitors. The answer may exist in your content, so the fallback should not pretend otherwise. Decline in one line, name who can help and offer the route there.

What should a chatbot say when it can't answer? Five rules

A chatbot that cannot answer should apologise once at most, name the limit, offer two ways forward, change course on a second miss and hand off with the context attached. Most come straight from usability research, and each shows up in the 20 examples above.

1. Apologise once, or not at all

One short apology is polite; three in a row read as a script. Google's guidance pairs a brief apology with a shortened version of the original question on the first retry, then adds options or examples on the second. Most of the 20 messages above contain no apology at all, because "I can't find that" states a fact rather than a fault. Save "sorry" for moments the bot actually failed the visitor, such as a second miss or a broken tool.

2. Name the limit in plain words

Say what is missing in the visitor's language, not the system's: "Our delivery pages don't cover Iceland" beats "No results found", and both beat an error code. Nielsen Norman Group's error-message guidelines ask for human-readable language, a concise and precise description of the problem, constructive advice and a tone that never blames the user. Google's guidance puts the principle in one line: "Be honest and transparent when explaining why something doesn't work." A fallback that names the limit also tells the visitor whether rephrasing could help, which a bare "I don't understand" never does.

3. Offer two ways forward

Every fallback should end with at least two things the visitor can do right away, such as a narrower question, a related topic or a person. In the Resilient Chatbots paper from IBM Research, presented at CHI 2019, 203 participants compared eight ways a chatbot can repair a breakdown in a scenario-based study and generally preferred two: offering options and explaining why the bot got stuck. The authors' reading is that both show initiative and give people something they can act on. NN/g's error-message guidelines agree: offer likely fixes as a short list to pick from. Two options beat one, because a single option is a funnel, not a choice.

4. Change course on the second miss

Never send the same fallback twice in a row. Google's guidance, written for Assistant voice apps that were deprecated in June 2023, describes a ladder: the first reprompt is brief, the second adds options or examples, and if that also fails the conversation ends, so users never hit more than three errors in a row. In a support chat, the right ending is a person, not a goodbye. The human handoff guide uses the same threshold: two replies that do not resolve the question, or one message with clear frustration, and the bot offers a person without being asked.

5. Hand off with the context attached

A handoff only works as a fallback if the person who picks it up can see what already happened. Pass the transcript, any contact details the visitor gave and the reason for the handoff, so nobody asks the visitor to start again. Then say honestly when a reply will come: a teammate shortly when someone is online, an email follow-up when nobody is.

Promising a person who never arrives turns a fallback into a broken promise. If nobody watches the chat live, the fallback should collect an email address once and set a realistic reply time; the offline chat message examples cover that wording in detail.

Where does fallback wording live in an AI support agent?

In an AI support agent, fallback wording lives in the instructions and guardrails the model reads before every reply, not in a fixed response typed once. An intent-based bot has a fallback intent or a no-match prompt, one slot where you write the exact sentence. A language model writes each reply fresh, so you describe the behaviour you want, give example wording, and let the model fit it to the question.

The difference changes how you write. A fixed script controls every word but cannot adapt, so the same sentence fires for a missing page and for a vague question. Instructions adapt to the cause but give up exact control, so you test them with real questions instead of proofreading them once. A fallback block in the instructions reads like a brief for a new hire:

When you can't answer from our knowledge base:
- Say what is missing in one sentence, in the visitor's words.
- Apologise at most once per conversation. No apology for a missing page.
- Offer two next steps: a related topic you can answer, and a person on our team.
- If a question could mean two things, ask which one, naming both.
- If the visitor has already rephrased once, stop asking and offer a person.
- Never say you are checking or looking into something unless you are.
For questions unrelated to Ledgerly, say in one line what you can help with.
For discounts, refunds, and legal or tax advice, decline briefly and say who can help.

The guide to writing chatbot instructions has persona and "never do this" examples by business type that pair with this block.

How fallbacks work in PepoChat

PepoChat's AI support agent searches the workspace's knowledge base for every reply. When it finds nothing relevant, it says so and offers to bring in your team rather than inventing an answer, as how answers work describes. The default line is "I don't have specific information about that in our knowledge base. Would you like me to connect you with a human support agent?" A yes escalates the conversation: the team inbox marks it escalated, and your team is emailed the reason and the last three messages, as the human handoff docs set out.

Three settings on the Widget Customization page shape the fallback copy around that default:

• How to sound and behave (up to 4,000 characters) sets persona, tone and wording. The instruction block above belongs here.
• Never do this (up to 2,000 characters) holds your policy restrictions. When a visitor asks for one, the agent declines briefly and offers what it can help with instead, which is the shape of messages 17 to 20.
• Away message (up to 300 characters, in the Team availability card) is what the agent says instead of promising a person whenever nobody can reply live. The agent then asks once for the visitor's email address and hands the conversation over.

What you cannot write away is the honesty. The agent instructions docs list the base rules that win over anything in your text: the agent still searches first, never invents an answer, offers a person when it cannot help and escalates when a visitor asks for one. Test every change in the Test the agent drawer, where chats are free and stay out of analytics, by asking questions you know your content cannot answer.

How do you measure fallback rate, and what is a good number?

Fallback rate is the share of visitor messages, or of conversations, that end in a fallback instead of an answer. No industry benchmark exists for it: the figures quoted online are starting targets or single studies, so treat your own trend and the mix of causes as the signal, not the level.

Definitions differ in their denominator, so numbers rarely compare. The chatbot analytics company Dimension Labs divides customer messages that hit the fallback by all customer messages, and its KPI documentation suggests under 5% as a starting target while stating that its targets are not industry benchmarks. The banking chatbot study counted stalled exchanges instead and found 9.4%. Neither describes a modern AI support agent, and PepoChat publishes no fallback figure of its own.

Last verified 19 October 2026.

Measure per conversation and split by cause. A per-message rate falls whenever conversations get longer, and a single overall number hides the difference between a missing page, which is a content task, and a policy refusal, which is the assistant working as designed. The post on AI chatbot metrics that matter calls the per-conversation version the "I don't know" rate and makes the key point: the right number is not zero. A near-zero rate alongside poor answer accuracy means the bot is guessing.

A weekly fallback review in five steps

1. Pull the week's escalations. In PepoChat, filter the team inbox by Escalated, or use Export CSV with a From and To date. Each escalation email also states its reason, such as "The agent could not find an answer or failed."
2. Tag each fallback with one of the five causes. Dimension Labs suggests reviewing the top 20 messages the bot did not understand, grouped by theme, which is a sustainable weekly size.
3. Fix by cause. Add content for missing answers, clearer pages for ambiguity, a sharper scope line for out-of-scope requests and an integration or verification rule for account data. Leave policy fallbacks alone unless the wording is wrong or your site promises what the assistant may not give.
4. Retest the exact questions in the test drawer after each fix.
5. Track the count per cause each week, next to the Topics card in Analytics, which labels every conversation from its first message.

What to do next

Pick five fallbacks from the tables, one per cause, and adapt them to what your content and integrations really support. Put the five rules into your assistant's instructions, then ask a test chat ten questions you know it cannot answer and read every reply. The guide to stopping chatbot hallucinations covers the grounding underneath, and the chatbot script examples show full conversations around these lines. To try it on your own content, create a free workspace: agent instructions, human handoff and the test drawer are on every plan, including Free.