Diesen Artikel gibt es noch nicht auf Deutsch. Zu den deutschen Artikeln
Every chat assistant meets questions it cannot answer. What it says then, its chatbot fallback message, decides whether the visitor gets what they came for or closes the tab. Much of the advice still describes button bots; here is the moment as it looks when an AI agent answers from your own content.
What a chatbot fallback message is, and what changed with AI agents
A chatbot fallback message is the reply a chat assistant gives when it cannot answer from what it knows. In a decision-tree bot it fires when the visitor's words match none of the options it was built with, hence the classic fallback response "Sorry, I didn't understand that. Could you rephrase?" Conversation designers call that trigger a fallback intent.
An AI agent understands almost any wording, so "I didn't understand" is rarely true. Its misses are different: the answer is not in what it learned, the question needs a person, or the visitor wants something done that it has no skill for. The fallback has to fit the miss and end in a next step.
Why "Sorry, I didn't understand" sends people away
People hate chat assistants that trap them, and they are right to. A visitor who got the same apology twice has learned that the chat is a wall; they call, email or buy elsewhere.
Classic builders cap the loop. WATI's help article on fallback messages (updated September 29, 2026, read November 3, 2026) lets you show the message 1, 2 or 3 times before the session ends. In Microsoft Copilot Studio, the default Fallback topic asks the user no more than twice to rephrase, then redirects to the Escalate topic, according to Microsoft Learn the same day. Both caps count how often the visitor's words matched no option or topic. An agent that answers from your content needs no counter: it says once what it cannot answer and moves the question on.
What is a fallback in AI?
In AI, a fallback is the planned behavior for when the main path fails: the model is unsure, a lookup returns nothing or a tool does not answer. For a customer-facing AI agent it decides what the visitor hears and where the question goes, instead of a guess. Four misses cover nearly every case:
- A knowledge gap. The fact is not in your pages, files or Q&A pairs. Say so, give the nearest fact, offer to forward.
- A figure it must not guess. Today's stock, a price for a custom job, a date that changes. No number; a forwarded request instead.
- A request only your team can handle. A quote, personal advice, a complaint. A teammate if one is online, else a forwarded request.
- A job without a skill. "Is there still room on Saturday?" needs a lookup in another system. Forward it today; give it a skill next week.
Chatbot fallback message examples for five situations
Take Fernhill Garden Center, a fictional family-run nursery that also runs weekend workshops. Its agent learned from the website, the care sheets and a few Q&A pairs, and a behavior makes it introduce itself as Fernhill's AI assistant. Each example is its first reply in a new website chat; the first also shows the generic reply it replaces.
The fact is not in its sources
A visitor asks: "Do you deliver bagged mulch to Lakeville?"
A generic bot says: "Sorry, I didn't understand that."
The agent replies: "I'm Fernhill's AI assistant. Our delivery area isn't in the information I have, so I won't guess. I can pass your question to the team so someone confirms it. Would you like that?"
A price or stock it must not guess
A visitor asks: "Do you have a 'Bloodgood' Japanese maple in a 15-gallon pot?"
The agent replies: "I'm Fernhill's AI assistant. We carry Japanese maples in spring, but I can't see today's stock, so I can't promise that size. Shall I pass your question to the team so they can check the yard?"
Advice only your team gives
A visitor asks: "Can you plan a shade bed for my backyard and tell me what it costs?"
The agent replies: "I'm Fernhill's AI assistant. A plan and price for your yard is something the team works out with you. Our care sheets suggest hostas, ferns and astilbes for shade. Want me to forward your request?"
A visitor asks for a person while someone is online
Tuesday afternoon; Dana has switched herself online in the portal.
A visitor asks: "Can I talk to someone?"
The agent replies: "I'm Fernhill's AI assistant. I'll bring in someone from the team now."
Once Dana clicks Join, the widget shows "Dana has joined the chat." (the name she set under Agent Identity), and she continues in the same window with the whole conversation.
The same request while nobody is online
Saturday morning in April: three people on the floor, nobody switched online.
A visitor asks: "Can I talk to someone about a bulk order?"
The agent replies: "I'm Fernhill's AI assistant. Nobody from the team is in the chat right now, but I can pass your request on so someone gets back to you. Would you like that?"
After a yes it asks for her name and email, sums up the request in one sentence and confirms: "Thanks, Maria. The team has your request."
What should a chatbot say when it doesn't know the answer?
When a chat assistant does not know the answer, it should say so in one short sentence, give the closest thing it does know, and offer one next step. Never a guessed price, date, stock level or opening time: a wrong answer costs more trust than an honest gap.
- Name the gap plainly. "That isn't in the information I have" beats "I'm sorry, I'm just a bot."
- Give the nearest fact. The plants you usually carry, the page on the topic.
- Offer one next step. A person if someone is online, a forwarded request if not, or a link.
Ultimo Bots answers from your content, not from guesses about your business, and you word its gaps yourself. In the AI Studio, open the Behaviors tab of the "Your agent" panel, click Add Behaviors, paste the text into "Describe how your bot should act" (up to 250 characters) and click Add. It is stored word for word and applies on every channel.
The first line deserves the same care: see our guide to the opening greeting, and our best practices for keeping answers right.
Who catches the question: a teammate online, or a forwarded request
A good fallback hands the question to someone, and who that is depends on whether your team is online. In Ultimo Bots the path is built in:
- A teammate is online. The website widget and chat link show a "Talk to a human" button. When a visitor asks for a person, the agent hands over right away; when it cannot help, it offers a person and hands over on a yes. Our human handoff guide covers the takeover itself.
- Nobody is online. The agent keeps answering. When it cannot answer reliably, or the visitor wants something only your team can give, it offers once to forward the request and asks for contact details only after the visitor says yes. The request lands in Leads with the full conversation, and by default the alert reaches you in the portal and by email (leads and analytics).
- You are online, but busy. If nobody joins within the wait time you set in the live chat settings (30 to 500 seconds), the widget tells the visitor that no one responded in time, and the AI assistant keeps helping.
- Teammate joins
- Request forwarded
The toggle "Offer proactively" on a lead skill's card in the AI Studio switches the offer off; the agent then forwards only when a visitor asks to be contacted or for a person. More on captured requests in lead generation with an AI agent.
When "Is there room on Saturday?" keeps coming back
At Fernhill, one spring week of forwarded requests tells the story: one maple, one delivery question, and the same question over and over, "Is there still room in Saturday's workshop?" The workshops are on the website, but tickets sell through Eventbrite and spots change by the hour, so the agent rightly would not guess. The owner opens the AI Studio and writes to Mo:
When a visitor asks about our workshops, look up our upcoming live events in Eventbrite and tell them the date, the time and whether tickets are still available, with the event link. If a workshop is sold out, say so and name the next one.
Ultimo Bots is the AI agent platform where you describe in plain words what your agent should do, and Mo, the builder in the AI Studio, builds it: skills on any public API your business already uses, on remote MCP servers, or on ready-made app actions. A question the agent forwards every week is a job waiting to be described.
Everything this job needs is public in the Eventbrite API v3 reference (read November 3, 2026): a private token sent as a bearer token, a list of an organization's live events by start date, and ticket classes whose sale status reads "AVAILABLE" or "SOLD_OUT". Mo asks for the token in a secure field, builds the skill on those two calls, tests it and reports what the agent can now do.
The next Saturday question gets an answer instead of a fallback: "I'm Fernhill's AI assistant. The Japanese maple pruning workshop is on Saturday, April 24, at 10 am, and tickets are still available. Here's the event page to book your spot." The date and status are invented here; in a real chat they come from Eventbrite.
Test answers and Mo's work use AI credits from one balance; Smart starts at $29 a month ($19 billed yearly) with 500 AI credits and 3 skills you add (pricing).
The same works for a class schedule or a seasonal event list: if the system has a public API with a key, describe the job to Mo (AI Studio).
Read your fallbacks every Friday: source, behavior or skill
Every fallback is a note about what your agent should learn next. Once a week, read the past days' conversations and the requests in Leads, mark each question the agent could not answer, and give it one of three fixes:
- A missing fact becomes a source. Add a Q&A pair or the page that answers it in the knowledge base. Fernhill added its delivery towns as one Q&A pair.
- A wrong tone or a guess becomes a behavior. If the agent apologized in circles or named a price it should not have, add a rule in your own words.
- A repeated "can you check" becomes a skill. When the answer lives in a system you already use, tell Mo what to look up.
Start today with the smallest version: in the Test tab, ask your agent the one question your website does not answer. If the reply guesses or apologizes in circles, paste the behavior above. Then switch yourself online, ask the widget for a person, and watch the handover arrive in your portal.
Questions and answers
What is a fallback message in a chatbot?
The reply a chat assistant gives when it cannot answer. In a decision-tree bot it fires when the words match no option ("Sorry, I didn't understand"). In an AI agent it fires when the answer is not in its sources or needs a person. A good one says so plainly, offers the closest fact and gives one next step.
What is a fallback in AI?
The planned behavior for when the main path fails: a model is unsure, a lookup returns nothing or a tool does not answer. For a customer-facing AI agent, the fallback decides what the visitor hears and where the question goes next: a person, a forwarded request or a link, instead of a guess.
What should a chatbot say when it doesn't know the answer?
Three things, in one or two sentences: that the answer is not in what it knows, the nearest fact it can give, and one next step, such as passing the question to the team. It should never guess a price, a date or stock. With Ultimo Bots you set this as a behavior in the AI Studio, in your own words.
How many times should a chatbot show a fallback before it hands over?
Once is usually enough for an AI agent. Decision-tree tools often repeat the fallback a few times, then end the session or escalate. Ultimo Bots offers to forward the request once per conversation, asks for contact details only after the visitor agrees, and hands over right away when a visitor asks for a person while a teammate is online.
What is a good chatbot fallback rate?
The figures vendors quote come from their own customers, so they rarely transfer to your site. Count your own: read a week of conversations and group the questions the agent could not answer. The ones that repeat are your to-do list: add the fact to your sources, add a behavior, or ask Mo to build a skill.
What are the four types of chatbots?
Lists differ, but a practical split is: menu or button bots, keyword or rule-based bots, AI assistants that answer from documents, and AI agents that also act, such as checking workshop availability. The first two need a classic fallback message; the last two need a plan for questions beyond their sources and skills.
