A chatbot that doesn't know the answer has four honest moves: refuse, ask a clarifying question, show the source it did find, or take contact details for a person. A confident guess is the only genuinely expensive option, because the customer acts on it.
Every chatbot spends part of its day outside what it knows. The design question is not how to shrink that part to zero, which is not available. It is what the bot does when it gets there.
Most teams never decide. They tune the answers and let the fallback be whatever the model produces under pressure, which is a fluent paragraph of nothing, or worse, a fluent paragraph of something.
Four ways of not knowing, which need four different responses
The answer was never published. Your delivery radius exists in the owner's head. No amount of retrieval will find it. The bot should say the material does not cover it and move to collecting a contact.
The question is ambiguous. "What does the annual plan cost" from a visitor who has not said which plan. Guessing which one they meant produces a right answer to the wrong question, which is indistinguishable from a wrong answer once they act on it. Ask which one.
The question is out of scope. A visitor asks a plumbing supplier to diagnose their boiler. The bot has no business answering, even if it could. Say what the business does cover.
The answer needs a human's judgment. Custom quotes, complaints, anything where the honest response is "someone has to look at it". Recognizing this class early is what stops a standard published rate being stretched over a job it was never meant to cover.
Treating all four as one "I'm sorry, I don't have that information" is why fallbacks read as broken. Three of the four have a better move available.
Why the confident guess is the expensive option
A refusal costs you a small amount of goodwill from one visitor.
A wrong answer gets acted on. A customer told there is a fourteen-day return window plans around it. A customer told the shop is open on Sunday drives across Chandigarh. A customer quoted ₹18,000 for a job that costs ₹26,000 arrives believing they have a price, and you now choose between eating the difference and explaining that your own website misled them. Both cost more than the enquiry was worth.
The asymmetry is sharper for small businesses than for large ones. A bank with a call center absorbs a wrong answer through a complaints process. A six-person firm absorbs it out of the owner's margin, and the customer tells people.
There is a second cost that shows up later. A bot that guesses is a bot you cannot trust to run unattended, so somebody starts reading transcripts every morning, and the tool you bought to save labor has quietly created some.
What a good refusal actually says
Four parts, in this order.
Name the gap plainly. "I can't find our pricing for a two-bedroom flat on the site" beats "I'm sorry, I don't have that information", because it tells the visitor what went wrong.
Offer the nearest published thing. If the exact rate is missing but the site states a starting price and what it includes, give that and label it as what it is.
Show where it came from. When an answer does come from your content, the visitor should be able to see which passage produced it. A source link converts a claim into something checkable, and it also lets you trace a strange reply back to the page that caused it. It does not make the answer correct, which is a separate problem.
Then hand over. On a business site at eleven at night, that means taking a name and a phone number, not routing to a live agent who does not exist.
Handoff without a support team
"Human handoff" in enterprise chat means transferring a conversation to a staffed queue. Almost no Indian small business has one, and pretending otherwise produces the worst pattern in the category: a bot that promises "let me connect you to an agent" into an empty room.
The honest version is smaller and works better. The bot says it cannot answer this one, asks for a name and number, confirms that someone will call, and ends. The visitor knows exactly what happens next. Chatterbox does this, and it is deliberately phone-first, because an Indian customer will hand over a number sooner than an email and you were going to call rather than write.
The obligation moves to you at that point. A captured lead nobody calls back is worse than no widget at all, because the customer believes they are in the queue.
Read the refusals
Every refusal is a question a customer asked that your website does not answer, and they arrive already sorted by frequency.
Go through a month of them and you get a content backlog written by your own customers. The top five are usually a delivery area, a price band, a timing, a policy detail and something about payment. Publish those five answers, and next month's refusal list is a different, shorter set of questions.
That loop is the practical next step here. Not a better model, not a longer prompt: read what your bot could not answer, and answer it on the site. The content checklist is the same job done in advance.