Writing the way your chatbot should speak

Most of what a chatbot says is decided by the facts you give it. The part you control separately is how it sounds, and the sentences worth writing carefully are not the greetings - they are the refusals, the bad news and the handovers. This guide is about those.

One principle: write what you would say on the phone

Read your knowledge out loud. If a sentence is something you would never say to a customer standing in front of you, it is wrong for a chatbot too. Marketing copy does not survive the transition: "We pride ourselves on a customer-focused philosophy" is unanswerable, while "We aim to come back to you the same working day" is a sentence a person can rely on.

The refusals are the hardest and the most important

Every business has questions its chatbot must not answer. A refusal can be written three ways, and only one of them keeps the customer.

The useful version does three things: it says no, it says why in one clause, and it offers the next step. Write every refusal in that shape and the chatbot's limits stop feeling like failures.

Say the bad news early and plainly

The sentences businesses most want to hide are the ones that convert best when stated clearly: the call-out charge, the minimum order, the fact that you do not cover that town, the lead time. A visitor who finds out at the end feels misled; one who is told in the second message either leaves - which costs you nothing - or carries on knowing the terms. Put the money sentence in the knowledge in the exact words you want quoted, because the assistant will quote exactly that.

Write the clarifying questions as questions, not as a form

An assistant asks up to three short questions, one at a time. How you word them decides whether people answer. Ask for one thing, give the reason, and allow "I don't know":

• Weak: "Please provide your project details and budget."
• Better: "Roughly what are you hoping to spend? A range is fine - it tells us which of our options to talk about, and it's fine to say you're not sure yet."

The permission to be unsure is what makes people answer the uncomfortable question instead of closing the tab.

Short, but not clipped

Two or three sentences is right for chat. A six-line paragraph is not read on a phone, and a four-word answer reads as impatience. If an answer genuinely needs a list, write it as a list in the knowledge and it will come out as one.

Personality: a little, and never in the wrong place

A warm first line is fine. Jokes are not, and the reason is that you cannot predict the context: the same quip lands differently for somebody whose basement is flooding, who has had a bereavement, or who is reporting a problem with your work. Keep any personality in the greeting and the small talk, and make the operational sentences plain. If your business has a genuine voice - a brewery, a tattoo studio - it can carry further, but the refusals and the money sentences stay plain in every business.

The handover sentence deserves its own draft

"Let me get someone" is where many businesses lose the thread. Say what will happen, who will do it, and when - and never promise a timescale your team will not meet. "I'll pass this to the team now. Someone will come back to you by the end of the next working day, and they'll have everything you've told me here" is a sentence the assistant can keep, because it is true by design: the whole conversation is visible to whoever picks it up. See handing a chat to a person.

Tone in other languages is not translation

A sentence that reads as friendly in American English can read as presumptuous in German or curt in Japanese. Formality is the usual trap - several languages distinguish a formal and informal "you", and picking the wrong one is noticeable in a way no English speaker would detect. If a language matters commercially to you, have the wording read by somebody who speaks it. See choosing your languages.

Test it by reading the transcripts

After a week, read what the assistant actually said rather than what you wrote. The gap between the two is the work. The questions it could not answer are listed for you, and correcting an answer creates a new version of the knowledge, so a fix is permanent. See chatbot best practices.