Chatbot Conversation Design: Writing Scripts and Prompts That Convert

One sentence, three times the replies

We changed a single line in a service business chatbot. The original opening was "Hi! How can I help you today?"

The replacement was: "Hi — I can check availability or answer a question about pricing. Which one?"

Replies to the first message went up by a factor of roughly three. Same platform, same model, same knowledge base. The difference was that the second version told the visitor what the assistant could actually do, and gave them two ways to start.

Chatbot conversation design is mostly this: removing the moment where someone has to invent what to type.

Design the job before the greeting

A booking assistant, a lead qualifier, and a support deflector need different flows. A bot attempting all three does each of them badly, and the visitor feels it as friction without being able to name it.

Write the one job at the top of the document. Then apply it as a filter: any question that doesn't move toward that job gets cut, and any path that drifts away from it gets removed.

The four-part shape that works

The opening

Who's speaking, what they can do, and a reason to spend thirty seconds. Two sentences. Specific beats friendly.

"We can usually book an appointment in under a minute — are you a new or existing patient?" is better than any variation of "Welcome! I'm here to assist".

The qualifying question

One question that lets the visitor describe their situation, either through options or free text. Service businesses generally need what they want and when.

The question should feel like the start of getting something done, not an intake form. "What needs fixing?" is better than "Please select a service category".

The value moment

Before asking for anything, give something: a real answer, a price range, an availability check, an honest explanation of the process.

This is the moment that separates an assistant from a form with extra steps, and it's the one most bots skip because the flow was designed around data collection.

The ask

One action, phrased as the natural next step. "Want me to book you in for Thursday at 10?" rather than "Please provide your contact details so that a member of our team may reach out."

One ask per conversation. Two competing asks produce no action.

Rules for the individual message

Two or three lines maximum. Anything longer reads as a document, and people skim documents rather than replying to them. If detail is needed, put it behind a link or save it for the follow-up.

One question per message. Three questions get one answer, usually the wrong one. Sequence them so each answer sets up the next question naturally.

Buttons where answers are predictable. Typing introduces ambiguity and effort. Free text where the visitor needs to explain, buttons where the answers are known.

Name the next step in every message. Nobody should have to wonder what happens if they reply. "I'll check Thursday and Friday — one sec" is better than a bare "Checking availability".

Read each message aloud. Scripts that look fine on screen often sound like a form when spoken. Anything you wouldn't say to a customer at a counter should be rewritten.

Tone, and the one disclosure that matters

Say it's automated, in one plain line. Not buried in a terms page, not implied by an "AI-powered" badge. It costs nothing and prevents the specific resentment that arrives when someone realises later that they poured their frustration into a bot that didn't say so.

Mirror the visitor's register. Short casual messages get short casual replies. Formal messages get formal replies. This single behaviour accounts for most of the difference between bots described as "surprisingly good" and bots described as "robotic".

No cheerfulness in the face of a problem. "I'm sorry about that — let's get it sorted. Can you tell me the order number?" is right. An emoji and "I'd be happy to help!" is not.

Never guess about money, legal, or health. Answer from an approved source or hand over. A confident wrong figure here costs more than any delay.

The hard cases, scripted

These are where conversation design is actually tested, and where most bots are left to improvise.

The angry visitor

Frustrated people are trying to resolve something that has already failed once. Acknowledge it plainly, no defensiveness, no exclamation marks, and go straight to a resolution or a human. Do not ask them to rate their experience.

Out of scope

Name the limit and offer a real alternative: a person, a number, an email, a callback. "I can't change a booking made last month, but Sam can — shall I have him call you?" reads as competence. A plausible-sounding invention reads as a scam.

The loop

If the visitor repeats themselves or says "that's not what I asked", that's the signal to escalate, not to rephrase. Script an explicit trigger rather than relying on the model to notice.

Silence and typos

People type "tommorow", "dr's apt", "how mutch". Design for it by not depending on exact words. And if the visitor says nothing for two messages, offer a human rather than sending a third prompt.

The handoff is part of the script

Decide in advance:

• Which topics escalate, as a written list: complaints, refunds above a threshold, anything legal, existing customers with an open issue, safety.
• What context transfers. The human should receive the conversation, the name, the account if known, and what's already been tried. If the customer has to explain twice, automation failed.
• What the visitor is told. "I'm passing this to Priya, who handles accounts — she'll call you within the hour." Specific, and only if it's true.
• What happens out of hours. Confirm it's logged, state when someone will reply, and mean it.

Testing before launch

Five people who don't work for you, each with a real task: book an appointment, ask a price, complain about something. Watch where they hesitate. Hesitation is where the script is wrong, and it's usually in the opening line or a question that assumes knowledge.

Then keep testing. Ten transcripts a week, read properly. You're looking for three things: questions it couldn't answer, answers that were plausible but wrong, and moments where a person would have changed the subject.

Where to start

Rewrite your opening line so it names what the assistant can do, and cut every question from your flow that doesn't change the outcome. Those two changes take an hour and typically move completion more than any model upgrade.

If the underlying site is where visitors are losing interest, a free digital presence audit on Webcheckhub scores conversion readiness and website health alongside Google presence, so you can see whether the bot is fixing a problem or inheriting one. When you want the flow designed, scripted, and tested against real conversations, Solutions Elites builds customer-facing assistants for service businesses.