Conversation design for small business owners: seven rules for how your AI agent should talk, with before and after examples

You have written your agent's knowledge and the basics of its instructions, and the answers are correct, yet some chats still feel clumsy. That is a conversation design problem: who speaks when, how much, and what happens at the tricky moments. Here are the seven rules that fix most of it, each with a before and after chat and a line you can paste into your instructions.

By Ritchie, ReplyKit · Published 17 October 2026 · 9 min read

Quick answer

Conversation design is planning how a chat flows, not just what it says. For a small business agent, the basics are: answer the question before asking anything, ask one question per message, keep replies short, read details back before booking, let customers change their mind, admit when you do not know and offer a person, and end each chat with a clear next step.

What is conversation design?Rule 1: answer first, then askRule 2: one question per messageRule 3: short turns that fit the channelRule 4: read it back before you commitRule 5: let people go off scriptRule 6: admit the gap and open a doorRule 7: end with a next stepThe paste-ready instruction blockA five-chat testPutting it into ReplyKitFAQ

What is conversation design?

Conversation design is deciding how a chat between a customer and an automated assistant should unfold: the order of turns, the length of each message, where the assistant confirms, and how it recovers when things go sideways. Large companies have specialists for it. A small business owner can get most of the benefit from a handful of rules.

It matters because the failures people remember are rarely wrong facts; they are clumsy flows. In Nielsen Norman Group's study of chatbot usability, bots worked when users followed the expected path and broke down when users skipped steps, used unexpected wording or changed their minds. AI agents are far more flexible than the scripted bots in that study, but the same moments still need planning.

Your agent's knowledge decides what it knows. Its instructions decide how it behaves. If you have not written the instructions yet, start with our instruction template, then come back to tune the conversation.

Rule 1: answer first, then ask

When a customer asks a question, the next message must answer it. Questions back to the customer come after the answer, never instead of it.

Before

Customer: How much is a facial?
Agent: Happy to help! Can I have your name and which branch you prefer?

After

Customer: How much is a facial?
Agent: Our classic facial is $65 for 60 minutes. Would you like me to check a time for you?

Paste: "Always answer the customer's question in your first reply. Only ask for their details after you have answered, and only when needed for the next step."

Rule 2: one question per message

Ask one thing, wait for the answer, then ask the next. Customers on a phone answer the last question in a list and ignore the rest.

Before

Agent: Sure! What date would you like, what time, how many people, any dietary requirements, and is it for a special occasion?
Customer: Birthday

After

Agent: Sure. Which date are you thinking of?
Customer: Saturday 14th
Agent: Great. How many people?

The exception: two tightly linked details can share a turn ("Which day and roughly what time?"). Lead qualification follows the same rule; our lead qualification questions put the questions in a sensible order.

Paste: "Ask at most one question per message. If you need several details, ask for them one at a time, in this order: [list]."

Rule 3: short turns that fit the channel

On WhatsApp, a good reply is usually one to three short sentences. On a website, it can be a little longer because the visitor is reading on a page, but walls of text lose people everywhere.

SituationAim forAvoid
Simple fact (price, hours)One or two sentencesRepeating the question back, a sign-off paragraph
Comparing optionsA short list of two or three itemsEvery option on the menu
Instructions (preparation, directions)Numbered steps, three to fiveOne long paragraph
Bad news (fully booked, out of stock)The news, then an alternativeApologizing three times before getting to it

Paste: "Keep replies under 60 words unless the customer asks for detail. Use a short list only when comparing options. No greetings or sign-offs after the first message."

Rule 4: read it back before you commit

Before anything that is hard to undo, such as a booking or a recorded order, the agent repeats the key details and asks for a yes. This one rule prevents most "that's not what I booked" messages.

Before

Customer: ok the 3pm one
Agent: Booked! See you then.

After

Customer: ok the 3pm one
Agent: To confirm: scaling and polishing with Dr Tan, Thursday 16 October at 3pm, at our Bangsar clinic. Shall I book it?
Customer: yes
Agent: Done. You'll get a reminder the day before.

Read-backs should include the what, who, when and where, in that order, and nothing else. For how booking itself works behind the scenes, see how an agent books without double-booking.

Paste: "Before making a booking, repeat the service, staff member if any, date, time and location in one sentence, and only book after the customer confirms."

Rule 5: let people go off script

Customers change their minds, ask a side question mid-booking, or answer a question you have not asked yet. The agent should follow them, then return to the task without starting over.

Before

Agent: How many people?
Customer: 4. oh wait do you have parking?
Agent: Sorry, please tell me how many people.

After

Agent: How many people?
Customer: 4. oh wait do you have parking?
Agent: Yes, free parking behind the building. So that's a table for 4 on Saturday; what time suits you?

The same applies when a customer gives everything at once ("table for 4 Saturday 7pm"): the agent should use it all and skip the questions already answered.

Paste: "If the customer asks something else in the middle of a booking, answer it, then continue from where you were. Never ask for information the customer has already given."

Rule 6: admit the gap and open a door

When the agent does not know, it should say so plainly and offer a route to a person, not guess and not loop. NN/g calls this "owning the failure and offering an escape hatch", and the WhatsApp Business Messaging Policy requires businesses that automate to have "prompt, clear, and direct escalation paths".

Before

Customer: Can I bring my own cake for the party?
Agent: Yes, outside food is usually fine!

After

Customer: Can I bring my own cake for the party?
Agent: I don't have our policy on outside cakes, so I've asked the team. Someone will reply here shortly. Anything else I can check while you wait?

The "before" is the dangerous one: a confident guess that the business may have to honor. Write the full list of handover triggers using our escalation rules template.

Paste: "If the answer is not in your knowledge, say you don't have that information, tell the customer a team member will follow up, and do not guess. Hand over immediately if the customer is upset, asks for a person, or asks about refunds or complaints."

Rule 7: end with a next step

Every chat should finish with the customer knowing what happens next: they are booked, someone will call, or they have what they need. Avoid both the abrupt stop and the endless "anything else?".

Before

Customer: ok thanks
Agent: Is there anything else I can help you with today? We also offer massages, body scrubs, manicures, pedicures...

After

Customer: ok thanks
Agent: You're welcome. You're booked for Friday at 2pm; just reply here if anything changes.

Paste: "When the customer signals they are done, close in one sentence that restates the next step. Do not list other services unless they ask."

The paste-ready instruction block

Here are the seven rules as one block for your agent's instructions. Adjust the details in brackets.

Conversation rules:
1. Always answer the customer's question first. Ask for details only after answering, and only when needed.
2. Ask at most one question per message, in this order: [details you need].
3. Keep replies under 60 words unless asked for detail. No sign-offs after the first message.
4. Before booking, repeat service, staff, date, time and location, and book only after a yes.
5. If the customer goes off topic mid-task, answer, then continue. Never re-ask for given information.
6. If the answer is not in your knowledge, say so, tell them a team member will follow up, and do not guess. Hand over at once for complaints, refunds, upset customers or requests for a person.
7. When the customer is done, close in one sentence that restates the next step.

For voice and personality, which sit on top of these rules, see writing an agent persona.

A five-chat test

After changing the instructions, play the customer in five short chats. Each one targets a rule that commonly breaks.

  1. The direct question: "how much for a haircut". Pass if the price comes first (rule 1).
  2. The all-in-one: "need a table for 6 sat 8pm, one vegetarian". Pass if no question repeats a given detail (rules 2 and 5).
  3. The booking: pick a slot. Pass if it reads back details and waits for yes (rule 4).
  4. The unknown: ask something not in your knowledge. Pass if it admits it and hands over (rule 6).
  5. The goodbye: say "thanks that's all". Pass if it closes in one line (rule 7).

Putting it into ReplyKit

In ReplyKit, an agent's instructions are plain English, so the block above goes straight in. The rest of the design is built in: the agent answers only from your knowledge and says so when the answer is missing, you are notified when a chat needs a person and can take over with one click (the AI pauses for that chat) and hand back the same way, bookings use real open slots with double-booking protection, and it remembers the conversation history with each customer, so returning customers are not asked the same things twice. Questions it could not answer are listed for you to teach once.

Try the five-chat test on the live demo on our homepage, or start a free trial and run it on your own agent. Plans run from $29 to $99 a month; see pricing. If what you need is a button-menu flow with fixed paths and no AI, a visual flow builder is a better match than ReplyKit.

Frequently asked questions

What is conversation design in simple terms?

It is planning how a chat with an automated assistant should flow: when it answers, when it asks, how long each message is, where it confirms, and how it recovers when something unexpected happens. Good design makes a correct answer feel easy to get.

Do I need conversation design if I use an AI agent instead of a scripted bot?

Yes, but less of it. An AI agent handles unexpected wording far better than a scripted bot, but it still needs rules for turn-taking, confirmations and handovers, which you set in its instructions.

How many questions should a chatbot ask at once?

One, in most cases. Two closely linked details, such as day and time, can share a message. Lists of questions get partial answers on a phone.

How should a chatbot say it does not know something?

Plainly, without guessing, and with a next step: "I don't have that information, so I've asked the team; someone will reply here shortly." A route to a person matters more than the exact wording.

How long should chatbot replies be?

On WhatsApp, usually one to three short sentences, and under about 60 words unless the customer asks for detail. Website replies can be a little longer, but short lists beat paragraphs everywhere.

Should a chatbot confirm details before booking?

Always. A one-sentence read-back of the service, date, time and location, followed by a yes from the customer, prevents most booking mistakes and disputes.

Paste the seven rules into an agent and run the five-chat test

Plain-English instructions, your own knowledge, and you one click away. 7-day free trial, 300 AI replies, no credit card.

Start free
Written by Ritchie, ReplyKitPart of the small team in Malaysia that builds and runs ReplyKit. Writes about WhatsApp automation, AI customer service and small business lead response.

How we know this. These rules come from reading and adjusting conversations in ReplyKit test accounts, where the same few patterns (too many questions at once, bookings without a read-back, dead-end refusals) cause most of the friction we see. The research points are checked against Nielsen Norman Group's chatbot usability study and the WhatsApp Business Messaging Policy, both opened in October 2026. All example chats are written by us for this article and are hypothetical. We make ReplyKit, so weigh our view accordingly. If you run a scripted button-menu bot rather than an AI agent, a dedicated flow-builder tool and its own design guides will serve you better than this article. About ReplyKit.

Sources and further reading

  1. Nielsen Norman Group, The User Experience of Chatbots (Raluca Budiu, 2018) · bots work on linear flows and break when users stray; recommends owning failures and offering an escape route
  2. WhatsApp Business Messaging Policy · automated experiences must have prompt, clear and direct escalation paths

ReplyKit is an independent product and is not affiliated with or endorsed by Meta or WhatsApp. Third-party pages were checked in October 2026 and may change. Example conversations are hypothetical.