Conversational AI Design for Sales: Flows That Convert

Conversational AI design decides what an AI agent says, how it makes a call, when it asks a question, when it hands over to a person and how it recovers from a miss. If you run Google or Meta ads, this is what turns an enquiry into a booked meeting. Get it wrong and the lead goes cold in your inbox.

This guide is for founders and heads of sales at $1M to $10M service businesses. It covers the principles that hold up, how to map a flow from first message to booked call, how to handle errors and handoffs, and how to test the result. The aim is simple: qualify and book the leads your ads already paid for.

What conversational AI design covers

Most people hear "conversation design" and picture a script. A script is the smallest part. Design covers the agent's job, its tone, which questions it asks and in what order, how it copes with an answer it did not expect, when a human takes over and how every exchange lands back in your CRM.

Yvonne Gando, a senior director of user experience at Salesforce, says conversation design "has more in common with product architecture than copy editing". It shapes how an agent reasons and how it handles uncertainty (Salesforce Australia). That suits sales. A qualification chat is a small workflow with a start, a decision and an outcome. The wording sits on top. The same thinking applies outside chat too. This piece on AI sales quoting needing human checks shows it applied to quotes.

Four decisions matter most for a service business:

• The job. Qualify, book, wake up a dormant lead or answer a pricing question. Pick one per conversation.
• The voice. Plain and Australian, like your best salesperson talking to a new enquiry.
• The rules. What counts as qualified, which answers rule a lead out and which go to a person.
• The recovery. What happens when the lead says something the agent did not expect.

Conversational AI design vs chatbot conversation design

People search for both phrases and they overlap. The real difference is scope. A chatbot usually lives in a text window. The user sends a message, the bot replies. Conversational AI design also covers voice, where the system has to handle turn-taking, interruptions and silence (Featurebase covers this well).

In sales, one lead might fill in a Meta lead form, get an SMS within a minute, reply by text, then take a call from a voice agent. One design has to hold across all of it.

Njin builds this layer as part of the wider sales system. For the build side, read our conversational AI for sales implementation guide or see the Conversational AI Agent. For a broader view, Devwiz has a good business guide to conversational AI.

Principles for a clear, natural conversation

You do not need to invent these. Several large design systems publish guidance that points the same way.

Set expectations early. Microsoft's guidance says to clarify what the AI can do and to expect the same intent in many phrasings (Microsoft Learn). For a sales agent that is one line. Who it is, what it can do (check your enquiry, book a time) and that a person is available.

Keep the purpose clear and the control with the user. Maersk's design system says to let users clarify or escalate to a human, to keep sentences short and to avoid implying the AI is human or has feelings (Maersk Design System). A lead who thinks they were tricked into talking to a person rarely books.

Be cooperative. Say true things, give the right amount of detail and stay on topic. ServiceNow publishes its own conversation design guidelines if you want a second view. In practice it means one question per message, no walls of text and no answers to questions nobody asked.

Treat privacy as part of the design. Maersk's guidance covers showing users that sensitive data is handled securely. If your agent collects contact details or a budget, say why you ask and what happens next.

Make it accessible. UXMatters recommends text alternatives for voice interactions and voice input for people who cannot type (UXMatters). Test the chat widget itself, not just the page around it. Devwiz covers this in more depth in its post on accessibility for chatbots.

How to map a flow from first message to booked meeting

Google's design guidance for business messaging (see the Google for Developers messaging docs) is a good place to start reading. The order of work below is our own, adapted for an ad-driven service business.

1. Pick one outcome. For most service businesses it is a booked discovery call with a qualified lead. Write it down first.
2. List your qualifying criteria. Need, budget, timing and decision authority are the usual four. We build qualification around the BANT framework, then tune the thresholds to your business.
3. Write the ideal chat as a dialogue. Read it aloud. If a real salesperson would never say a line, cut it.
4. Draw the flow. Mark each decision point: qualified, not qualified, unclear, wants a person now.
5. Add suggested replies. Buttons or short prompts cut typing and keep leads on the path.
6. Write the fallbacks and the handoff. More on that below.
7. Connect it to the CRM. Every answer should land in a field, so your team sees what the agent learnt before they call.
8. Test, launch and keep tuning.

Want a head start? Our conversation flow templates give you ready-made flows to adapt.

A qualify-and-book flow, step by step

One published framework for qualifying leads is to identify intent, ask one or two qualifying questions, then route the lead or answer. Booking adds confirming the service, offering time slots, collecting contact details and confirming the booking (Omago). Here is how that looks for a lead who just clicked a Google ad for an accounting advisory service:

• Opening: thank them for the enquiry and name the service they asked about.
• Intent check: confirm what they want help with, in their own words.
• Question one: timing, such as when they need this sorted.
• Question two: size or scope, such as the number of entities or staff.
• Route: qualified leads see available times. Unclear answers get a follow-up question. Poor fits get a polite, helpful close.
• Confirmation: the time, who they will speak to and what to prepare.

Keep it to two questions before you offer a time. Each extra question is another chance to lose someone who clicked an ad a minute ago. We go deeper on question design in our guide to chatbot lead qualification.

Email enquiries follow the same logic

Not every lead arrives by chat. Plenty come by email, and the same thinking applies. Identify intent, pull out the key details, reply fast and route what needs a person. This overview of inquiry response automation covers the email side.

Fallbacks, errors and handoff to a person

This is where most designs fail. The happy path is the only one anyone wrote down.

Microsoft's guidance says unexpected, unclear or off-topic input should trigger a clarifying question, alternative suggestions or a redirect, not a dead end (Microsoft Learn). Gando at Salesforce puts it bluntly: expect that conversations will not go as planned.

Our rule of thumb for sales is a short ladder:

1. First miss: rephrase the question and offer two or three suggested answers.
2. Second miss: own the confusion and offer a person or a callback time.
3. Handoff: pass the full transcript and captured fields to the CRM, so the salesperson never asks the lead to repeat themselves.

Some moments skip the ladder and go straight to a human. A complaint, a high-value account, an angry message or a lead who asks for a person. Write those triggers in as hard rules. Our post on chatbot human handoff covers the routing in detail, and AI Orchestrators has a useful take on AI agent handoff.

Voice raises the stakes. There is no screen to fall back on, so interruptions, silence and background noise all need a plan. Our voice AI sales assistant guide walks through it.

Write the agent's instructions with the same care

The design only works if the instructions behind it are clear. Vague instructions give vague behaviour. Give the role, goal, allowed actions, banned topics and escalation rules a line each. We unpack what good instructions look like in our post on lessons from a leaked system prompt.

Also keep a short list of things the agent must never do. Quoting prices that are not published, promising outcomes and giving regulated advice are the usual ones. For a professional services firm, that list matters as much as the script.

Test and measure the design

Launch is where the work starts. Two Australian examples show what measurement looks like.

Accuracy on average is not enough. iTnews reported in 2018 that UBank saw about a 15 percent uplift in conversion for customers who used its Robochat, and that it answered correctly on the first attempt about 80 percent of the time. The same report quoted UBank's head of digital saying some customers did far worse, depending on their grammar, whether English is their second language or the slang they use (iTnews). So test with colloquial and misspelt input, not just tidy questions.

Test again and again. Inside Retail Australia reported that retail AI tool Preezie runs roughly 500 tests before a retailer validates and goes live, and that its A/B tests typically show 5 per cent to 10 per cent incremental revenue (Inside Retail Australia). That is retail, not services. Treat it as an example of method, not a benchmark for your business. Devwiz also has a practical post on conversational UI design.

For a service business, these are the numbers to watch:

• Reply time from enquiry to first message.
• Engagement rate, meaning the share of leads who answer the first question.
• Qualification rate and booking rate from engaged leads.
• Show rate for booked meetings.
• Handoff rate and the reasons behind it.
• Fallback rate by question, to find the lines that confuse people.
• Pipeline value and cost per booked meeting, tied back to the campaign that sourced the lead.

Read real transcripts every week for the first month. Change one thing at a time so you know what moved the numbers. A dashboard that ties conversations to ad spend makes this routine.

Njin designs and runs AI sales systems inside the CRM you already use. Start with the Lead Qualification Agent if qualification is where your leads leak, or the Instant Reply Agent if slow first replies are the problem.

Common questions

What is the difference between conversation design and conversational AI design?

Conversation design shapes how an AI and a person trade messages. Conversational AI design adds the reasoning, the actions the agent can take and the channels it works across, such as chat, SMS and voice. For sales, the second view is the useful one. The agent has to do something at the end, like book a meeting.

How many questions should a sales chatbot ask before booking a meeting?

One published lead qualification framework uses one or two qualifying questions before it routes the lead or answers (Omago). That is a good start for ad-driven leads. Add a question only if it changes who takes the call or whether the meeting happens.

When should a chatbot hand over to a person?

Our rule is after two misses in a row. Hand over straight away when a lead asks for a person, raises a complaint or is a high-value account.

How do you test whether a conversational flow is working?

Track booking rate, qualification rate, fallback rate and handoff reasons. Read real transcripts each week. Change one element at a time and compare. Test with the slang and misspellings your leads really use, since accuracy on clean questions can hide problems.

Take the AI Consulting Fit Scorecard to see where an AI sales conversation would pay off first in your business. It takes five minutes.