Conversational AI is technology that lets a customer write to your business in ordinary language and get an answer, or a completed task, back across as many turns as the conversation needs. In customer service it replaces menus and keyword matching with a model that understands what was meant, asks about what is missing, and uses your own information to respond.
The model is the part everyone talks about, but the quality of the customer's experience depends on the conversation design around it. This guide covers the parts of a system, the design principles that separate a good one from a frustrating one, a worked example, and the measures to track.
The five parts of a conversational AI system
The language model: reads the customer's message, in any phrasing or language, and writes the reply.
Knowledge: your prices, policies and product facts, retrieved when relevant so the reply is grounded in what is true for your business.
Memory: what was said earlier in this conversation, and what is already known about the customer, so the system does not ask twice.
Connections: the calendar, store, CRM or order system the conversation needs to read from or change.
Channels and handover: where the conversation takes place, and the route to a person when it needs one.
Weakness in any of the five shows up as a customer complaint. A fluent model with no knowledge invents answers. Good knowledge with no memory repeats questions. Memory with no connections cannot finish anything. A model with no handover leaves customers stuck. When you evaluate a product, ask about all five parts, the model included.
Where it came from, and why it matters now
Earlier customer chat tools matched keywords or followed trees that someone drew, so they worked only for the questions their builders predicted. Language models changed that by handling phrasing nobody planned for. They also changed the risk: a tree fails by saying "I did not understand," while a model can fail by saying something confident and wrong. Design for that difference. The comparison of AI agents and chatbots explains the levels in more detail.
Eight rules of conversation design
Answer first, then ask. If the customer asks about opening hours, give them before asking for their name.
One question at a time. A message with four questions gets one answer and three silences.
Never ask for what you know. If the customer already gave a date, or the contact record has their email, use it.
Mirror the customer's language and length. A one-line question deserves a short answer. A reply in the customer's language, including switching mid-chat, is basic courtesy.
Confirm before acting. "I will move your booking from Thursday 10:00 to Saturday 11:00. Shall I go ahead?" is cheap insurance.
Say what happens next. After any action, name it: what was done, what the customer will receive, and when.
Say when you do not know. "I do not have that, so I am asking the team and you will hear back today" beats an invented answer.
Hand over early and cleanly. Tell the customer who is taking over, and make sure that person sees the whole thread.
A worked conversation
A patient writes to a clinic with three branches: "Hi, can I see Dr Tan this week? My back has been bad since Monday."
The system understands three things at once: a named physician, a time frame, and a symptom. It does not ask for any of them again.
It checks which branch Dr Tan works at on which days, and finds two open slots at the branch the patient visited last time.
It replies: "Dr Tan has Wednesday 15:00 or Friday 10:30 at the Central branch. Which suits you?"
The patient picks Friday. The system books it, adds the symptom to the record, sends the confirmation and notes that a reminder will arrive the day before.
If the patient had mentioned chest pain or something similar, the system would not have booked at all. It would have said that this needs a person's attention now and handed the conversation over at once, with the message attached.
The last point is the part that distinguishes a safe design from a merely smooth one: define in advance the phrases and situations that trigger handover, and test them.
Channel behavior
WhatsApp: short messages, quick replies, voice notes and photos from customers. Free-form replies are allowed within 24 hours of the customer's last message, and approved templates are needed after that. See the 24-hour rule guide.
Web chat: visitors are anonymous at first, so asking for a phone number or email early lets the conversation continue elsewhere. Cards, such as time slots to pick from, shorten bookings.
Email: longer, more formal replies, with attachments the system should read.
Instagram and Messenger: casual, often prompted by a post or an ad, and frequently pre-sale.
One agent and one inbox across these channels keeps the customer's history together, whichever channel they use next time.
What to measure
| Measure | What to look for |
|---|---|
| Resolution without a person | Rising over time, without rising complaints |
| Turns per resolved conversation | Fewer turns usually means better questions |
| Repeat questions | Cases where the system asked something already known |
| Handover quality | Whether staff needed to ask the customer to repeat |
| Satisfaction on AI-handled chats | Compared with human-handled ones |
| Wrong answers in a weekly sample | The number that stays honest |
Testing a conversation before customers see it
Write twenty test scripts. Each script is a short exchange with a goal, such as "change a booking and ask about a deposit", and a twist, such as the customer changing their mind halfway or answering a different question from the one asked. Run each script against the system and mark where it drifts.
Three scripts with a customer who writes in a second language, or mixes two in one message.
Three with an upset customer, to test tone and handover.
Three with a request outside what the system should do, to test refusal.
Three with missing information, to test that it asks and does not guess.
The rest with ordinary requests phrased badly: no capital letters, half a sentence, a photo instead of words.
Keep the scripts. Each time you change the knowledge or the instructions, run them again, and add any real conversation that went wrong. After a few months they are a regression test for the whole experience, and running them takes minutes.
Common failures and fixes
Too chatty. Long preambles and repeated apologies. Fix the instruction to keep replies short and drop the filler.
Too eager. Booking before the customer has chosen. Add a confirmation step.
Too stubborn. Refusing to hand over. Add explicit handover triggers and test them with annoyed messages.
Out of date. Answers from last year's price list. Keep the knowledge current and review it monthly, as set out in the AI knowledge base guide.
A clinic group that handled a flood of WhatsApp enquiries
Yong Kang TCM Clinic, a multi-branch traditional Chinese medicine clinic, saw its WhatsApp messages grow faster than its staff could reply. Founder Elson described it in the case study: "Simple WhatsApp queries turned into an overwhelming flood of inquiries. With multiple branches, our staff struggled to keep track of our diverse offerings, especially when it came to answering customer queries." Volty now routes each WhatsApp inquiry to the right branch, physician and slot, hands over to reception with the history, and answers Messenger and Instagram too. The case study reports booking resolution time at peak down from 30 minutes to under 5, bookings up 20%, and no headcount growth despite 60% growth in chat traffic. Read the Yong Kang TCM case study.

Doing this in Volty
Volty is built around the five parts: models routed to the right task, knowledge from your website and files, memory in the contact record, connections to calendars, CRMs, stores and any MCP server, and one shared inbox with handover rules your staff set. See Volty's WhatsApp AI page and the web chat page; every plan begins with a 7-day free trial.
Frequently asked questions
What is conversational AI in customer service?
It is AI that lets customers write or speak in ordinary language and carries the conversation across several turns: understanding the request, asking what is missing, answering from your information and, in the stronger systems, acting in your tools. It replaces menus and keyword matching with understanding.
What is the difference between conversational AI and a chatbot?
A chatbot is one kind of product built with conversational AI, and older chatbots followed scripts. Conversational AI describes the capability, which can also power an agent that books, looks up and updates records.
What are the main parts of a conversational AI system?
A language model that understands and writes, a knowledge source to answer from, memory of the conversation and the customer, connections to your tools, and channels and handover rules around them.
How do you design a good conversation for customer service?
Answer first, ask one question at a time, never ask for what you already know, confirm before an action, and hand over early when the customer is upset or the request is outside what the agent should do.
Which channels suit conversational AI?
Messaging channels where customers already write, such as WhatsApp, Instagram, Messenger, web chat and email. Volty runs one agent across all five from one inbox.










