In Singapore, most customer enquiries arrive on WhatsApp. That is why so many owners end up searching for a WhatsApp chatbot — something to handle the same twenty questions about price, hours and availability while they are busy doing the work. The trouble is that the word covers everything from a keyword auto-reply that irritates everyone who meets it to an assistant that reads your own price list and books the appointment.
This is a guide to telling those apart, picking the one that suits your business, and setting it up without the usual mistakes. It is for small businesses — salons, clinics, tuition centres, contractors, agents — not enterprise contact centres. Parts of it will tell you that you need less than you think.
Why WhatsApp is the default business channel here
Nearly everyone here already has WhatsApp open. Customers do not want a contact form, an email thread or another app — they send a message and expect an answer. For a small business that is both a gift and a burden: enquiries arrive constantly, at all hours.
It also means the bar is set by chat, not by email. A four-hour reply reads as neglect on WhatsApp, and most enquiries are comparison shopping — the customer has messaged three businesses and is waiting to see who answers. We have written separately on why response time decides who wins; the first useful reply usually takes the job.
Three kinds of WhatsApp bot, and which one you need
Keyword bots watch for a trigger word and fire back a canned message. Type "price" and you get the price list. They break the moment someone writes the way people actually write: "hi how much for full set ah, can do this Sat?" The result is the worst kind of experience — obviously automated, and answering the wrong thing.
Menu or flow bots offer numbered options: press 1 for prices, 2 for hours, 3 to speak to someone. These are honest about what they are, and for a business with a genuinely small set of questions — fixed hours, fixed menu, one location — a well-built flow bot is often enough. There is no sense paying for intelligence you will not use.
AI chatbots read the message as written and answer from your own information. They cope with typos, Singlish, a sentence that switches language halfway, and enquiries that bundle two things at once. You want one when the answer depends on the situation: which service, which branch, which staff member.
A useful test: write down the ten questions you were asked most this week. If nine have one fixed answer, a flow bot will do. If most begin with "it depends", you want an AI chatbot.
What a WhatsApp chatbot should be able to do
Once you are comparing AI tools the demos look identical. These are the questions that separate them.
- Does it answer from your prices and policies, or from general knowledge? A bot that invents a price is worse than no bot, so ask how your documents get in and how you update them.
- Can it book, or only talk about booking? Writing the appointment into your calendar is where the time saving lives, and a clinic with three doctors on one calendar needs more than a one-person studio does.
- What happens when it does not know? A good assistant recognises the edge of its competence and hands you the conversation instead of guessing and leaving you to apologise.
- Which languages does it handle, and how naturally? Conversations here move between English, Mandarin, Malay and Tamil without warning, and a stiff translated reply is obvious.
- Where is the data processed, and what is kept? Ask whether NRIC and card numbers are stripped before storage, how long messages are retained, and whether your conversations train someone else's model — our plain-English PDPA checklist covers the rest.
The Business App and the Business Platform are not the same thing
The WhatsApp Business App is the free app you install on a phone. It gives you a business profile, labels, quick replies and a greeting message. What it does not give you is a way for outside software to read and send your messages. Anything marketed as a chatbot for the app is usually a workaround, or the away-message feature under a nicer name.
The WhatsApp Business Platform — what most people call the API — is the official route for software to send and receive on your number. There is no app to open; you connect the number to a provider and your messages flow through it. Any real WhatsApp chatbot runs on this. A number can generally only be attached to one setup at a time, so if you also want to keep replying by hand from your phone, ask the provider whether that is supported.
How to set up a WhatsApp chatbot, step by step
The mechanics vary by provider — we walk through the connection itself in how to add AI to WhatsApp — but the sequence is the same everywhere.
- Choose the number first. Use one that is not tied to a personal WhatsApp account, or be ready to migrate the number your customers already have. Settle this early, because changing a published number later costs you enquiries for months.
- Connect it to a provider. This registers your number on the WhatsApp Business Platform. Expect a verification stage involving your business details, and set aside an afternoon.
- Feed it your real information. Prices, packages, hours, service areas, cancellation policy, and the things you do not do. Owners underestimate this, then blame the AI for vague answers. If the tool can read documents you already keep, point it at that folder instead of retyping everything.
- Connect your calendar and set the booking rules. Buffer between jobs, how far ahead people may book, which staff member does what, whether two bookings can share a slot. If you have no online booking system yet, sort that out first.
- Write the handover rule down. Decide what the bot must never handle alone — complaints, refunds, medical or legal questions, anything involving a discount — and let it attempt the rest.
- Test with real questions, not polite ones. Take fifty genuine messages from your chat history and send them through: the blunt ones, the vague ones, and the ones that switch language mid-sentence.
- Read the first week's transcripts. Every conversation, for one week, correcting what went wrong — usually by fixing the underlying information rather than the bot.
Most bad chatbot answers are not an AI problem. They are a documentation problem — the bot is repeating a gap that was already there.
Common mistakes
- Automating everything on day one. Start with prices, hours and availability, keep the judgement calls, and widen the scope as you watch it perform.
- Pretending it is a person. Customers work it out and resent it; a plain line saying an assistant is replying and a human can take over costs you nothing.
- Loading it with marketing copy instead of an operational price list. Brochure language produces brochure answers, and someone asking about price wants a number.
- Setting it live and never looking again. It goes out of date the way a printed menu does.
How to tell whether it is working
Judge it on outcomes you already care about, not on how clever the replies sound. Four numbers are enough: how many enquiries got a useful answer without you touching your phone, how many became bookings, how fast the first reply was, and how many conversations were handed to you.
If handovers fall week on week while bookings hold steady or rise, it is working. If handovers stay flat, the cause is almost always missing information rather than a weak model. If bookings rise but complaints rise too, you have automated something that needed a person — put that topic back on the handover list. Keep reading a few transcripts a week either way. You will find the question you never realised customers ask.
Where microcrew fits
microcrew is our own answer to this: an AI assistant for Singapore small businesses that replies to WhatsApp and Telegram in about 30 seconds, around the clock, in the customer's language. It answers from your own prices, policies and documents by indexing the Google Drive files you already keep, and books into Google Calendar — several staff on one calendar, multiple branches, and capacity-style slots for places that seat more than one booking at once. It sends the customer a night-before reminder and you a digest the next day, and hands the conversation over when something needs your judgement.
On data: NRIC and card numbers are redacted at ingestion, before anything is stored or sent to a model; AI processing runs in Singapore under a no-training agreement; retention is bounded. Plans are SGD $99, $199 and $399 a month, and access is invite-only for now — if that sounds like the gap in your setup, you can read more about how it works and request early access.