Start with a week of real conversations
Before you choose any tool, export or scroll back through one normal week of customer chats and tally what repeats. Founders are usually surprised twice: by how much of the inbox is three or four questions asked in different words, and by how many messages were answered hours late because nobody was on the phone.
That tally is your brief. It tells you which flows are worth automating, what your assistant needs to be able to look up, and where delays are costing you money. Building from a feature list instead of from real messages is how teams end up with a chatbot nobody uses.
Match the tooling to your stage, not to your ambition
The expensive mistake is jumping to a custom build before your process is stable, or clinging to one phone long after the inbox stopped coping. Three rough stages:
| Where you are | What usually fits | Signal that you have outgrown it |
|---|---|---|
| One founder answering everything, still validating | The free WhatsApp Business app: greeting message, away message, quick replies, labels | Two people need the same inbox, or messages sit unanswered overnight |
| First hires, the same questions every day | A shared inbox on the official WhatsApp Business Platform with two or three scripted flows and templates for reminders | Replies need live data: a free slot, a stock level, an order status |
| Operations run through the channel | Automation wired into your booking system, store, CRM or internal API, with an AI agent answering from your own content | You are editing flows by hand every week, or the assistant cannot act, only answer |
The split between the free app and the official platform is the decision most teams get stuck on. We walk through it in detail in our guide to WhatsApp Business: app or API, so this article assumes you will pick the one that matches the row above.
The flows that pay for themselves first
Score each candidate flow against four questions before you build it:
- Volume. How many times a week does someone type roughly the same thing?
- Decision complexity. Can the answer be produced from data you already hold (price list, opening hours, calendar, order status), or does it need judgement?
- Cost of being slow. A late reply to a booking request loses the booking. A late reply to an invoice query rarely does.
- Reversibility. If the automation gets it wrong, can a person fix it easily? Confirming a slot is safe. Issuing a refund is not, unless a human approves it.
Three flows usually clear that bar early:
- Lead qualification. The first reply asks the two or three questions your team always asks anyway (what, when, where, how many), then writes a structured lead into your CRM instead of leaving it buried in chat.
- Booking and reminders. The customer picks from live availability, gets a confirmation, and gets a reminder before the appointment. That is the pattern behind our own booking product, Artover Appointment, which sends reminders over WhatsApp.
- Order and delivery updates. Paid, packed, on the way, delivered. Four short messages that remove most of the "where is my order" traffic.
Buy a tool, build on the API, or have it built
For a startup, engineering time is scarcer than money, and whatever you build you also maintain. Decide on that basis:
| Option | What it really costs you | Fits when |
|---|---|---|
| Off-the-shelf chat platform | Recurring fees that scale with seats or conversations, and limited control over your own conversation data | You need a shared inbox and simple flows this month, with no engineering time to spare |
| Building on the API yourself | Weeks of founder or engineer time, then ongoing maintenance of templates, retries, opt-outs and hand-over logic | Messaging is part of the product itself, not a support channel |
| Having it built and maintained for you | A scoped project, plus a maintenance relationship afterwards | WhatsApp has to talk to your booking system, store, ERP or CRM and follow your own procedures |
A sensible sequence: free app while you validate, an off-the-shelf tool once two people share the inbox, a custom integration once the channel is doing real operational work. There is no prize for skipping to the last step, and plenty of startups have paid for a build they outgrew in a different direction six months later.
Design the hand-over before you design the bot
The quality of an automated channel is judged at the point it fails. Decide in advance what the assistant does when it is unsure, when the customer is angry, when money is involved and when it is 2am and nobody is on shift.
- Define the exact triggers for hand-over: low confidence, a refund or complaint, a repeated question, or the customer simply asking for a person.
- Pass the full conversation and a short summary to whoever takes over, so the customer never repeats themselves.
- Say honestly that a human will reply, and when. Silence after a bot says "let me check" is worse than no bot at all.
- Keep anything that commits the business (a discount, a cancellation, a refund) behind human approval until the logs give you reason to relax that.
Our own assistants are built this way: when the AI cannot handle a request it hands over to a person with the context, and in supervised mode it proposes actions rather than taking them.
Notifications people actually want
Outbound messaging is where goodwill gets spent. Two habits protect it. First, only message people who contacted you or opted in. Second, treat notifications as a service, not a campaign: order updates, booking reminders and document confirmations are welcome, repeated promotional blasts to the same list are not. WhatsApp also sets its own rules on business messaging, and Meta updates them, so check the current requirements on its official documentation before you plan a launch date rather than relying on what a blog said last year.
One rule is worth building around from day one: reminders to customers who have not written to you in the last 24 hours go out as WhatsApp-approved templates, not free text. That shapes how you phrase reminders and how far ahead you schedule them.
Treat your WhatsApp number as infrastructure, not a marketing channel. If access is interrupted you lose support, bookings and sales at the same moment, so keep a fallback your customers already know: a website contact form, email, or a second channel.
Expanding into a second market
The flows rarely transfer unchanged. Opening hours, delivery promises, payment methods and the way people actually write all shift across borders, and a template that reads naturally in one market can sound stiff in the next. Plan for separate copy per market rather than a translation pass, and keep a human reviewing the first weeks of conversations in each new country.
Language is part of this. Customers often mix two or three languages in a single message. Our AI agents work in English, French and Kreol Morisien, including mixed messages, and reply in the language the customer used. Whatever you build, test it with the messy, abbreviated way your customers really type, not with clean sentences written by your own team. We have a version of this written for teams running on WhatsApp automation in South Africa, and the local differences there are a good illustration of why copy does not travel for free.
The data questions you will be asked
WhatsApp threads contain names, phone numbers, addresses, order history and sometimes identity documents. Sooner or later a larger customer, a partner or an investor will ask where that sits. Have answers ready for five questions: where conversation data is stored, who on your team can read it, how long you keep it, whether any of it reaches a third-party model, and how an opt-out is recorded and honoured.
We design these systems with the Mauritius Data Protection Act (2017) and GDPR in mind, and offer private deployment so client data stays in an environment the client controls and is not used to train public models. Deciding this early costs a conversation. Retrofitting it after you have a year of chat history costs a project.
Measure two things, not twelve
Dashboards full of message counts prove nothing. Track how many conversations closed without a human touching them, and how many of those customers came back unhappy or asked the same thing again. If the first number rises while the second holds steady, automate the next flow. If both rise, your assistant is deflecting rather than resolving, and the fix is usually better access to data, not better wording.
When you are ready to connect the channel to your booking system, store or CRM, our WhatsApp automation page sets out the integrations we build most often.
Frequently asked questions
Where should an early-stage startup begin with WhatsApp automation?
With one flow, chosen from a week of real conversations. Lead qualification, booking confirmations and reminders, or order status updates are the usual first candidates because they repeat often and can be answered from data you already hold.
Should we build on the WhatsApp API ourselves or use an existing tool?
Use a tool while messaging is a support channel and your engineering time belongs in the product. Build, or have it built, once WhatsApp has to read and write to your booking system, store or CRM and follow procedures that are specific to your business.
Can a WhatsApp assistant reply in French or Kreol Morisien?
Yes. Our AI agents work in English, French and Kreol Morisien, including messages that mix languages, and reply in whichever language the customer used.
Can we keep the WhatsApp number our customers already have?
Usually yes. An existing WhatsApp Business number can often be connected to the official platform rather than replaced, so you do not lose the number people have saved.
How do we send reminders without annoying customers?
Only message people who contacted you or opted in, keep notifications useful (confirmations, reminders, delivery updates) rather than promotional, and make opting out easy. Reminders to customers who have not written in the last 24 hours go out as WhatsApp-approved templates.
What drives the cost of a WhatsApp automation project?
Mainly how many systems it has to talk to, how many flows go live at launch, and whether the assistant needs to answer from your own documents and data. We give a fixed written quote after a free discovery call.