How WhatsApp customer engagement actually gets automated

Most descriptions of WhatsApp automation are a list of things a tool will do for you: reply instantly, follow up, broadcast, recover carts. The parts that decide whether any of it works are the things it declines to do.

Meta accepts a free-form message sent more than twenty-four hours after the customer's last one, returns a success, and then drops it. Nobody is told. The business believes it answered; the customer believes it was ignored. So the first piece of real machinery is a window check that runs before the message reaches Meta at all, refuses the send, and names the approved template that would have worked instead. A tool that reports 'sent' there is not automating your engagement, it is automating a lie.

The same shape recurs. A marketing template sent to somebody with no matching opt-in risks the number, so the send has to be refused at launch rather than apologised for later. An authentication template containing a link is a phishing pattern that Meta rejects, so it has to be refused at composition time, before a submission is spent on it. Automation that only ever says yes hands you a faster way to make expensive mistakes.

A customer's message arrives on a webhook. Before anything answers it, several things have to be recorded, because each one is a question somebody will ask later.

The conversation's recency moves, which is what an unassigned queue sorts by. The response clock starts — from this message, not from the third one they send while waiting, or the wait resets every time they get impatient. If the workspace has asked for automatic routing, an owner is picked: round-robin, or whoever holds the fewest open conversations, skipping anyone marked away or outside working hours, and the reason for the choice is written down so 'why do I always get the difficult ones' has an answer.

Then something replies — a keyword flow, an AI responder working from the business's own knowledge, or a human. When it is automation, the customer has to be able to tell, and asking for a person has to work immediately rather than after another round of deflection.

When the reply goes out, the response time is stored on two clocks. Wall time is what the customer experienced. Business time counts only open hours. A message arriving at 23:00 and answered at 09:05 was a ten-hour wait and a five-minute wait, and a tool that reports only one of those is either slandering the team or hiding a real failure.

Chasing someone who went quiet is the most valuable automation in this category and the easiest to get wrong. The mechanism is a short ladder — a nudge after an hour, another the next day, a last one three days later — and the part that matters is the exit.

The moment the customer replies, or buys, or books, the rest of the ladder has to be cancelled. Not paused, not skipped at send time: cancelled, in the same breath as the event. Everyone has received the message that says 'still thinking about it?' the day after they bought the thing, and it does more damage than the sale was worth.

Consent has to be re-checked at every rung, not once at the start. Somebody who opts out on Tuesday must not get Wednesday's nudge, which means the ladder cannot be a loop over a list captured at enrolment. It has to be a set of scheduled steps that each consult consent when their time comes — and a step skipped for that reason should be recorded as skipped, with the reason, because 'nine due, four sent, five stopped by an opt-out' is an answer and a silently shorter send is not.

In India most of this ends in cash on delivery, which changes the shape of the automation. There is no payment to wait for, so the customer's confirmation *is* the order: a button reply moves it to confirmed, records that they agreed, and tells the shop. No payment link, because creating one and then ignoring it is how somebody pays twice.

After that the useful automation is not promotional at all. Placed, shipped, out for delivery, delivered — each arriving from the shop's own system and going out as a utility template with the tracking link. Utility, deliberately: a discount smuggled into a shipping notice gets the whole account recategorised and rate-limited.

Two details separate a working version from a demo. Paid and shipped have to be different facts, or a cash order looks unconfirmed for a week. And a redelivered webhook — which every commerce platform sends eventually — must change nothing and message nobody, or your customer is told twice that their parcel shipped.

Since July 2025 Meta charges per message, by category. A marketing template costs materially more than a utility one, and a reply inside the customer's own twenty-four-hour window is free today — which changes on 1 October 2026, when in-window service messages begin to be charged.

That has a practical consequence for how automation should behave: the cheapest correct action is usually to answer inside the window, and the expensive one is to start a conversation with a marketing template. A tool that treats every outbound message as the same thing will quietly spend a multiple of what it needed to.

It also means a projected cost belongs on the screen before a broadcast goes out, not on the invoice afterwards — reachable recipients multiplied by that category's rate, from a rate table with dates on it. Dates matter because the rates expire: any tool quoting you a price without saying when it stops being true is quoting you a number it has not checked. Wavy charges 0.10 credits for an in-window reply, 0.35 for a utility or authentication template and 1.25 for marketing, where one credit is one rupee, GST is added when you top up, and those figures hold until 30 September 2026.

Four questions separate tools that have built this from tools that have described it.

What happens if I try to send a free-form reply twenty-six hours after the customer's last message? The right answer is a refusal naming the template alternative, not a success.

What does a rejected template tell me? 'Rejected' alone means somebody has to open Meta's Business Manager to find out why. The reason, the quality rating, and whether Meta re-categorised your utility template as marketing — that last one changes what you are charged — should all be on the row.

What does the follow-up sequence do when the customer buys? If the answer is anything other than 'the rest of it is cancelled', it will message people who have already converted.

And what will this broadcast cost? A tool that cannot answer before sending is asking you to find out from your bill.