Robots Answering WhatsApp? They've Been Here for 3 Years and Almost Nobody Noticed

It's been three years since any company could connect a bot to its WhatsApp using Meta's official infrastructure, but it still surprises a lote of people. The typical reaction is "but isn't that a recent thing, with generative AI?" No. The technical door has been open since 2022, with the Cloud API, and for quite a while before that with the On-Premises version that Meta finally retired in October 2025. What's changed over the last two years isn't the ability to automate, but the quality of what's behind the automation.

This article is about that: why something that's been available for so long still seems like science fiction to many small businesses, what the platform actually allows, where the real limits are (not the ones invented in forums) and how you can tell — or not — when you're talking to a bot on WhatsApp without knowing it.

A timeline almost nobody knows

WhatsApp opened its business API long before generative AI was on everyone's lips. The first version, known as the On-Premises API, required the company (or its provider) to maintain its own servers running Meta's software. It was cumbersome, required serious technical maintenance, and was mainly aimed at large integrators.

In 2022 the Cloud API arrived, hosted directly by Meta. There was no longer any need to maintain your own servers for the basics: it was enough to connect your system to Meta's servers through a standard API. According to Meta for Developers' own documentation, since its launch the Cloud API "has seen strong adoption and better performance" than the previous version. And on October 23, 2025, Meta permanently retired the last supported version of the On-Premises client, leaving the Cloud API as the only official access route.

In other words: for years an old, heavy infrastructure coexisted with a new, accessible one, and the migration happened almost silently. No marketing campaign, no big announcement. One day it simply stopped being necessary to have a dedicated server to get a bot answering on WhatsApp. And that lowered the barrier to entry considerably: any medium-sized company with a reasonably competent developer — or an external provider — could have one running within weeks, not months.

This historical detail isn't just a curious anecdote. It explains why public perception lags well behind technical reality: the infrastructure had been ready for a while by the time the social conversation about "AI arriving" hadn't even started.

Foto de Sanket Mishra en Pexels

What "handling" WhatsApp with a bot actually means

When we talk about WhatsApp automation, there are two things people tend to mix up: the open conversation window and the sending of template messages.

WhatsApp for Business's official policy makes it clear: within 24 hours of a customer writing in, the company can reply with whatever it wants, including automation, as long as there's a clear, direct way to escalate to a human if the customer asks. Outside that window, a conversation can only be initiated using a message template previously approved by Meta, with a fixed format.

This distinction explains why some bots seem much more limited than others. A bot that can only send templates outside the window sounds rigid, like a fancy answering machine. A bot working within the 24-hour window, with well-defined business rules and access to real customer data, can hold a fairly fluid conversation without the user noticing anything odd — until it reaches a point where it needs to hand off to a person, and if it's well designed, that handoff is immediate.

Why nobody has noticed

There's a simple explanation and a less comfortable one.

The simple one is that a well-built bot doesn't try to look like a bot. It responds fast, handles the usual stuff — hours, availability, order status, an appointment — and steps out of the conversation the moment human judgment is needed. If it works well, the customer never even wonders whether they spoke to a person or a system; they simply got what they wanted. Invisible automation doesn't make headlines.

The less comfortable explanation has to do with how this kind of technology has generally been perceived: rigid menus of "press 1 for this, press 2 for that," clumsily adapted to a free-text channel, with canned responses that understood nothing outside the script, have long been the visible face of messaging automation. That kind of bot is spotted a mile away, and it has poisoned general perception: "WhatsApp bots are annoying." The result is that many people assume automation on this channel is new, clumsy, or experimental, when in reality it's had three years of constant improvement — first with fixed rules, and now, increasingly, with language models that better understand the intent behind a casual message.

Hypothetical examples of what's already routine

To be clear about what kind of tasks we're talking about, here are a few hypothetical examples — not real cases from any client, just illustrations of what the technology allows today without any invention:

A dental clinic could have a bot that confirms or reschedules appointments via WhatsApp, sends the reminder 24 hours in advance, and, if a patient writes outside office hours asking about specific pain, automatically forwards an emergency message with the on-call phone number, without waiting for reception to open.

A car repair shop could automate the "your car is ready" notice and let the customer reply with simple questions about the cost or pickup time, while anything that sounds like a complaint or a complex technical issue goes straight to a person.

A small online store could handle a good chunk of "where's my order?" questions via WhatsApp using real data from its shipping system, without anyone on the team having to check their phone every five minutes.

None of these examples require sophisticated artificial intelligence. They require good integration with the company's systems (calendar, ERP, carrier) and clear rules about when the bot should stay quiet and hand off.

Foto de Amina Filkins en Pexels

The pricing shift that really is recent

Here's a change that genuinely is from the last few months and worth keeping in mind if you're considering getting into this now. As of July 1, 2025, according to Meta for Developers' official documentation, billing is done per delivered template message, replacing the previous model based on open conversations. This changes the numbers for any project: if your bot sends a lot of template messages to re-engage conversations outside the 24-hour window, the cost is calculated differently than it was until recently.

It's a technical detail, but it has a direct practical consequence: designing the flow so that most of the interaction happens within the free 24-hour window (once the customer has written first) remains the cheapest strategy and, moreover, the one that gives the best experience, because that's where the bot can be most flexible.

Any project being considered today should start by understanding this billing mechanic well before designing the conversation flow, not after. A bot that depends on constantly re-engaging the user with templates outside the free window can end up much more expensive than expected, and that cost is only caught if the numbers are worked through carefully from the start.

What a WhatsApp bot can't do, even if you want it to

It's worth being honest about the limits, because part of the noise around this topic comes from inflated expectations.

You can't freely start a conversation with someone who hasn't written to you first, except with a Meta-approved template, with a fixed format and subject to review. This immediately rules out any fantasy of "mass WhatsApp like email marketing": the channel is deliberately designed so that the initiative to make contact rests mostly with the user.

You also can't automate everything friction-free: the policy itself requires that the user have a clear way to request a human, and that means designing that escalation path — it's not optional or a "nice to have." A bot that doesn't allow going off-script violates the policy, on top of being a bad experience.

And there's a practical limit, not a regulatory one: the more complex what you're asking the bot to resolve — a negotiation, a delicate complaint, a legal case — the more the design should account for handing it off early. Automation performs best on repetitive, well-defined tasks, not on everything.

How to set one up without overcomplicating things

The technical part has stopped being the main obstacle. Today the bottleneck usually lies in designing the conversation and integrating it with the company's real data, not in the infrastructure.

1. Choose two or three specific use cases

   Not "automate customer service" in general, but specific tasks: confirming appointments, informing about order status, answering very frequently repeated questions.

2. Define the escalation point

   Before writing a single line of flow, decide which phrases or situations should immediately hand off to a person, and to whom.

3. Connect to real data

   A bot without access to the calendar, the ERP, or the shipping system can only give generic answers. Real usefulness comes from the integration, not from the bot itself.

4. Test it with real customers before announcing it

   Let it run in parallel with human supervision for a few weeks before trusting it without a safety net.

5. Review and adjust based on real conversations

   The first flows are almost never fully right; you need to review transcripts and fix the points where the bot gets stuck or the customer gets frustrated.

Common risks when designing the flow

Beyond the concrete steps, it's worth keeping in mind some general risks that tend to come up when planning this kind of project, regardless of industry.

One is treating the bot as a closed project that gets launched and forgotten. Customer conversations change over time, new questions come up, and a flow that isn't reviewed periodically tends to become outdated.

Another is not measuring anything. Without data on what proportion of conversations the bot resolves on its own and how many end up escalated to a person, it's hard to know whether the system is working well or just sitting there taking up space.

A third risk, already mentioned above, is underestimating the 24-hour window as a design constraint: treating the project as if WhatsApp were a free messaging channel, without accounting for when you can write without a template and when you can't, usually translates into cost surprises and worse-designed experiences.

These risks aren't exclusive to WhatsApp or to this particular technology. They're the same ones that show up in any automation project designed without measuring or reviewing, and they're worth keeping in mind precisely because they're easy to overlook when the technical part seems solved.

Why this matters beyond the anecdote

The fact that this technology has been available for three years isn't just a historical curiosity. It has a simple practical implication: the barrier to automating WhatsApp today is not mainly technical or about infrastructure cost, but about the time spent designing the conversation flow well and integrating it with your own systems. That changes the question any small business should be asking: not "is this viable yet?" but "do we have a clear idea of what we want to automate and how it should hand off to a person when needed?"

And conversely: if a company already has a poorly designed bot, with rigid menus and no clear way out to a human, it's probably generating more frustration than savings. The platform's maturity is no longer an excuse for a bad experience. Meta has been providing the pieces for three years; what's usually missing, in practice, is sitting down to design the flow properly, rather than waiting for a newer tool to come along.

A final thought, without selling smoke

You don't need cutting-edge artificial intelligence to have a useful WhatsApp bot. You need clarity about which questions come up repeatedly in your business, access to the data needed to actually answer them, and an honest way out to a person when the bot can't go further. That could be built in 2022 and it can be built today, with better tools but the same underlying approach.

The surprise shouldn't be that robots are answering WhatsApp. If anything, the surprise is that after three years of technical availability, it's still so common to come across poor implementations or businesses that haven't even considered it, when the technology to do it right has been waiting all this time.

Cover image: Matheus Bertelli on Pexels