AI customer service on WhatsApp: what to expect.

An AI agent on WhatsApp is a virtual employee that reads the messages arriving at your business's WhatsApp, understands what the customer wants and replies —or resolves the request— autonomously, in plain conversational language and at any hour. It isn't the "press 1, press 2" menu of years ago, and it isn't magic that answers anything without ever getting it wrong. It's software that takes on the repetitive part of your customer support so your team can handle what genuinely needs a person.

What an AI agent on WhatsApp is (and isn't)

Behind it sits an LLM —the same technology that powers ChatGPT or Claude— connected to your business WhatsApp and to the tools where the information lives: your CRM, your database, your order or booking system. The AI provides the conversation and the judgement; the integrations provide the hands. Without those connections you have a parrot that talks well but can't check an order or book anything. With them, you have something that actually closes tasks.

The difference from the classic chatbot is autonomy. The decision-tree bot followed a script: if the customer stepped off the menu, it broke. An AI agent understands what people ask even when they write it badly, with typos or half-finished, decides what to do, checks wherever it needs to, and answers. And when it doesn't know, or the case turns delicate, it says so and hands the conversation to a person —with a summary of what was said— instead of inventing an answer.

What it resolves in customer support today

Far more than you'd think. Where it fits perfectly right now:

• Repetitive questions. Opening hours, address, order status, "do you have this?", "how do I do that?". It's the bulk of the volume that eats your team's day, and it's answered instantly without making anyone wait.
• Capturing and qualifying whoever writes in. Every message answered at minute zero: the agent asks the right questions, discards what doesn't fit and books the call or the quote with the ones that do. In support and in sales, response speed is where most opportunities are won or lost, and a machine doesn't sleep or take Sundays off.
• Simple end-to-end tasks. Checking an order, changing a booking, collecting the details of an incident and opening the ticket. Concrete work, not just "informing".

At Tec2020 we run these agents in our own products before we sell them: the assistant handling the chat and WhatsApp of ListingOK is exactly this. We've been building software for years —the apps we've developed add up to more than 130 million downloads— and AI is the new layer on top of that same craft, not a leap into the void.

Why WhatsApp is the channel that matters

In Spain, when a customer wants something from a business, they open WhatsApp before email or the phone. It's where they already are and where they expect a fast reply. That's why putting the agent there —and not hidden in a website chat nobody uses— multiplies what it resolves.

To do it right —reliably, and without getting your account shut down— you use Meta's official WhatsApp Business API, not the phone app with a program running behind it. That gives you a verified number, approved templates to start conversations and the reliability to connect the agent to your systems. It's a technical detail, but it's the difference between a fragile setup and something you can lean a business on. If you want the background, the WhatsApp Business Platform documentation explains it.

Where I wouldn't put it yet

This is the part almost nobody tells you when they want to sell you the project.

Don't let an agent make decisions that cost money or reputation without supervision: approving a large refund, giving legal or medical advice, promising something the business then can't deliver. And in a delicate conversation —a serious complaint, an angry customer, a sensitive case— the machine should recognise it early and hand over to a person, not insist on solving it alone. AI works, but it can fail, and we should only let it fail where the failure is cheap and reversible.

The practical rule: start with high-volume, low-risk messages. You'll quickly see the system gets things right well beyond what you expected, and that's when you widen its responsibilities. The other way round —handing it the critical stuff on day one— is like giving someone you've just met the keys to the till.

How to know if it's worth it

Before spending a euro, answer these three questions:

1. Do you have a type of request that repeats on WhatsApp? If you answer the same thing many times a month, there's a case. If it's four one-off, unrelated enquiries, there isn't.
2. Is the information accessible? If the answer lives only in someone's head, it has to be written down first. If it lives in a system you can connect to, all the better.
3. Who supervises? The agent needs a human owner who reviews the doubtful cases and tunes it early on. It isn't "set it and forget it".

If you answered yes to the first two and you're clear on the who of the third, it will probably pay off.

And there are three situations where it pays off for sure:

• You want to answer 24/7. If people write at odd hours, at the weekend or in the middle of the night and those messages currently sit unanswered until the next day, the agent answers them on the spot. It doesn't rest or take days off.
• You don't want anyone left waiting. An agent is never busy: it handles many conversations at once, so nobody is left on "read" or gets a "we're receiving a lot of messages, we'll get back to you soon".
• You serve more than one language. If people write in several languages, the agent replies in the customer's own language without you needing a person for each one.

How to roll it out without making a mess of it

The most expensive mistake is framing it as a six-month mega-project. You don't need to. The sensible way: you start with one concrete, measurable type of request —the most repeated one—. You set up the WhatsApp agent and connect it to the systems it needs, and you give it your business's knowledge. You test it small, with real conversations and a person watching. You measure: how many cases it solves on its own, how many it escalates, where it gets things wrong. And you correct. In two-week sprints you see results — or problems — fast, which is exactly what you want, instead of discovering six months in that you were on the wrong path.

And one detail that matters more than it seems: that the project doesn't tie you down. No mandatory maintenance contracts, with the code and the data being yours. If it works, you'll continue because you want to, not because you're trapped.

Where to start

Don't start by buying "an AI for WhatsApp". Start by choosing the type of request: the most repeated one, the highest-volume one, the one your team answers on autopilot. That's your first candidate —order taking, for example. The rest —which model, which integrations, how to measure it— is fairly predictable engineering once the task is well chosen. If you want the full picture, we cover it in our guide to AI virtual employees.

If you already have a case in mind and want to know whether it makes sense to automate it, the fastest thing is to talk it through on a call: you tell us what comes in on WhatsApp and we tell you frankly whether we'd build it, roughly what it would cost, or whether you're better off not doing it. No commitment, no hype. You can also take a look at the AI virtual employees we build.

← Back to Blog