AI agents on WhatsApp: what actually works in real operations

Actually solve what it can

An AI agent on WhatsApp is worth it when it resolves things, not when it just replies. Replying is telling someone your opening hours. Resolving is looking up the customer’s order in your system and telling them it ships tomorrow, with the tracking code.

For that, the agent has to be connected to your company’s systems. A few examples of tasks that usually work well:

• Giving the status of an order, delivery or ticket by checking the system.
• Sending a copy of an invoice or a payment slip.
• Booking, rescheduling and confirming appointments.
• Collecting details and documents to open an account or a ticket.
• Answering questions about products, policies and timelines based on official company information.
• Qualifying a sales lead before passing it to a salesperson.

These are just examples. If a task repeats and follows clear rules, it can be automated. The test is simple: does the agent have the information and the permission to resolve it? If so, it resolves it. If not, it hands it off.

Hand off to the right person, at the right time

The handoff to a person is where most bots fail. Either it takes too long, leaving the customer stuck in a conversation with no way out, or it lands in the wrong queue and the customer has to explain everything again.

A good agent hands the conversation off in situations agreed in advance:

• The customer asks to talk to a person. No pushback, no asking three times if they’re sure.
• The issue needs a human decision, such as a negotiation, a policy exception or a serious complaint.
• The agent isn’t sure of the answer or couldn’t find the information.
• The customer is getting frustrated or the conversation is going in circles.
• The topic is sensitive, such as health, a legal matter or detailed financial information.

And it hands off to the right person. A billing question goes to finance, a technical problem goes to support, a sales proposal goes to that customer’s account manager. That means the agent has to classify the issue and the routing rules have to be set up, rather than everything dropping into one generic queue.

An agent that knows when to step out of the conversation is worth more than one that tries to solve everything.

Keep the context from start to finish

Nothing annoys a customer more than repeating the same story. When the agent hands the conversation to a person, it should include a summary: who the customer is, what they asked for, what was already looked up, what was already tried and why it’s being escalated. Whoever picks it up reads it in seconds and carries on from there.

Context also has to survive over time. A customer who messaged yesterday and comes back today shouldn’t be treated as a stranger. The agent should know what was discussed, whether there’s an open ticket and whether someone promised to get back to them.

It works the other way too. If a person resolved the case, the agent should know that next time, instead of offering the same fix that didn’t work.

WhatsApp rules and privacy laws you need to know

For a business to run agents on WhatsApp reliably, the way to do it is the official WhatsApp Business Platform, usually through a provider authorized by Meta. Setups that automate a regular number with an unofficial app risk getting banned and come with no guarantees.

A few practical points about the official platform:

• Messages the business starts outside the customer service window use templates pre-approved by Meta.
• There are rules about what can be sent and what consent the customer has to give.
• Pricing varies by conversation type, so it’s worth understanding the billing model before you scale.

On the privacy side, the agent handles personal data all the time: names, phone numbers, addresses, sometimes documents. Under laws like GDPR in Europe and LGPD in Brazil, the business needs a legal basis for each use, should collect only what it needs, protect what it stores and be able to handle requests from the people the data belongs to, such as access and correction. It’s also good practice to make clear to customers that they’re talking to an automated assistant.

Common mistakes when you’re getting started

• Trying to automate everything at once. Start with two or three frequent, well-defined topics. Expand once you’ve seen how the agent behaves.
• Leaving the agent without official information. If it answers from outdated or scattered texts, it will be wrong with confidence. Keep a knowledge base that someone on the team reviews.
• Hiding the way out to a person. A customer who can’t find how to reach someone gets angrier than they would with no bot at all.
• Not reading the conversations. Review a sample every week at the start. That’s where the questions nobody anticipated show up.
• Measuring volume only. The number of conversations handled says little. Also look at how many were actually resolved and how many came back with the same problem.

How to tell if it’s working

A few simple questions help you evaluate the agent after the first weeks:

• What share of conversations did the agent resolve without anyone else?
• How long does the customer wait for an answer, with and without a handoff?
• Does the team receiving handoffs say the summary helps or gets in the way?
• Are customers coming back about the same issue?
• Which questions does the agent hand off most, and could it resolve any of them?

The answers show you where to adjust. A good WhatsApp agent is almost never ready on day one. It gets good because someone follows the conversations and improves the rules, week after week.