Nexly AgentsBlog

Nexly Agents › Blog › Guide

Guide

WhatsApp AI agent for customer service: how to set it up in 2026

Published 2026-10-14·6 min read·Author: Jakub Milewski, Nexly Lab

Photo: Unsplash

A customer messages you on WhatsApp at 9:15 pm because that is where they already are. They want to know whether Tuesday's order ships before the weekend. If the answer arrives the next morning, the conversation still happens, just later and with a less patient customer. That is the case for a WhatsApp AI agent: it answers straight away from your own knowledge and hands over to a person when it does not know. Before you set one up, you need to understand the rules Meta applies to every business on the platform, whether a human or a language model is typing the replies.

The app is not the platform. The free WhatsApp Business app is fine when one or two people answer by hand. An AI agent needs the WhatsApp Business Platform, which means API access; Meta offers it as the Cloud API. Your business number is connected to a WhatsApp Business account and incoming messages are delivered to the software that replies. If you already use the app, the same number can be connected to the API while you keep using the app.

Choose a partner, often called a BSP. Few companies connect to the API on their own. Most go through a Meta partner. Meta's documentation now calls them Solution Partners and Tech Providers; the industry still says BSP, Business Solution Provider. The difference matters for billing. A Solution Partner can invoice you directly for platform usage. With a Tech Provider, you add your own payment method, Meta bills you for messages and the provider bills you for everything else. Settle this before you sign, because it decides who has access to your account and what your invoice looks like.

The 24-hour customer service window is the rule everything else depends on. When a customer messages or calls you, a 24-hour timer starts, and every new message from them resets it. While the window is open you can reply freely: text, images, interactive lists. Once it closes, you can only send a pre-approved template. For an AI agent this means it answers questions while the window is open, but it does not start a new conversation three days later on its own.

Prepare your templates before launch. A template is a message with fixed wording and fields you fill in, such as a ticket number. Each template must be categorised as marketing, utility (status updates, confirmations) or authentication (one-time codes). Meta reviews templates before you can use them, and the category affects the price. You will need at least a few utility templates: ticket status, a request for missing details, a note that a person has taken over. Without them, the colleague who picks up the case the next day cannot reply at all.

You pay per message, not per conversation. Since 1 July 2025 Meta charges for each delivered template message, at a rate that depends on the category and the recipient's country code. Replies inside the customer service window have been free since November 2024, and utility templates sent inside an open window since July 2025. That is changing. In its developer documentation Meta announced that from 1 October 2026 it charges per message for service replies and for utility templates inside the window, at the same rate as utility messages in that market and without volume tiers. The business-facing pricing page still describes service replies as free, so ask your partner for current rates for your markets before you budget. Meta changes prices at most once a quarter: on 1 January, April, July or October.

Per-message pricing leads to one practical setting: the agent should answer in one complete message rather than three short bubbles. There is also an exception worth knowing. When a customer arrives from a Click to WhatsApp ad or a call-to-action button on your Facebook Page and you reply within 24 hours, all messages are free for 72 hours.

Get opt-in before you message first. Since the November 2024 policy update, opt-in can be general rather than WhatsApp-specific, but it must clearly say that the person agrees to receive messages and from which business, and it must comply with applicable law. When the customer writes first, things are simpler, but the agent should not turn that conversation into a mailing list.

Human handoff is a requirement, not an extra. The WhatsApp Business Messaging Policy says that when you use automation, customers must have prompt, clear and direct escalation paths: an in-chat transfer to a human agent, a phone number, e-mail, web support or a form. Three rules work well in practice:

  1. The customer types "agent", "human" or is clearly frustrated, and the AI hands over without arguing.
  2. The AI cannot find the answer in approved knowledge and says so instead of guessing.
  3. The case involves money, a complaint or personal data, so a person decides.

The handoff has to carry context. A support rep who asks "how can I help?" after ten messages with the AI wastes everything the customer has already said. And mind the window: if a person returns to the case the next day, they need a template.

Tell customers they are talking to AI. From 2 August 2026, Article 50 of the EU AI Act requires it, at the latest at the first interaction. One sentence in the agent's first reply is enough.

Two changes from 2026 to know about. Since 15 January, Meta allows so-called AI Providers to offer general-purpose AI assistants on the platform only where it is legally required to, and Meta states that this policy does not change how it charges other businesses. An agent that answers questions about your own products and orders is a different case from a general chatbot, but ask your vendor to confirm it before launch. And since July, Meta has its own agent platform, Meta Business Agent, billed per token from 1 August. It is one more option to compare.

A setup order that saves trouble:

  1. Pick the number and the partner; agree who pays Meta and who can access the account.
  2. Collect the knowledge: common questions, terms, delivery and returns rules. The agent should answer only from these.
  3. Write your utility templates and submit them for review.
  4. Define handoff rules and who on your side picks up conversations, including evenings.
  5. Test the agent on questions from real conversations before you give customers the number.
  6. After launch, read conversations weekly and fill the gaps where the agent had no answer.

Nexly Agents is one of the options. It runs in production on WhatsApp, as well as on your website, e-mail, Messenger, Telegram, Signal, Microsoft Teams and Slack, with one knowledge base for all channels (omnichannel AI agent for WhatsApp, Teams and Slack). It answers only from knowledge your company has approved and shows the source; when it does not know, it says so and hands the conversation to a person with full context. Conversations from messaging apps land in one inbox with a customer wait-time counter and one-click takeover. How we protect data is described on our security page.

Want to see this on your own website rather than on slides? Enter your website address and e-mail in the demo form. In 20 minutes you get an agent trained on your offer, ready to be quizzed.

Sources

See the AI agent on your own website

Give us your website address and e-mail, and in 20 minutes you get an agent trained on your offer, ready to be quizzed.

Get a free demo

Read also

GDPR-compliant AI chatbot: a checklist for EU companies Omnichannel AI customer service: one agent for website, email, WhatsApp, Teams and Slack Don't start AI with a revolution