All articles

AI agent for WhatsApp customer messages: draft, then approve

An AI agent for WhatsApp customer messages reads every chat, tells you which ones need you now, and has a reply drafted in your voice before you open the app. The version I build and use every day does not send anything on its own. A person reads the draft, changes it if needed, and approves it, and only then does the message go out.

I built this because I was the one drowning. Agency work runs on WhatsApp groups, Telegram chats and Slack channels, and the clients are in the US while I am in the Philippines. Their working day is my night. The message that matters arrives at 3am, sits under a pile of others, and by the time I find it the client has been waiting for hours.

What an AI that replies to client chats actually does

The job is less “chatbot” and more “very fast assistant who reads everything”. Here is the loop my agents run:

  1. A client message lands in a group, a DM or a channel.
  2. The agent reads the thread, not just the last line, and writes a short summary.
  3. It flags what needs action: a question waiting on an answer, a deadline, a complaint, a payment.
  4. It drafts a reply that answers the actual question first.
  5. A person approves, edits or skips it. Nothing sends until that happens.

On the Slack side there is a second agent for operations. It raises urgent client issues, turns texts into tickets so they do not live only in someone’s head, and posts one recap each morning. That recap is the first thing I read. I can see which clients are waiting, which drafts are ready, and what can wait until later.

You can see the setup in the AI chat agents case study. These agents have run live every day across client channels since October 2026.

Why approval-first beats auto-reply for client relationships

Auto-reply sounds like the bigger win. It is not, at least not for clients who pay you.

A client chat is not a support queue. The same message can need three different answers depending on who sent it and what happened last week. “Can we push the launch?” from a client who is happy is a scheduling question. From a client who is already frustrated, it is a test. An AI that replies instantly and gets the tone wrong does damage that a slow reply never would.

Approval-first keeps the speed and removes that risk. The slow part of replying is reading the thread, remembering the context and finding the words. The agent does all of that. What is left for the person is a quick read and a yes.

There is also a plain trust reason. Clients notice when replies start to sound like a machine. The agent writes each draft against voice rules for that relationship: how I greet them, how formal I am, what I never say. A person still reads it before it goes, so the client always hears from a human who chose those words.

I made the send step hard on purpose. In my WhatsApp setup, every outgoing message is an immutable, single-use draft. If the text, the recipient or an attachment changes, it becomes a new draft that needs a new approval. A failed send is never retried blindly. The agent checks the chat first, so a client never gets the same message twice.

What to automate fully, and what never to

Some messages are safe to automate end to end. Most client messages are not. This is where I draw the line:

Fine to automate fully

  • Confirming receipt outside working hours, with an honest time for a real reply.
  • Fixed facts: opening hours, an address, a booking link, a file someone asked for.
  • Internal work: summaries, tickets, the morning recap, reminders to your own team.

Draft and approve

  • Any answer to a real question from a client.
  • Status updates, especially when something is late.
  • Anything that commits you to a date, a price or a deliverable.

Never automate

  • Money: invoices, refunds, discounts, payment chasing.
  • Complaints and anything emotional.
  • Anything you would be embarrassed to see screenshotted.

A fair note: if you run a store or a clinic that gets hundreds of the same questions from the public, a fully automated chatbot can be the better fit for the front line. For that case, a bot on your own business number through Meta’s official WhatsApp Business Platform is the right route. Approval-first is for relationships, where every client knows your name.

How the agent is built: Claude, n8n and custom MCP tools

The pieces are not exotic. What matters is how tightly each one is limited.

  • Claude reads the threads, writes the summaries and drafts the replies.
  • Custom MCP servers give Claude its tools. An MCP server is a small program that hands an AI a fixed set of actions, like “read this chat” or “create a draft”. Claude cannot send a message just because it wants to. The send tool needs an approved draft.
  • A WhatsApp bridge connects the agent to WhatsApp so it can read chats and queue drafts.
  • n8n runs scheduled jobs, like the morning recap.
  • Small Python services on a DigitalOcean server keep the watchers running when my laptop is closed.

The same safety idea runs through my other tools. My GHL workflow kit builds HighLevel automations through MCP, and every write there needs an allowlist entry, a confirmation phrase and a backup first. An agent that can touch real clients or real accounts needs rules that it cannot talk its way around.

If you want this connected to your own channels, the AI agents and custom MCP servers page explains how I scope it. For the wider automation around it, like tickets, follow-ups and CRM updates, see AI automation.

What I learned running it every day

A few things I did not expect when I started:

  • Summaries matter more than drafts. Knowing which few chats out of many need me is most of the value. The draft is the second half.
  • The agent must read the whole thread. Early drafts answered the last message and missed a question asked an hour before. Reading the full thread fixed it.
  • Voice rules need real examples. “Be warm” produces generic text. Pasting in how I actually write to a client produces drafts I barely edit.
  • The approval step catches real mistakes. Not often, but enough. A wrong date, a promise I could not keep, a reply to the wrong group. Each one would have cost more than the seconds the approval took.

For agencies, this pairs well with white-label fulfillment. If you resell builds to your own clients, the marketing agencies page covers how I deliver those under your brand.

Getting started

Send the project brief with the channels you use, roughly how many client chats you handle, and what slows you down today. I reply with a written plan: what the agent would read, what it would draft, and what stays fully in your hands. Then we talk it through on WhatsApp.

Questions

Can an AI agent reply to my WhatsApp customers for me?

It can draft every reply in your voice. In the agents I build, a person approves each message before it is sent, because a wrong message to a client costs more than a slow one.

Which chat apps can an AI agent read?

The agents I run read WhatsApp groups, Telegram chats and Slack channels. Each one summarises the conversation, flags what needs action and has a reply ready.

Should I let AI auto-reply to clients?

Only for answers that are fixed and low risk, like opening hours or a booking link. Anything about money, a complaint, a deadline or a promise should be drafted by the AI and approved by a person.

What is an AI agent built with?

Mine run on Claude, with custom MCP servers that give it a limited set of tools, n8n for scheduled jobs, and small Python services on a server. The tools are portable, so the model can change later.

Does the agent remember how I talk to each client?

It drafts from the thread itself and from written voice rules for each relationship, so a reply to a long-time client reads differently from a reply to a new one.