WACrew

Blog

How to use AI to reply to WhatsApp messages (ChatGPT, Claude and Gemini)

Three ways to use ChatGPT, Claude or Gemini to answer WhatsApp messages, from copy and paste to a connected agent, with prompts, costs and what AI should never answer alone.

Updated 4 min read
On this page
  1. Level 1: copy, paste and review
  2. Level 2: AI inside a support platform
  3. Level 3: an agent that reads the history and writes like you
  4. Rules that apply at every level
  5. Which model to use
  6. Where WACrew fits

Using AI to reply to WhatsApp messages is no longer something only big companies do. Anyone with ChatGPT, Claude or Gemini can speed up their replies today without writing code. What changes between setups is how much of the conversation the AI can see, and who presses send.

This guide covers three levels, from the simplest to the most automated, and the rules that apply to all of them.

Level 1: copy, paste and review

The simplest way is to use the AI chat as a drafting pad.

  1. Copy the last messages of the WhatsApp conversation (WhatsApp Web or the desktop app makes this easier).
  2. Paste them into ChatGPT, Claude or Gemini with a clear instruction.
  3. Review the draft, adjust what you need and send it yourself.

A prompt that works well:

You help me answer clients on WhatsApp.
Write the way I write: short sentences, no emoji, first name only.
Don't promise deadlines, prices or discounts. If the client asks for one, say I'll confirm.
Don't end with a stock sign-off.
Below are the last messages. Write only my next reply.

[paste the conversation here]

Good for: a few conversations a day, when you want help with long or delicate replies. Cost: the free plans are enough. Risk: none for your number.

The weak spot is the manual copying every time, and the AI only sees what you pasted.

Level 2: AI inside a support platform

The second level is a platform that connects WhatsApp to an AI model. Most of them run on the official WhatsApp Business Platform, and many already include a configurable AI agent: Wati, respond.io, SleekFlow, Twilio-based setups and others. We compare several in WhatsApp AI agent platforms and Wati alternatives.

You feed the AI a knowledge base (FAQs, catalog, policies) and it answers incoming messages, handing over to a human when it doesn't know.

Good for: high-volume support with similar questions, like e-commerce, clinics and schools. Cost: the platform subscription, plus Meta's per-message pricing for template messages. Risk: low, since it uses the official API.

The limit shows up in consultative work. A knowledge base can answer "what is the delivery time?", but not "what about the change we agreed in yesterday's call?". For that, the AI has to read the history of that specific conversation.

Level 3: an agent that reads the history and writes like you

The third level is an agent connected to your number that reads the whole conversation before writing. It knows what the client asked yesterday, what you promised and what is still open.

There are two ways to build it:

  • With automation tools like n8n or Make, wiring a WhatsApp connector (Evolution API, WAHA and similar) to an AI model. You design the flow, and the flow decides when to reply.
  • With a coding agent over MCP. The Model Context Protocol lets Claude Code, Codex and other clients read and answer WhatsApp through tools. We walk through it in build a WhatsApp AI agent with MCP.

Good for: consultative support, client group chats, technical support. Cost: hosting plus the model API. Risk: these connectors usually rely on WhatsApp's unofficial connection, so your number is exposed if you send in bulk or to people who never messaged you.

Rules that apply at every level

1. Give the AI real examples of your writing. "Write in a friendly way" produces the assistant tone everyone recognizes. Pasting ten of your own messages as examples gets much closer to you. Our guide on how to make AI write like you shows the full method.

2. Block the tells. Praising the question, a stock sign-off, em dashes in the middle of sentences, bulleted lists in a chat bubble. These habits show up in almost every model and clients notice. We list the common ones in the tells that make AI text read like a bot.

3. Decide what the AI never answers alone. Prices, deadlines, discounts, serious complaints, meeting requests, system access. In those cases, the AI at most acknowledges the message and hands it to you.

4. Don't let the AI deny being an AI. If a client asks, you answer. Lying about it destroys trust in a way no fast reply makes up for.

5. Start in draft mode. For a few weeks, let the AI write and you approve. When most suggestions go out unedited, it may make sense to let it reply on its own in specific chats.

Which model to use

For replying to messages, the gap between the big models is smaller than the gap between a good prompt and a bad one. Claude, GPT and Gemini all write well. What should drive your choice is price at your volume, speed, and whether your tool already integrates with it. Smaller, cheaper models handle triage and summaries well; save the bigger model for writing the reply.

Where WACrew fits

WACrew is an open-source level 3 agent. It reads the latest messages before writing, learns your style from your own messages and runs every reply through a code-based review that blocks robotic tells. You choose per chat whether it only drafts or replies on its own, and its MCP server works with Claude Code, Codex, Cursor and other clients. The code is on GitHub.