--- title: "How to automate customer replies with AI (email, WhatsApp, Telegram)" url: https://memocore.ai/memos/6ad88c6cbdec34b5f895d979 project: "AI in Everyday Life" tags: "business" updated_at: "2026-10-30T02:02:04.054Z" --- # How to automate customer replies with AI (email, WhatsApp, Telegram) Most customer messages repeat themselves: prices, timing, availability, address, order status. AI can answer them in seconds and leave you only the tricky cases. Below are two ways to do it: **partial automation** right in ChatGPT or Claude, and **full automation** with n8n. ## What you get - Routine questions are answered automatically or in one click - Answers are accurate: the AI takes them from your knowledge base instead of making them up - Complex and heated conversations go to a human ![](02-two-ways.png) ## Step 1. Build your knowledge base This is the most important step. Without a knowledge base, the AI will answer with vague phrases or get things wrong. **What to write down:** - frequent questions and the right answers to them - prices, delivery and return terms, opening hours - tone of voice: formal or casual - what the AI must never promise (discounts, deadlines, compensation) ![](01-knowledge-base.png) **Where to keep it:** | Option | Pros | Cons | |---|---|---| | **Google Sheets** | Simple: two columns, "question" and "answer" | Awkward for long texts and complex conditions | | **Notion** | Easy to write and edit; AI workflows can read it | Needs access setup; search by words, not by meaning | | **RAG on a vector database** (Pinecone, Qdrant, Supabase) | Most accurate search by meaning | You'll need a developer | | **Memocore** | Easiest, no setup; search by meaning built in | — | With Memocore, create a project, for example "Support", connect Memocore to ChatGPT or Claude, and just write in the chat: > *"Save our prices and delivery terms to the Support project: …"* If your knowledge base already lives in Notion, you can import it. ## Step 2. Write your reply rules The knowledge base answers *what* to tell customers. The rules answer *how*: in what tone, what never to promise, and when to bring in a human. You'll need them for both ways of automating. Here's an example. Replace the rules with your own, tailored to your product or service: ``` You are the support assistant for [company name]. Reply to customers on behalf of the company. Rules: - Before replying, search the knowledge base. Use only that information. If the answer isn't there, don't make it up — say you'll pass the question to a manager. - Tone: [friendly, polite]. Short, no corporate jargon. - Never promise discounts, compensation or deadlines that aren't in the knowledge base. - If the customer is unhappy, rude or asks for a refund — don't argue, apologize and hand the conversation to a human. - Reply in the customer's language. At the end of the reply add a tag: [AUTO] if the answer comes fully from the knowledge base, [HUMAN] if a person is needed. ``` ## Step 3. Partial automation: replies in ChatGPT or Claude The AI writes the reply; you paste in the customer's message and send the answer yourself. Good for getting started and for a small flow of messages. Setup takes about 10 minutes. ### Create a skill A skill is a saved instruction the AI follows when you call it. Both ChatGPT and Claude let you create one right in the chat. Send this to the chat: ``` Help me create a skill called "Customer replies" based on the rules below. The skill should look up the answer in our knowledge base ([the Support project in Memocore / the uploaded file]), write a reply to the customer following the rules, and add the [AUTO] or [HUMAN] tag. If anything is missing, ask me clarifying questions. Rules: [your rules from step 2] ``` The AI will ask a couple of questions, build the skill and offer to install it. ### Reply to customers When a customer message comes in, send this to ChatGPT or Claude: ``` Use the "Customer replies" skill. Customer message: [text] ``` Check the reply and send it to the customer. If it has the `[HUMAN]` tag, reply yourself. Work this way for a week. Every question the AI couldn't answer — add it to the knowledge base. ## Step 4. Full automation with n8n Messages are handled on their own, without you. n8n is an automation service where you build a workflow from blocks, no code required. ### How the workflow runs ![](03-workflow.png) 1. **A new message** from a customer (email, Telegram or WhatsApp) starts the workflow. 2. **The AI looks up the answer** in your knowledge base and writes a reply following your rules from step 2. 3. **Tag check.** If the reply has `[HUMAN]`, the message is forwarded to you. 4. **Reply to the customer.** At first the AI shows you each reply and sends it only after you approve it. Later you switch on automatic sending. ### Let the AI walk you through setup Service interfaces change often, so instead of screenshots that quickly go out of date, use this prompt. The AI will guide you step by step for your channel and your knowledge base. Fill in the brackets and send it to ChatGPT or Claude: ``` Help me set up AI auto-replies to customers in n8n. I'm not a developer: guide me one step at a time and wait until I confirm each step worked. My setup: - Channel: [Gmail / Outlook / Telegram / WhatsApp] - Knowledge base: [Google Sheets / Notion / Memocore / vector database] - n8n account: [already have one / not yet] - Mode: [I approve each reply before sending / automatic sending] Workflow logic: 1. A new customer message starts the workflow. 2. The AI looks up the answer in the knowledge base and writes a reply following the rules below. 3. If the reply has the [HUMAN] tag — forward the message to me on Telegram. 4. Otherwise — remove the [AUTO] tag from the reply. Then send the reply to me on Telegram for approval and send it to the customer only after I approve it, or, in automatic mode, send it to the customer right away. Rules for the AI in the workflow (tell me where to put them): [your rules from step 2] If I need to choose or sign up for something, explain it in plain words and recommend an option. Before each step, check the current n8n documentation and the docs of the service we're connecting: the interface may have changed. If I can't find a button or field, ask me what I see on the screen. ``` ## Step 5. After launch The workflow is built, but don't let it run on its own right away. ![](04-after-launch.png) 1. **Test it yourself.** Message your channel from another account as a customer: a simple question, a question that isn't in the knowledge base, and a complaint. Make sure the reply to the first one came to you for approval, and the second and third were forwarded to you. 2. **Run in approval mode for two weeks.** Read every reply before approving it and note where the AI got it wrong. 3. **Fix the cause, not the reply.** Wrong fact — add it to the knowledge base. Wrong tone or promised too much — refine the rules. 4. **Turn on automatic sending** once replies barely need edits. Ask the AI that helped you with setup to switch the mode. 5. **Once a week, review** a few random replies and every question that was forwarded to you. Add the answers to the knowledge base. ## Launch checklist - [ ] Knowledge base is in one place - [ ] Reply rules are written for your business - [ ] Skill is created, and you've checked reply quality for a week - [ ] n8n workflow is built and tested on sample messages - [ ] Two weeks in approval mode with no serious mistakes - [ ] Automatic sending is on - [ ] Unanswered questions are regularly added to the knowledge base --- Every new answer in your knowledge base is a question you'll never have to answer by hand again. In a month, AI handles routine questions around the clock, and you spend your time only on the customers who really need a human.