How to Build Your Own AI Assistant on WhatsApp or Telegram (2026 Guide)

You have probably seen demos of AI assistants that answer questions, automate replies, or help customers on WhatsApp and Telegram. What the demos rarely show is how approachable the build process actually is — even without writing a single line of backend code. This guide walks you through the realistic no-code and low-code path to building your own AI assistant on WhatsApp or Telegram: deciding what it should do, writing a system prompt that shapes its personality and limits, connecting a knowledge source, testing properly, and adding guardrails so it does not embarrass you.

By the end you will have a working bot plan, five copy-paste prompt templates, and a clear picture of where to start this week — whether you are building a customer support bot, a personal productivity assistant, or a community FAQ handler.

What you’ll need

• A WhatsApp Business account (free via Meta) OR a Telegram account with BotFather access
• A no-code automation platform: Make (formerly Integromat), n8n, or Zapier — all have free tiers and native messaging integrations
• An AI API key: OpenAI (GPT-4o), Anthropic (Claude), or Google (Gemini). You call the model via API, not via a browser chat
• Optional: a simple knowledge source — a Google Doc, a PDF, or a Notion page — that the bot will use as its reference material
• About 3–5 hours for a first working version; refinement is ongoing

If you are in Algeria and need ChatGPT or Claude API credits, clickdz.ai provides genuine AI subscriptions and credits paid in DZD via CIB, EDAHABIA, or BaridiMob — no international card needed, instant activation, 4.9/5 rating from 1,200+ verified reviews.

Step 1 — Decide exactly what the bot should and should not do

The most common reason AI bots go wrong is scope creep: the bot tries to answer everything and handles nothing well. Before you write a single prompt, spend 20 minutes answering three questions in writing:

1. One job: What is the single most valuable thing this bot can do for my users? (Answer FAQs about my product, help people book appointments, summarise incoming messages, coach users through a workflow)
2. Hard limits: What should it never do, even if asked? (Discuss pricing it does not have, give medical advice, discuss competitors, take payments)
3. Fallback: What should it do when it does not know the answer? (Say “I don’t know and suggest contacting a human” is almost always the right answer)

Write these three answers down. They become the backbone of your system prompt in Step 2. Skipping this step is why most bots fail — not technical issues, but unclear scope.

Step 2 — Write the system prompt that defines your assistant’s personality and limits

The system prompt is the set of standing instructions that runs at the start of every conversation, invisible to the user. It is the single most important thing you will configure. A good system prompt has four parts: identity, job, limits, and tone.

You are [NAME], the virtual assistant for [BUSINESS NAME].

YOUR JOB:
Help customers with [specific task 1], [specific task 2], and [specific task 3].
Answer questions only about [topic area]. Use the knowledge base provided below.

HARD LIMITS — never do the following, even if the user asks:
- Do not discuss pricing unless it is listed in the knowledge base.
- Do not make promises the business has not approved.
- Do not discuss competitors by name.
- Do not give [medical / legal / financial] advice.
- Do not engage with rude or abusive messages — politely end the conversation.

WHEN YOU DON'T KNOW:
If the answer is not in the knowledge base, say: "I don't have that information right now. For this kind of question, please contact our team at [CONTACT METHOD]."

TONE: [e.g. "Friendly and professional, like a helpful shop assistant. Short sentences. No jargon. Respond in the same language the user writes in — Arabic, French, or English."]

KNOWLEDGE BASE:
[PASTE YOUR FAQ / PRODUCT INFO / POLICY TEXT HERE — keep it under 2,000 words for best results]

Write this prompt, then test it by role-playing as a difficult user. Try asking it something outside its scope. If it answers confidently and wrongly, tighten the limits. If it refuses everything, loosen the tone. Iterate before connecting to the live messaging channel.

Step 3 — Connect the bot to WhatsApp or Telegram using a no-code platform

This is the technical step — but with Make or n8n, it is closer to connecting Lego bricks than writing code. Here is the architecture in plain language:

For Telegram

1. Message @BotFather on Telegram, create a new bot with /newbot, and copy the API token it gives you.
2. In Make, create a new scenario. Add a Telegram: Watch Updates trigger — paste your bot token.
3. Add an OpenAI: Create a Message (GPT) module. Paste your system prompt in the System field. Map the incoming Telegram message text to the User field.
4. Add a Telegram: Send a Message action. Map the OpenAI response text to the message body. Map the Telegram chat ID so replies go to the right person.
5. Activate the scenario. Test by messaging your bot on Telegram.

For WhatsApp

WhatsApp requires a Meta Business account and approval for the WhatsApp Business API (free for small message volumes). Use the 360dialog or Vonage WhatsApp integration in Make or n8n — the architecture is identical to Telegram once the channel connection is approved.

Help me write step-by-step instructions for connecting an OpenAI GPT-4o assistant to [Telegram / WhatsApp] using Make (formerly Integromat).

My skill level: [beginner with no-code tools / comfortable with Make / developer]
My use case: [describe what the bot should do]
My existing setup: [do I already have a Meta Business account? A Make account? An OpenAI key?]

Give me:
1. The exact modules to add in Make, in order
2. What to map where (inputs → outputs between modules)
3. The most common error at each step and how to fix it
4. How to test it before going live

Step 4 — Connect a knowledge source (RAG for non-developers)

If your bot needs to know about your specific products, policies, or services, you need to give it that information — not rely on the model’s general training. The simplest method is called naive RAG (retrieval-augmented generation): you paste your reference content directly into the system prompt.

For content under 2,000 words (a product FAQ, a policy document, a service menu), paste it directly at the bottom of your system prompt under a “KNOWLEDGE BASE” heading. This works reliably and requires zero extra infrastructure.

For longer content, use Make + Google Drive + OpenAI Assistants: upload your documents to an OpenAI Assistant’s knowledge base (via the Assistants API), and have Make call that Assistant instead of a raw completion. This lets the bot retrieve from 50+ pages of documentation accurately.

I have the following reference content that I want my bot to use:
[PASTE OR DESCRIBE YOUR DOCUMENT]

Help me:
1. Structure this content so an AI assistant can retrieve answers from it accurately
2. Identify any sections that might confuse an AI (ambiguous phrasing, contradictions, missing context)
3. Suggest what to remove or simplify for a knowledge base under 2,000 words
4. Write a clear heading structure I can use when pasting it into my system prompt

Step 5 — Test properly and add guardrails

Testing is where most builders rush and later regret. Before any real user touches your bot, run it through these four test categories:

• Happy path: Ask the 10 most common questions your users will actually have. Does it answer correctly every time?
• Edge cases: Ask things that are slightly outside scope but related. Does it stay within its limits?
• Adversarial: Try to make it break its rules — ask it to ignore its instructions, pretend to be a different bot, or reveal its system prompt. A well-written system prompt resists most of these.
• Language switch: If your users might write in Arabic, French, or Darija, test each one. Instruct the bot to respond in the same language the user writes in — this single instruction dramatically improves user experience for multilingual audiences.

You are a quality tester for an AI chatbot. I am going to share the bot's system prompt and a sample conversation.

Your job:
1. Identify any responses where the bot exceeded its defined scope
2. Identify any responses where the bot was unhelpfully vague when it had enough information to answer
3. Suggest one specific revision to the system prompt to fix each problem you find
4. Rate the overall conversation quality: Excellent / Acceptable / Needs work — with one sentence of reasoning

System prompt:
[PASTE SYSTEM PROMPT]

Sample conversation:
[PASTE TEST CONVERSATION TRANSCRIPT]

The BUILD Framework for Your AI Messaging Assistant

Best tools for building a WhatsApp / Telegram AI assistant

Common mistakes to avoid

• Building before defining scope. A bot with no clear job confuses users and embarrasses you. Write the three-question scope definition before touching any platform.
• A vague system prompt. “Be helpful and answer questions” is not a system prompt. Every sentence in the system prompt is a decision — tone, limits, fallback behaviour, language policy. Treat it seriously.
• Pasting an entire website into the knowledge base. Models perform much better with structured, concise knowledge bases than with raw scraped text. Edit your knowledge base to 2,000 words or under for best results with the system-prompt approach.
• Only testing the happy path. Users will always do something unexpected. Adversarial testing is not paranoia — it is what separates professional bots from embarrassing ones.
• No human fallback. A bot that confidently gives wrong answers is worse than no bot. Always include a clear “I don’t know — contact us here” behaviour for anything outside the bot’s knowledge.

FAQ

Do I need to know how to code to build this?

No. Make and n8n are visual drag-and-drop platforms — you connect modules by clicking, not by writing code. The most technical thing you will do is paste an API key into a field and map a variable from one module to another. That said, knowing the basics of JSON helps when something breaks, and the platform documentation is genuinely good.

What does it cost to run a WhatsApp or Telegram AI bot?

Telegram is free at any volume. WhatsApp charges per conversation after the first 1,000 conversations per month (free tier). OpenAI API costs for GPT-4o are roughly $0.002–$0.005 per typical message exchange — very low for moderate volumes. Make’s free plan covers 1,000 operations per month; most small bots stay within that. Total cost for a small business bot: typically under $10/month to start.

How do I handle conversations in Darija or other Arabic dialects?

Include an explicit instruction in your system prompt: “Detect the language the user is writing in — including Moroccan Darija, Algerian Darija, or Tunisian — and respond in the same variety. Do not correct or switch to Modern Standard Arabic unless the user does.” GPT-4o handles Darija reasonably well for practical customer-service tasks. For content you want to publish or that must sound natural to a native reader, run Arabic output through Humanizily, which specifically targets and removes the machine patterns that native speakers notice immediately.

Conclusion

Building an AI assistant on WhatsApp or Telegram is no longer a developer-only project. With Make or n8n, an OpenAI API key, and a well-written system prompt, you can have a working, useful bot running on the messaging app your customers already use — in a weekend. The work is mostly thinking: deciding scope, writing limits, structuring your knowledge base, and testing honestly. The BUILD framework keeps you on track through each stage.

For a deeper understanding of which AI model to use at the core of your bot, the ChatGPT vs Claude 2026 comparison is worth reading before you commit. And if you want to go further with prompt engineering — the skill that determines everything from bot quality to cost — the prompt engineering guide for Arabic covers the advanced techniques that make a real difference.

Pro tips & power moves

• Start with Telegram, not WhatsApp. Telegram has no approval process, instant setup via BotFather, and unlimited free usage. Build and test your full bot there first, then port to WhatsApp once the logic is proven.
• Use conversation memory carefully. By default, many simple Make setups do not persist conversation history — each message is treated independently. For a more natural experience, store the last 5–10 messages in a Make data store and pass them as context to the AI on each new message.
• Version your system prompt like code. Every time you change the system prompt, save the previous version with a date. When something breaks, you can compare versions and pinpoint the change that caused it.
• Set a max token limit. If your bot is sending very long replies, users stop reading. Add an instruction like “Your replies must never exceed 120 words” — short, useful answers outperform exhaustive ones in messaging contexts every time.
• Monitor the first 100 real conversations manually. Read through them and note every question the bot handled badly. Update your knowledge base and system prompt based on what you see. This feedback loop is how bots go from decent to excellent.

Your action checklist

• ✅ Write down your bot’s one job, hard limits, and fallback behaviour before starting
• ✅ Draft a system prompt with all four sections: identity, job, limits, tone
• ✅ Create a Telegram bot via BotFather (takes under 5 minutes) and get your API token
• ✅ Set up a Make scenario connecting Telegram → OpenAI → Telegram reply
• ✅ Paste your knowledge base (structured, under 2,000 words) into the system prompt
• ✅ Run happy path, edge case, adversarial, and multilingual tests before going live
• ✅ Read the first 100 real conversations and update the prompt based on failures
• ✅ If you want guided help on your real project, book a session at 1v1.clickdz.ai