BOTS · SERVICE
WhatsApp & Telegram bot: AI answering on the channels your clients actually use
Working bot on WhatsApp Business or Telegram, tested and live. Appointment booking, FAQ, lead qualification, connected to CRM or RAG knowledge base. Flat fee CHF 1,800.
Researched & fact-checked by: DuneDive LLC · As of: 2026-05
What is a WhatsApp or Telegram bot?
A bot is an automated counterpart in a messaging app. When your clients write on WhatsApp or Telegram, a program – not a human – responds first. It understands the question, searches your knowledge base or CRM, returns a grounded answer, and hands over to a human as soon as the conversation steps outside the defined frame.
Technically there are two stable paths in May 2026. WhatsApp runs through Meta Cloud API (the simple path since late 2024; before that only the Business Service Provider route was possible). Telegram runs through the official Bot API, stable and free since 2015. Both platforms support webhook-based bots which we couple to your LLM gateway and CRM with Node.js or Python.
As a service from us: you name the use case (appointment booking at a fiduciary, FAQ for a practice, lead qualification for an architecture firm), we build the bot in 2–3 weeks, test it with real messages, go live on your phone number (WhatsApp) or under your bot name (Telegram). Flat fee: CHF 1,800.
Why it matters
WhatsApp has over 80 % reach in the DACH region in 2026 and is the most-used messaging app. A practice, firm or SME reachable via WhatsApp lowers the barrier to first contact significantly versus phone or email. Telegram dominates in eastern markets and is the preferred choice when bots are to be richer (inline buttons, mini apps).
The problem without a bot: messages arrive around the clock – evenings, weekends, lunch break. Answering all personally costs time or speed. Letting them sit costs leads. Bots close that gap without disappointing the clientele – provided the bot clearly states what it can do and hands over cleanly when needed.
Two hard numbers from our own practice. First: response time. A bot answers in 2–5 seconds – a human response in an SME averages 4–8 hours with wide spread. Second: triage share. 60–80 % of inquiries in a fiduciary practice are FAQ-shaped ("Which documents for the tax return?", "Where is receipt X?", "Is appointment Y possible?"). A well-built bot catches those; the remaining 20–40 % go to a human – and that human suddenly has room to breathe.
How we build it
The bot consists of four layers: receive (messaging API), understand (LLM with routing), know (CRM/RAG search), and answer (with escalation logic).
Receive: For WhatsApp the webhook runs through Meta Cloud API. You need a Meta business account, a verified phone number and approved message templates for outbound conversations. We help with setup – the process typically takes 5–10 business days due to Meta reviews. For Telegram setup takes 10 minutes: BotFather, token, webhook.
Understand: Each incoming message passes through a small classifier (LLM, e.g. Claude-Haiku or GPT-4o-mini) recognising intent – "appointment request", "status question", "complaint", "smalltalk", "out-of-scope". The branch then differs by intent.
Know: For FAQ branches the bot looks up a RAG knowledge base (Qdrant + LLM, see RAG service). For CRM branches it queries your CRM via API (mandate status, open items, calendar). Important: the bot cites the source when answering from a document – no invented answers.
Answer & escalation: Replies are short (1–3 sentences), polite, in the formal Sie form. Three hard rules: (1) If confidence is below threshold, the bot hands over to a human – visibly, with reason. (2) If the topic is sensitive (health, legal advice, money) it hands over always. (3) When a human is replying, the bot knows from the CRM status and stays silent.
The architecture runs on the same server as n8n and LiteLLM. Conversations are logged in a Postgres table – seven days of live logs for debugging, then pseudonymised. Telegram alerts on errors go to on-call.
From use case to go-live
- 01Use-case workshop (2 hours): list of inquiries, estimate triage share, define escalation points, write out-of-scope answers.
- 02Platform setup: WhatsApp Cloud API (Meta business account, verify number, submit templates, 5–10 days Meta review) or Telegram (BotFather, 10 minutes).
- 03Build bot architecture: webhook endpoint in Node.js or Python, intent classifier via LiteLLM, RAG or CRM integration, answer generator, escalation branch.
- 04Conversation logging in Postgres: 7 days raw, then pseudonymised. Telegram alert on errors or escalation to on-call.
- 05Test operation 2 weeks: 20–50 test messages from us plus real messages in a soft launch. Walk through logs, fold in edge cases.
- 06Go-live: flip the switch, inform management, 30-day guarantee active. 90-minute training for 2–3 internal staff.
When to use
A bot is the right choice when (a) your clients already want to reach you on WhatsApp or Telegram, (b) a high share of inquiries is recurring, (c) you want to improve response speed without adding staff.
Concrete cases we built or would build: appointment booking in a fiduciary practice (bot asks the matter, proposes three free slots from Cal.com, books, sends confirmation); FAQ for a law firm with RAG knowledge base (bot answers general questions on Swiss inheritance law from the guidelines, escalates concrete cases); lead qualification for an architecture office (bot asks project type, budget range, location, routes qualified leads to sales, politely declines unfit ones); status inquiry for an insurer (bot shows the processing state of a claim from the CRM, sparing a phone call).
The flat CHF 1,800 covers: one use case, setup on WhatsApp Cloud API or Telegram, integration with your LLM gateway and one existing system (CRM or RAG), 2 weeks of test operation with logs, 90 minutes of training. For multilingual bots (DE/FR/IT/EN) or multiple use cases an add-on is needed – we discuss it in the audit.
When not to use
A bot is the wrong choice when first contact must be human for substantive or regulatory reasons. Examples: first interview in criminal law, crisis intervention in psychotherapy, counselling vulnerable persons. Here a bot harms – it creates distance where closeness is needed.
It is also wrong when you do not have the volume. If your practice receives 3 inquiries a week, the CHF 1,800 plus running cloud cost is not justified. Rule of thumb: from about 30 messages per week a bot becomes economical.
Be careful in legally sensitive sectors. A bot delivering "advice" risks a supervisory question in regulated sectors (FINMA, medical professions, lawyer monopoly). We solve this by clear framing: the bot provides information, not advice; advice comes from the human. For FINMA awareness we recommend legal review of the bot scope upfront.
The bot is also unsuitable if your knowledge base is chaotic. A bot answering from contradictory documents is worse than no bot. Order knowledge first (RAG pilot, see RAG service), then bot.
Trade-offs
STRENGTHS
- Round-the-clock reachability – the bot answers in 2–5 seconds
- Triage relieves the team – 60–80 % of routine questions bypass humans
- WhatsApp Cloud API since 2024 without BSP middleman – direct, cheaper
- Telegram as a playground for rich bots: inline buttons, mini apps, multi-channel logic
WEAKNESSES
- WhatsApp setup takes 5–10 business days due to Meta review
- A dedicated phone number is mandatory
- For sensitive topics (law, health, money) clear boundaries are required, otherwise regulatory risk
- Without a clean knowledge base the bot hallucinates – RAG is a prerequisite, not an add-on
FAQ
Do I need a dedicated phone number for WhatsApp?
Yes. Meta requires a unique, verifiable number not concurrently registered in the WhatsApp app on a phone. Common solutions: a second SIM, a VoIP number (Vodafone, Skype, or providers like Twilio with WhatsApp number provisioning), or an existing landline with SMS reception for the OTP. We help with the choice during setup.
What does WhatsApp Cloud API cost ongoing?
Meta bills per 24-hour conversation, not per message. The rate depends on country and conversation type (service, utility, marketing, authentication categories). For Switzerland 2026, service conversations are around EUR 0.07–0.12 per 24h window, opened by the client. At 500 conversations/month that is EUR 35–60. Telegram is free.
How does the bot prevent hallucinations?
Three layers. First: for knowledge-based answers, RAG runs with a hard refusal instruction ("answer only when the source is in the context"). Second: a confidence threshold – if the model drops below it, it hands over to a human. Third: out-of-scope branch – certain topics are explicitly excluded ("I cannot provide concrete tax advice, a colleague will contact you").