An AI assistant that answers your customers on WhatsApp or Telegram: what it costs, and what it shouldn't do

Yes, a small business can have an AI assistant that answers customer questions on WhatsApp or Telegram, asks the qualifying questions you’d ask, books the appointment and hands a warm lead to a person. Running it is cheap: with a current small model, the AI part costs under a cent to a few cents per conversation. Telegram charges nothing for bot messages. WhatsApp changed its pricing on 1 October 2026 and now charges for replies beyond the first 1,000 a month per number.

The expensive part isn’t running it. It’s building it so that it only says things that are true about your business, knows when to stop and pass the customer to you, and can’t be turned into a free chatbot for strangers on your bill. That’s what this post is about.

What an assistant like this does well

It’s good at the repetitive front of the conversation — the part you or your staff answer twenty times a week:

• Answering from your own information: prices, opening hours, areas you cover, what’s included, how long things take, your cancellation policy.
• Qualifying: budget, location, dates, what they actually need — the questions you’d ask before deciding whether a call is worth it.
• Booking: offering free slots from your calendar and confirming one.
• Handing off: passing the conversation, with a short summary, to a person when the customer is ready to buy, upset, or asking something it shouldn’t answer.

This is what businesses are paying for right now. On Upwork this week, a real estate company posted a $4,000 budget for exactly this — an assistant on Telegram and WhatsApp that answers investor questions, qualifies budget and goals, and hands warm leads to brokers. A clinic posted for an AI receptionist working inside its patient system. Searches like “AI automation for small business” and “WhatsApp bot for business” now autocomplete by industry: real estate agents, restaurants, HVAC, healthcare.

What it shouldn’t do

Invent things. A language model will produce a confident answer whether or not it knows one. Asked about a discount you don’t offer, it may offer it. The fix is design, not a better model: the assistant answers only from a document you control, and when the answer isn’t there, it says so and hands over to a person.

This matters more than it sounds. In February 2024 a Canadian tribunal ruled in Moffatt v. Air Canada that the airline had to pay out a bereavement-fare refund its website chatbot had wrongly promised, and rejected the argument that the chatbot was responsible for its own words. What your assistant says in your name is treated as what you said.

Take payments inside the chat. Send a payment link to your normal checkout instead. Card details have no business in a chat log.

Give advice you’d need a licence for. Medical, legal, financial: the assistant can book the appointment, not replace it.

I learned the “only from your documents” rule building VesselRef, a system that answers engineers’ questions from ship manuals and drawings. The answers are only useful because each one shows the page it came from. A customer assistant needs the same discipline, just with a smaller document.

What it costs to run

Three costs: the AI model, the messaging channel, and a server.

The model. Assume a typical customer conversation of six back-and-forth messages. Each reply sends the model your business information (say 8,000 tokens — roughly 15 pages of text, cached after the first message), the conversation so far, and gets back a short answer. At Anthropic’s published prices on 28 October 2026:

Model	Price (input / output per million tokens)	Per conversation (approx.)	1,000 conversations a month
Claude Haiku 5.5	$0.10 / $0.50	under 1 cent	about $5–10
Claude Sonnet 5.5	$2 / $10	about 3–5 cents	about $30–50

For answering questions from a fixed document, the small model is usually enough. I’d start there and move up only for the conversations where it demonstrably falls short.

The channel.

• Telegram: free. Bots message users at no cost, with a rate limit of about one message per second per chat that a customer conversation never comes near.
• WhatsApp: since 1 October 2026, replies inside the 24-hour window after a customer writes to you are free only for the first 1,000 delivered messages per month, per phone number. After that, each reply is charged at your customers’ country rate. Messages you start yourself — reminders, confirmations, marketing — use approved templates and are charged separately.

That WhatsApp change is easy to underestimate. 1,000 messages sounds like a lot, but at six replies per conversation it’s about 160 conversations a month. An assistant that answers in one clear message instead of three short ones costs a third as much on WhatsApp — that’s a design decision, and it should be made before launch, not after the first invoice.

The server. The assistant needs somewhere to run: a small VPS, typically €5–30 a month, or a server you already pay for.

So for a small business, the running cost is usually tens of euros a month, not hundreds. The build is the real investment, and the build is mostly not the AI.

WhatsApp or Telegram?

Go where your customers already are. That’s the whole decision, but some practical differences:

• WhatsApp dominates in most of Europe, Latin America and much of Asia. It needs a verified Meta business account, approved message templates for anything you send first, and the per-message pricing above.
• Telegram is free, a bot can be set up in an afternoon, and it’s strong in Ukraine and Eastern Europe.
• A chat on your own website costs nothing per message and needs no third-party approval, but it only reaches people already on your site.

One assistant can serve all three. The logic sits on your server; each channel is a thin connection to it.

The parts that cost money if you skip them

I run an AI chat tutor in production inside ProWord, my language-learning app on the App Store and Google Play. The model calls are cheap. What protects the bill is everything around them:

• The API key never leaves the server. If it’s inside an app or a web page, someone will find it and run their own traffic through it.
• A rate limit per user. ProWord’s backend caps every client at 60 requests a minute. Without a limit, one script can call your assistant in a loop all night.
• A hard monthly spending cap at the model provider, with an alert well before you hit it.
• Cache what repeats. ProWord’s text-to-speech audio is generated once and stored; every later user gets the stored file, so each phrase is paid for once. The same idea applies to the twenty questions every customer asks.
• A narrow job. An assistant told to answer only about your business, and to decline everything else, can’t be used by strangers to write their homework on your account.

What I’d build first

1. Write down the 30 questions customers actually ask, from your inbox and WhatsApp history — not the ones you think they ask.
2. Write the answers in one document. Prices, areas, times, policies. This is the assistant’s entire knowledge, and keeping it current is the only ongoing work.
3. Make it answer only from that document, and hand over when it can’t.
4. Build the hand-off before anything clever. A person gets a notification with a three-line summary and takes over the chat.
5. Read the conversations every week for the first month. Every bad answer is a missing line in the document.
6. Add booking, CRM entries and reminders after that — once the answers are right.

The honest limitation

An assistant answers the repetitive questions; it doesn’t close the sale and it doesn’t replace the person who knows the business. If your customer conversations are mostly unique — custom quotes, complicated problems, negotiation — the assistant’s job shrinks to collecting the details before you reply, and that may not be worth building. It also isn’t fire-and-forget: when your prices or policies change and the document doesn’t, the assistant will confidently tell customers the old ones.

Questions people ask

Does it need ChatGPT? It needs a language model, and there are several good ones — OpenAI’s, Anthropic’s Claude, Google’s Gemini. For answering questions from a document the differences are small; price, data terms and how well it handles your customers’ languages matter more than the brand.

Can it reply in several languages? Yes. Current models reply in the customer’s language by default, and your document can stay in one language. Test the languages your customers actually use before launch — quality varies by language.

Is my customers’ data safe? It depends on the setup. Use a model provider whose API terms say your data isn’t used for training, sign their data processing agreement if you’re in the EU, and don’t send the model anything it doesn’t need to answer the question. If your data can’t leave your building at all, the model can run on your own hardware instead — that’s slower and costs more up front, but it’s the setup I build for clients with confidential documents.

How long does it take to build? A first version that answers from your document and hands off to a person is a matter of weeks, not months. Most of that time is getting the answers right, not the code. Booking and CRM integrations add to it depending on which systems you use.

---

If you’re answering the same customer questions every day and want to know whether an assistant is worth it for your business, send me a short description for a free project review. Tell me which channel your customers use, roughly how many conversations you get a month, and what you’d want it to handle. You’ll get a straight answer — including “not worth it” when that’s the answer.