← Back to Blog
Support8 min read

Telegram FAQ Bot

Most of the messages you answer are not really questions. They are the same ten questions, arriving in a different order, from different people, forever. That is a solvable problem and it does not need AI.

Find the real list first

Do not write your FAQ from memory. Memory over-represents the interesting questions and forgets the boring ones, and the boring ones are the volume.

Open your last hundred messages, wherever they live, and tally them. Almost everyone who does this finds the same thing: a very short head and a long thin tail. Six to ten questions cover most of it, and half of those are price, location, hours and "is this available".

That tally is your build list, in order. Anything below about three occurrences in a hundred is not worth automating yet.

Why rules beat AI here

An FAQ is the one support problem where a language model is usually the wrong tool. The questions are known, each has exactly one correct answer, and that answer is something you want to have approved rather than generated.

Keyword rules cost nothing per message, always give the answer you wrote, and never invent a refund policy you do not offer. A model costs money on every message and introduces a category of failure that is expensive in a customer conversation. Keep the model for the tail, if at all, and see AI Telegram bots for where it does earn its place.

Commands and keywords do different jobs

  • Commands appear in a menu and serve the people who explore before they ask. They also give you a structure you can point at: "the price list is under Prices".
  • Keyword replies fire when a message contains a word you chose. They serve everyone who just types their question, which is the majority, and they work without the person learning anything.

Build both. The menu is for browsing, the keywords are for catching, and the same answer can sit behind both.

Choosing keywords that actually fire

The mistake is picking the word you use rather than the words your customers use. For each answer, write down every way somebody might ask it:

  • Price: price, cost, how much, fee, rate, charge, expensive.
  • Location: where, address, located, how do I get, directions, map.
  • Hours: open, closed, hours, today, weekend, holiday, what time.
  • Delivery: delivery, shipping, ship, send, courier, how long.

Include the misspellings you actually see, and include the short forms. People type "hw much" more often than you would like.

One caution: a rule fires when the message contains the word anywhere, so avoid triggers that appear inside ordinary sentences. "How" on its own will fire on almost everything and make the bot feel broken.

Writing answers that end the conversation

A good FAQ answer is complete. If the reply produces another question, it has failed and you have made things worse by adding a step. Four rules:

  • Give the number. "Prices start from X, a typical Y costs around Z." Not "prices vary depending on requirements", which is a non-answer everyone recognises.
  • Answer the question behind the question. "Are you open?" usually means "can I come now?". Include today's hours and the closing time.
  • Pre-empt the obvious follow up. The delivery answer should mention cost and time together, because whoever asked one was about to ask the other.
  • Sound like a person. The advantage of chat over a web page is that it does not read like a policy document. Write what you would say out loud.

The fallback is the most important message

Someone will ask something you did not plan for. What happens next decides whether they think your bot is helpful or an obstacle.

A good fallback admits it plainly, gives the working hours, and offers one clear way through: a button to leave a message, or your phone number. A bad fallback repeats the menu, or worse, says something confident and wrong.

Say when a human will reply and be honest about it. "We answer within a few hours on weekdays" sets an expectation people accept. Silence makes them message three more times.

The tail is a list, not a failure

Once a month, look at what people asked that nothing matched. That list is the most useful document in your support operation: it is either questions to add, or things wrong with your product that keep generating questions.

Most of the time it is the second one, and fixing the cause removes the question entirely, which is better than answering it faster.

What to measure

  • Messages you personally answered per week, before and after. The whole point, in one number.
  • Fallback rate. How often nothing matched. Above roughly a fifth means your keyword lists are too narrow.
  • Which answers get used, which tells you what customers actually care about and often what your website should say more loudly.
  • Repeat questions after an answer, which means that answer is incomplete.

A realistic build

  1. Hour one: tally your last hundred messages and pick the top eight.
  2. Hour two: write the eight answers. This is the slow part, because writing a complete answer is harder than answering ad hoc.
  3. Hour three: add keywords for each, five to eight variations, plus the menu and the fallback.
  4. Week two: read the tail and add two more answers. Then leave it alone.

Frequently asked questions

Does an FAQ bot need AI?

Almost never. Known questions with fixed answers are exactly what rules are for, and rules cannot invent a policy you never approved.

How do I find what people actually ask?

Tally your last hundred messages instead of writing the list from memory. The real list is shorter and duller than you expect.

Commands or keywords?

Both. Commands for people who browse a menu, keywords for the majority who just type their question.

What if the bot does not know?

Say so plainly, give your hours, and offer one route to a human. Never repeat the menu and never guess.

Answer the top eight questions once

Keyword replies, a command menu and a fallback that hands over to you, all editable from your phone. No code, no server, no per message cost.

Build Your Bot Free

Keep reading