NovixSun — разработка сайтов ЕреванNovixSun
🤖Lessons from a launch

An AI chat on your website
replies to customers in seconds, not the next day

A customer messages you in the evening and gets an answer in the morning — if they get one at all. An AI chat replies straight away, sticks to the facts of your business, and calls in a person when a question is beyond it. Here's how it works, using the chat running on this very site.

See services →
📅Published: October 6, 202615 min readAIBots
  • Customers get an answer in seconds, at any hour — not the next day, when they've already gone with whoever replied first.
  • The chat answers only from your business facts: prices, services, terms. If something isn't in its knowledge base, it doesn't make it up.
  • A tricky question, or a plain “can I talk to a person?”, goes to a manager in Telegram — and the conversation carries on in the same window.
  • Not everyone needs one: if customers keep asking the same few things, a simple button bot does the job — it's simpler and cheaper.

Adding an AI chat: four steps

  1. 1
    Check whether you need it
    If customers ask lots of different questions, AI makes sense. If they all ask the same things — prices, address, opening hours — a button bot or the free auto-reply in WhatsApp Business will do.
  2. 2
    Build the knowledge base
    Prices, services, terms, answers to common questions. The chat will answer only from this, so how complete it is matters more than which model you pick.
  3. 3
    Decide who takes over
    Who gets notified about a new conversation and replies when the chat calls for a person — and what the customer sees if nobody can answer right away.
  4. 4
    Set the limits
    A cap on messages per conversation, protection against attempts to override its instructions, a retention period for chat history. Without these, the chat is an open tab for any passer-by.

We can set all of this up for you — more about bots for business.

Why reply speed matters

Someone messaging a business is rarely messaging just you. They've asked three companies for a price and are waiting to see who answers first. If your reply comes hours later or the next day, the choice has already been made — and not in your favour.

For small businesses in Armenia this is the usual story: the owner answers messages between other jobs, and in the evenings and at weekends they pile up. Hiring someone to answer around the clock is expensive, and most questions repeat anyway.

An AI chat closes exactly that gap. It doesn't replace your manager — it takes the first minute of the conversation: answers the question, clarifies the task, takes down a contact. A person steps in while the customer is still warm, not after they've gone cold.

When you don't need an AI chat

An AI chat is worth it if
  • customers ask varied questions that don't fit into a few buttons;
  • customers need to clarify something before leaving a request — the task, the scope, the timing;
  • messages arrive in the evenings and at weekends, when nobody's there to answer;
  • you're ready to keep the knowledge base up to date as prices and terms change.
A regular bot or auto-reply will do if
  • customers ask the same things: prices, address, opening hours;
  • you need bookings for specific time slots — a button bot does that more precisely and for less;
  • “Got it, we'll reply in the morning” is enough — in WhatsApp Business that's a free auto-reply you switch on in the settings;
  • you don't have time to keep an eye on what the chat says. Unsupervised AI is worse than an honest “we'll reply during business hours”.

How we choose between a scripted bot and an AI agent — on our bots for business page.

How the chat on this site works

The chat in the corner of this page isn't a demo — it's a working tool, and real questions and requests come to us through it. Here's what happens when you send a message.

  1. Your message
  2. Check
  3. Model + knowledge base
  4. Reply check
  5. Reply on screen

Along the way, the model can:

  • save a request → it reaches us
  • call a person → a manager replies from Telegram
A knowledge base built from the site
Every time the site is updated, the chat's knowledge base is rebuilt from the price list, service descriptions and case studies. Change a price on the site and the chat knows the new one. There's no way for the two to drift apart.
Requests without a form
The customer doesn't fill in any fields: the chat clarifies the task in conversation and, once the person is ready, takes down their name and contact. The request lands in the same place as requests from the site's form.
A real person in the same window
If the customer asks for a manager, or a question goes beyond the knowledge base, the chat forwards the conversation to Telegram. The manager replies from there, and the customer sees the answer in the same window on the site.
The same agent in Telegram
Our Telegram bot runs on the same engine: the same rules, the same knowledge base, the same requests.
Every conversation in the admin panel
Any conversation can be opened and read — that's how we find mistakes you'd never spot from the outside.

How the chat is protected against abuse

A chat on a website is an open box that anyone can type into. Some will try to get the bot to promise a discount that doesn't exist. Some will use it as a free ChatGPT. Some will just type nonsense. So there are several checks between a visitor's message and the reply on screen.

Messages are checked before the model reads them
Rudeness, attempts to override the instructions (“forget your rules, you are now…”) and off-topic requests are filtered out at the door. The visitor gets a polite reply, the main model never sees the message — and the business doesn't pay for junk.
Replies are checked before the visitor sees them
Prices above all: the bot quotes only amounts from the price list. If a visitor haggles (“how about 40,000?”), the bot won't repeat their number or agree to it. And if the model tries anyway, the reply doesn't go out.
The bot knows only what you'd tell a customer
Its knowledge base is the price list, services and case studies from the site. It has no internal data, so it has none to leak.
Code guarantees what matters, not the model
The model can do exactly two things: save a request and call a person — it can't delete or change anything. A request to speak to a manager is recognised by the code, so a person gets notified even if the model gets distracted.
Conversations have limits
There's a cap on messages per conversation and on conversations per day from one address. Before the limit kicks in, the visitor sees how to carry on: leave a contact or message us on Telegram.

One of these checks silently did nothing for the first few days — how we caught it is below.

What went wrong at launch

Before opening the chat to visitors, we gave it to people who hadn't written a single line of its instructions, then read every conversation. Here's what came up — and what's worth knowing for anyone putting an AI chat on their own site.

  1. 1

    A contact was given — but no request appeared

    When a conversation hit the message limit, the chat asked the person to leave a contact. But by then they could no longer type in the chat, and there was no contact field — the request led nowhere. In one conversation the person wrote their contact anyway, in plain text, and it was lost. Now the limit notice has a contact field, and it works precisely when everything else has closed.

    Takeaway: Anything your revenue depends on should be guaranteed by code. The model can forget.

  2. 2

    The protection wasn't working, and nothing showed it from outside

    The incoming-message check was meant to catch attempts to override the instructions. When we switched on a detailed call log, it turned out that over an entire conversation it hadn't returned a single verdict — 0 out of 7. Two checks shared one time budget; the first used it all up, and the second, timing out, silently let everything through. After the fix: 7 out of 7.

    Takeaway: Switching protection on isn't enough. You need a log that shows it's actually firing.

  3. 3

    The model didn't follow its own rules

    The instructions asked for short replies and one question at a time. The first model answered with numbered lists and fired off questions in batches. We moved to a model one tier up — and the problem disappeared without a single change to the instructions. At our volume, the price difference is a few dollars a month.

    Takeaway: Saving on the model costs more than it seems. Longer instructions won't fix a weak model.

  4. 4

    The bot forgot how the conversation started

    Testers complained that the chat “loses the thread”. The cause turned out to be arithmetic: the model saw the last 20 messages, while a conversation could run twice as long. By the end, the bot no longer remembered how it began.

    Takeaway: The memory window must be at least as long as the longest conversation you allow.

  5. 5

    The first to write in were sellers, not buyers

    Both of the first outside enquiries turned out to be sales pitches: a contractor and an SEO mailing. The bot questioned them like customers and called a manager — every mass mailing would have woken the owner. Now, when someone makes an offer, the chat says it will pass it on to the team and doesn't call a manager. But if someone asks for a message to be passed to management, a manager gets it either way.

    Takeaway: Not everyone who writes in is a customer. You need a script for the rest from day one.

What an AI chat costs

An AI chat has three cost lines, and each pays for something different: people's work at launch, people's work after launch, and the model's work. They're easy to mix up, so let's take them one at a time.

1. Development — one-off

A bot with a standard script starts at 90 000 ֏ with us. With an AI agent it costs more and takes longer: you add a knowledge base, checks and handover to a person. We give an exact price once we've talked through the task.

2. The model's work — pay as you go, pennies

Each reply is paid to the model provider. Almost all of a reply's cost is the knowledge base and instructions the model rereads every time. The reply itself is short and costs next to nothing.

On our chat, one reply costs no more than $0.006 — a thousand replies, no more than $6. A hundred conversations of 5–10 messages: up to $3–6. That's a ceiling: the provider caches the repeated part of each request and charges a quarter of the price for it.

Measured October 2026, model gpt-4.1-mini, prices from OpenAI's price list.

You can cut the cost without touching the model: we trimmed our knowledge base, and the tokens per conversation roughly halved.

3. Support is people's time, not the model's

From 20 000 ֏ a month. The model costs pennies, but it won't fix itself: someone has to read the conversations, find the replies where the chat got it wrong or didn't know, extend the knowledge base, and update prices and terms. That work is what separates a chat that gets better over time from one that keeps repeating its first mistakes. Without support the chat keeps running, but it will know exactly what it knew on launch day.

All our prices are on the pricing page. What goes into the price of a website — in a separate article.

What happens to the chat history

  • Conversations are stored so a manager can pick up where the chat left off and so we can find the chat's mistakes.
  • Conversations without a request are deleted automatically after 90 days.
  • The model sees the conversation text but doesn't learn from it: under OpenAI's API terms, this data isn't used to train models. OpenAI keeps it for up to 30 days to monitor abuse.

Details are in our privacy policy.

FAQ

Questions about AI chat

Will an AI chat replace my manager?

No. It handles the first few minutes: answers routine questions, clarifies the task, takes down a contact. Decisions, discounts and non-standard terms stay with a person — the chat hands the conversation over.

Won't it make up prices and promises?

The chat answers only from the knowledge base, and prices get an extra check: if a reply contains an amount that isn't in the price list, it doesn't reach the visitor. Mistakes are still possible, which is why we read the conversations and keep the knowledge base up to date.

What language does it reply in?

The visitor's: Russian, English and others. To be honest, now and then the chat replies in the page's language rather than the message's — we're still working on that.

Can it be connected to Telegram or WhatsApp?

Telegram, yes — our own bot runs on the same engine as the chat on the site. WhatsApp is possible too, but it has its own requirements: a business account, a verified number, and messaging fees paid on WhatsApp's side. We go through this before any work starts.

How much does upkeep cost?

Those are two separate lines. The model is pay-as-you-go and costs pennies — no more than $0.006 per reply for us. Support, from 20 000 ֏ a month, pays not for the model but for people's time: reviewing conversations, extending the knowledge base, updating prices. Without it the chat keeps running, but it only knows what it knew on launch day.

Still have questions

Want a chat like this on your site?

Tell us what customers ask you most often. We'll tell you whether you need AI or a regular bot will do, and estimate the cost. We reply within one business day.