An AI assistant on your website: where it really helps and where it gets in the way

More and more often people come to me with the same request: "put an AI on the site so it answers customers". Sometimes that is a sensible idea. Sometimes it is a way to spend money and end up with a polite chatterbox on the site that mixes up prices and annoys people. Below I look at where the line is, using the assistant I built for my own site, profweblab.ru, as the example. This article is for small business owners, so there is no code here, and the only numbers are ones I can stand behind.

What the assistant on my site does

First, how the site itself works. It has a brief sheet. A visitor picks what is not working for them (for example, leads get lost, or customers write at night and nobody answers), ticks the project type and the extras they need, and the site immediately shows a price and timeline range. At the bottom of the sheet there is a form: the person leaves their contacts, a PDF proposal arrives in their inbox, and I get a message in Telegram.

The assistant lives behind an "Ask AI" button in the corner of the screen. You can talk to it in plain words: "I run a small clinic and want patients to book appointments on their own". It answers in a few sentences, quotes "from" prices from my price list and ticks the matching items on the brief sheet by itself. When the person says they are ready to send a request, the assistant opens the sheet and points to where the contacts go.

One detail everything else depends on: the assistant does not make up prices. It sees the same price list as the brief sheet, and the total is calculated not by the model but by the site, with an ordinary formula. The model only picks items and writes text. If it picks the wrong item, the person will see it and untick the box themselves.

Where the assistant actually helps

Repetitive questions

How much does it cost, how long will it take, do you handle this kind of task. If you have a price list and clear rules, a model is good at retelling them in human language and for the specific question asked. The visitor does not have to read the whole services page to find one line.

Collecting the details

Many people find it easier to describe their task in their own words than to deal with a form with twenty fields. The assistant turns the story into structure: project type, required features, timeline. As a result, what reaches me is not "make me a website" but a request with ticked items, and the first call starts with the substance.

Early filtering

If someone asks about something you do not do, the assistant says so right away. If their budget is far below your "from" prices, they find out in the first two minutes rather than after two calls. That saves time for both sides, and nobody feels awkward.

Evenings and weekends

People often choose a contractor, a clinic or a craftsman in the evening, when they are finally free. The assistant answers at any hour, and the request can calmly wait until morning.

Where it gets in the way, and what it must not be trusted with

A language model writes plausible text. Plausible does not mean correct. So there are things I do not trust it with at all.

• Prices, discounts and timelines beyond the price list. If the model "promises" a discount, the customer will come to you with that promise. My assistant's instructions say it directly: prices only from the list and only with the word "from", no discounts, no adding things up on its own. And the site still calculates the sum.
• Collecting contacts in the chat. There is a form with consent to data processing for phone numbers and email. The chat is for talking.
• Diagnoses, legal and financial advice. For a clinic or a lawyer this is the main risk. The assistant can tell people which services exist and how to book, but it should not assess symptoms or someone's chances in court.
• Pretending to be a human. The window is honestly labelled as AI. People are fine with that as long as there is a way to call a real person next to it.
• Off-topic conversations. Visitors test the assistant: they ask it to write poems, to "forget all the rules", to quote a price ten times lower. My assistant brings the conversation back to the topic and does not do those things. Nobody can guarantee the model never slips. So what matters more is this: everything with consequences (the price, the ticks on the brief) is double-checked by the site itself.

Something people mention less often: an assistant can get in the way just by how it behaves. If the window pops up over the page by itself after five seconds, people close it in irritation. On my site it is a button in the corner that does nothing until it is clicked. And the assistant answers briefly, in one to four sentences. Hardly anyone reads long answers on a website.

Handing over to a human

In my view, this is the most important part of the whole thing. The assistant panel always shows a "Call Evgeny" button and my status: "online" or "will reply when he's free". The person clicks it, and the conversation moves to me.

I get a Telegram message marked "Called from the AI assistant". It contains the last lines of the conversation with the assistant and what the visitor has already ticked on the brief sheet. I do not need to ask what it was about. I reply straight from Telegram, and the answer appears for the visitor in the same window. While the live conversation is going on, the AI stays quiet and does not interfere.

The button works even when the assistant itself is unavailable: the model did not answer, the daily limit ran out, something broke at the provider. This is a matter of principle. If the AI goes silent, the person should not hit a wall. They see an honest "the assistant is unavailable right now" and two ways forward: fill in the brief themselves or message me.

If you are planning an assistant of your own, do not start by choosing a model. Start with the question: who will answer real people, and where, when they ask for it. Without an answer to that, the assistant turns into an answering machine that only delays the conversation.

What it costs

Language models are paid for by the amount of text, counted in tokens. You pay both for what you send to the model and for what it answers. For Russian models, YandexGPT and GigaChat, the order of magnitude is from a fraction of a ruble to a few rubles per thousand tokens, with lite versions cheaper than full ones. Rates change, so open the provider's current price list before you calculate anything.

The main thing to understand about spending: with every visitor question, the model receives more than just the question. It receives the assistant's entire instructions, your price list and the whole conversation so far. Long instructions are paid for in every single message. When I worked this out, I cut my assistant's instructions by about 40 % while keeping every rule.

How I limit spending on my site:

• the main model runs within a free allowance, the backup one is paid;
• the paid model has a monthly cap in rubles: once spending reaches it, the site stops calling it;
• one address can ask no more than 20 questions in 10 minutes;
• there is an overall daily limit of answers for the whole site.

When a limit is reached, the assistant says exactly that: that's all for today, fill in the brief yourself or message me on Telegram. The site keeps working as usual.

For a small site, the model requests themselves usually cost little compared to development and, especially, maintenance. Maintenance is what people often forget: prices change, services get added, and the assistant has to learn about it the same day. On my site the instructions are assembled from the same database as the brief sheet. I change a price in the admin panel, and the assistant quotes the new price from the next message. If your price list is pasted into the instructions by hand, sooner or later the assistant will start quoting old prices.

Personal data

Everything a visitor writes to the assistant goes to the model provider's servers. That is a transfer of data to a third party, and it is worth thinking about in advance, not after a complaint.

What I did on my site:

• the assistant does not ask for name, phone or email, there is a separate form with consent for that;
• conversations are not stored on the server, only the number of answers and the spending are counted;
• the conversation history is kept in the visitor's browser while the tab is open;
• the model receives nothing except the price list, the service descriptions, the ticks on the brief sheet and the conversation itself.

General advice I give clients. Choose a provider that processes data in Russia: for Federal Law 152-FZ on personal data this makes life noticeably easier. Do not pass the model anything it does not need to answer: your customer base, order history, phone numbers. If the assistant has to answer about a specific order, let the site find the order itself and pass the model only what is needed, without extra fields. Describe the assistant in your privacy policy. And remember that a visitor may type a phone number or a diagnosis into the chat even if nobody asked. For your particular field, it is worth consulting a lawyer once.

How to tell whether the assistant helps

"People like it" is not a metric. I would look at these things:

• how many conversations a day, and how many of them end in a request;
• how often people call a real person, and at what point in the conversation;
• whether requests have become more complete: fewer clarifying questions on the first call or not;
• how much a month of the assistant costs, and how many requests came through it in that time.

My site has an "Under the hood" page that shows how many answers the assistant gave today and how much was spent this month. I do not store the conversations themselves, so I judge quality by the requests and by the times I was called in. If you want to read conversations, that is data storage, and it has to be reflected in your policy.

And the cheapest test worth doing before launch: ask the assistant twenty questions your customers really ask you, plus a dozen awkward ones. Ask for a discount, ask about a service you do not offer, try to get it talking about the weather. If at this stage it lies confidently, it will lie to visitors too.

When to add one, and when not to

An assistant makes sense if you have a clear price list or rules, lots of repetitive questions, and a person who will pick up the conversation when needed. Then it saves time for you and for your visitors.

I would not add one if every price is individual and the assistant simply has nothing to say, if you have few visitors and it is easier to answer yourself, or if a mistake is expensive and you cannot restrict what the conversation is about. In those cases a good request form and a quick reply in a messenger work better than any model.