AI assistants and chatbots where they pay off

Artificial intelligence for business works with your data. It answers a customer, reads a document, finds a product by meaning, forecasts demand and writes product descriptions. Below are five such tasks, and for each one we say plainly when you don't need it. A build that won't pay for itself hurts both you and us. The last section covers where to start.

A website chatbot that answers from your data

What the bot does on the site and when it hands over to a manager

The bot lives in a widget on your website and, if needed, answers in Telegram as well. It takes its knowledge from your catalog, price list and FAQ. Before every reply the bot checks the knowledge base and never makes text up. The model isn't retrained on your data, it reads the relevant fragment each time.

Then the bot clarifies what the person needs and passes a sorted request to the manager in your CRM or Telegram. It doesn't invent answers to questions outside the knowledge base, it hands them straight to a person. If your customers call more often than they write, you need a voice bot. What a website bot can do in general is covered in our article why a business needs a website chatbot.

How much a website chatbot costs and what drives the price

There are two types of bot, and they are priced differently. A scripted chatbot guides the person with buttons and ready-made replies. It costs UAH 15,000–35,000 and takes 1–2 weeks to build. An AI consultant with a knowledge base answers in free text from your materials. It costs UAH 50,000–150,000 and takes 3–6 weeks.

Within the range, the price depends on the size of the knowledge base, the number of channels, CRM integration and handoff to a manager with the conversation history. Separately, each month you pay for support, UAH 3,000–10,000, and for model tokens at actual usage. We fix the amount after the brief and you pay in stages, so there's no need to pay for the whole project up front. Ordering a chatbot starts with a brief that shows your data and channels. To get an estimate, just fill in the brief online.

When you don't need a chatbot

Inquiries are few, and each one is unique. There's no knowledge base to answer from yet. Your customers prefer to call. In these cases a bot won't pay for its development. Start with a CRM or a simple business-card bot: a Telegram bot takes requests and invents nothing. Or do nothing at all, and we'll tell you so directly.

Document recognition without manual entry into 1C

A supplier sends an invoice as a PDF or a photo, and someone types it into 1C by hand. AI reads the document in any form: PDF, scan, a phone photo taken at an angle. It returns exactly the fields you need, for example the number, date, counterparty, line items, amounts and VAT. The data goes into 1C, your CRM or a Google Sheet, and we write the integration for your system.

Regular OCR turns a scan into text, and fields are then pulled out by template, so the system breaks on an invoice from a new supplier. AI reads the meaning, so a new layout of the same invoice doesn't stop it. Where AI isn't sure about a figure, the field is highlighted and goes to a person, the rest passes through on its own. Every document lands in a log that shows what was extracted and what went for review. We described the mechanics in more detail in the article how AI reads documents.

We start by reviewing your real documents and running a pilot on one type, say invoices. Other types are added once the pilot works. We quote the price after the review, because formats and volume set it. It's not worth it if you get a few documents a week or each one is a new type with different fields. AI will handle a new layout of a familiar invoice, but the list of fields is set in advance. And if the data already arrives in digital form, there's nothing to recognize, an integration is enough.

AI catalog search that understands the customer's words

A customer types "spring jacket with a hood", and the product card says "windbreaker, mid-season". Regular search looks for matching words, shows "nothing found", and the customer leaves. Semantic search works with the meaning of the query: it understands synonyms, descriptions by purpose and typos. If the exact product isn't there, it shows the closest one by meaning. We collected the typical reasons for empty results in the article why search can't find the product.

Technically, both the product and the query are turned into vectors, and search picks the nearest ones. We index your catalog or knowledge base: course materials for a school, instructions for support, internal rules for your team. Search returns only what you actually have and invents nothing. Add a product and the index updates, nothing needs retraining. The log shows queries with no results, that is, what your catalog is missing. A separate article covers semantic search on a website in more detail.

First we review the catalog, then run a pilot on your problem queries with a before-and-after comparison. We'll name the price after the review. Search finds and shows, a person makes the decision. If the system also has to answer or act in your CRM, you need AI agents. And for a catalog of a few dozen items with filters, regular search will do the job on its own.

Demand forecasting and predictive analytics on your sales

Several years of sales sit in spreadsheets, purchasing is planned by eye, and then leftover stock gets written off. Demand forecasting calculates what will sell and how much, taking season, day of the week and promotions into account. Sales analytics shows what brings margin and what people buy together. Inquiry data shows when load peaks and where customers drop off.

Predictive analytics starts not with a model but with a data audit: how much history there is and how cleanly records were kept. Then we bring 1C, the CRM and spreadsheets into one source, and often that is the main part of the work. We test the model on history, checking how accurately it would have predicted what already happened. We show a range and confidence limits rather than a single nice number. You get a dashboard with the metrics you already use, refreshed on fresh data. Where plain statistics are enough, we don't push AI.

It's too early if records aren't kept or history is short: seasonal demand needs several full cycles. In that case, set up data collection first, most conveniently through CRM implementation, and come back to forecasting later. The cost becomes clear after the audit, because it depends on the state of your data.

Product descriptions and meta tags generated and proofread by a person

A thousand product cards without descriptions, empty meta tags, a catalog that needs a second language. Here content generation takes on the volume. Product descriptions are generated from the fields in your database using a template that sets length, tone and what must be mentioned. First comes a trial batch that you approve, then the full volume. The finished texts go back into your catalog or CMS. Where to start is laid out in the article how to fill an online store with products.

The model tends to add a feature the product doesn't have. That's why a person proofreads every text and checks the facts, and we never skip this step. Google judges how useful a text is, not how it was produced, so we don't publish an unchecked batch. Sales copy, brand voice or an expert article are better handed to copywriting, where AI gives you a rough draft at most. We'll tell you the cost once we see your catalog.

Where to start and what happens to your data

Every lead from this website goes through our own Apros CRM. AI agents with 15+ tools work inside it, strictly within their access rights. If something breaks in the system, we are the first to feel it. We've been building websites and business systems since 2016, with more than 500 projects behind us. Offices in Kharkiv and Kyiv.

We connect models through the OpenAI or Anthropic API under enterprise terms, so your data isn't used to train public models. Access, storage terms and an NDA are fixed in the contract before we start. AI is added as one more service to your website, CRM or 1C, with nothing rewritten. Often it becomes part of wider business process automation.

AI development with us starts with a 30-minute call. You tell us which routine eats your time, and we tell you whether it's worth tackling and where to start. You can book it with the form below.