Skip to content
dvnt.io

Case study · AI · chatbot · CRM + calendar

An AI chatbot that answers from your website's content and books calls.

Vi answers only from what the company has published, saves the lead in the CRM and offers real slots from the calendar — after hours too.

Vianswers from dvnt.io contentAI
  1. How much does a web application cost?
  2. A web application costs from PLN 30,000 — small PLN 30–60k, medium PLN 60–120k, complex PLN 120–300k+. An analysis gives you the exact price.↳ /co-budujemy/…/aplikacje-webowe
  3. Can we talk tomorrow?
  4. A person will join during business hours. I can book a call for you right now in one of the free slots.

// Challenge

160 articles and a contact form

dvnt advises companies on how to use AI and implements it for clients. Yet for a long time its own website worked like most small-business websites: more than 160 articles, service offers, price calculators — and a contact form that leaves you waiting for a reply.

A measurement at the end of September 2026 showed what that means in practice. Of the contacts that reached the company's CRM over 120 days, 92% had never interacted with the website. People found the company by other routes. The website had the answers but could not give them.

On the other side is the visitor. They want to know whether the company builds online shops, roughly how much it costs and when they can talk to someone. In the evening or at the weekend, the most they can do is leave an email and hope someone gets back to them.

An AI chat seemed the obvious answer, but it carried the risks dvnt writes about for its clients. A bot that makes up a price does more harm than no bot at all. From 2 August 2026, the AI Act requires that the person knows from the start they are talking to AI. Conversations carry personal data. And a heavy chat script slows the website down for everyone, including people who never open it.

The assumption written down before the first line of code was simple: a conversation at 11 pm should end with "an entry in the calendar with an email confirmation and a reminder, not a promise".

// Approach

It knows only what we've published

Vi answers only from what the company has published. Every page, article and glossary entry published in Polish goes into the knowledge base the moment it is published. A page that is unpublished, redirected or hidden from search engines disappears from it in the same way. Nobody "trains" anything by hand, and Vi's knowledge is exactly what the visitor can see. In a production test, every amount and every timeframe in Vi's answers was backed by published content.

An assistant that says only what is on the website also keeps an eye on the website itself: in the first test it caught draft notes that had slipped into published pages, and they were corrected the same day.

Vi says what it is and knows when to hand the conversation over. Before anyone types anything, they see: "Vi is an AI assistant. You can ask to speak to a person at any time" — plus a link to the privacy policy. The name is there to help, not to pretend to be human; the prompt forbids that.

The conversation goes to the team when someone asks for it, when the knowledge base has no answer, when an existing client writes, in the case of a complaint or an invoice dispute, and when the enquiry looks like a concrete project. A general question about VAT or GDPR stays with Vi — tested on the real model. During working hours a person replies and the bot goes quiet. After hours, Vi says plainly that nobody will join right now and offers three available video call slots. It books them through the same module as the website, so the same calendar event is created, with an email and a reminder.

Models chosen like items in a budget. The conversation is handled by Claude Sonnet at a low effort level: it writes Polish well and understands website excerpts supplied as context. Assessing the enquiry — project type, budget, timeframe, contact details, a score from 0 to 100 — is done by the cheaper Claude Haiku, because it is filling in a form, not holding a conversation, and it happens after every message. If Anthropic does not respond or the model refuses, an OpenAI model takes over the reply. If all of them fail, the visitor gets a prepared message and the conversation goes to a person.

Data where it is needed, and no further. The website does not connect to the chat until someone clicks the button. The full conversation record stays in the company's own customer service inbox. Only a summary with a link goes to the CRM, without the transcript. The model providers work under data processing agreements and, under their terms, do not train models on this data. A person always decides on the offer.

Safeguards

  • GUARD_01 · KNOWLEDGE

    Answers only from published content

    Every answer has a source — a link to the page it comes from.

  • GUARD_02 · HUMAN

    Hands the conversation over when needed

    With the chat history, so nobody has to start from scratch.

  • GUARD_03 · LIMITS

    Limits set in the system

    The token budget and the number of steps live in the configuration, not in the prompt.

  • GUARD_04 · PERFORMANCE

    Loads only after a click

    0 requests to the chat server before the click.

When it hands over to a person

  1. 01When someone asks to talk to a person
  2. 02When the knowledge base has no answer
  3. 03When an existing client writes in
  4. 04In a complaint or an invoice dispute
  5. 05When the enquiry looks like a real project

// results · figures

What came of it

An answer at any hourOutside working hours, Vi says plainly that no one can join right now, tells the visitor when that will be possible and offers a time for a call
24/7
Real call slots instead of "we'll get back to you"From the same calendar and by the same rule as booking on the website; a slot taken in the meantime produces a new list, never a false "booked"
3
Connections to the chat server before a clickThe website loads nothing from the chat until the visitor opens it (browser measurement, 13 October 2026)
0
Model providers in the response chainClaude from Anthropic and, if the model fails or refuses, an OpenAI model; in testing, conversations carried on despite a misconfigured key
2

Since October 2026, Vi has been answering on digitalvantage.pl at any hour. A visitor asking about an online shop gets an answer with prices taken from the published pages, and when they want to talk — a specific time slot instead of a promise.

The company receives the enquiry already organised. When someone gives an email address or phone number, a contact is created in the CRM, and once it is clear what the project is about — a sales opportunity with a score. In a production test, that is exactly how a conversation with a sample lead ended: a contact and an opportunity in the CRM, a score above the threshold, and a booked slot that immediately disappeared from the available ones. Whoever takes over the conversation sees a summary, the opportunity stage and the score in the inbox, so they do not have to read everything from the beginning.

Before launch, Vi went through 12 conversation scenarios in production — from a question about price to an attempt to extract the bot's instructions — and model failures are reported automatically.

Messenger and Instagram are already connected to the same inbox, and Vi is assigned to them with the AI notice. It will start replying publicly once Meta has verified the app; WhatsApp is awaiting the same verification.

There are no business results yet — an observation period of at least four weeks with a weekly review of conversations is under way. The answer arrives in full after 5–9 seconds rather than word by word. Enquiry assessment does not yet have a fallback model. Automatic deletion of records after 12 months, as promised in the privacy policy, has yet to be built.

15 minutes. Zero selling.

Want an assistant like this on your site? In 15 minutes we'll tell you whether you need a chat, an automation, an agent — or nothing at all.