It often happens like this: a plugin goes onto the site, answers a few questions confidently, gets one of them wrong in front of a customer, and is switched off within a month, with the verdict that chatbots do not work. The chatbot in that story knew nothing about the company and did not know it knew nothing, and a chatbot like that always ends the same way.
This piece is about the other kind: an assistant that answers from your company’s own knowledge, under your name, and hands a question over when it reaches the edge of what it knows. I build them, and I run one of my own, Asystent PV, for solar installation companies. Before reading further, ask it something at asystentpv.pl. A few answers in, the rest of this article will be easier to judge.
Start with your customers’ questions
Before anyone talks about models or prices, write down the questions your office answers every week, the ones somebody sighs at because it is the fourth time since Monday.
For a solar installer the list writes itself: how net billing works this year, whether a battery is worth it, which grants are still open, how many panels fit the roof, how long installation takes. A dental practice has its own ten, a heating installer another ten, an accounting office another.
If you can write ten such questions in five minutes, an assistant has something to take off your people. If you struggle to find three, you probably do not need one, and a better FAQ page will do the job for a fraction of the money.
Calculate it on your own numbers
The calculation fits on a napkin, and it is worth doing before any conversation with a supplier, including me. Three numbers:
- how many times a week someone explains the same thing (by phone, email or chat)
- how many minutes one such explanation takes, counting the time to get back to what they were doing
- what an hour of that person costs the company, gross salary plus contributions, not the net figure
A worked example, the one my own calculator opens on:
| Explanations a week | 30 |
| Minutes each | 8 |
| Hours a month | about 17 |
| Cost of an hour | PLN 75 |
| Monthly cost of repeating yourself | about PLN 1,300 |
If the assistant costs less a month than that last line, and takes even half of those conversations, it pays. If you explain the same thing five times a week for two minutes, the result comes out negative and you should not buy one, whatever a supplier’s calculator says.
There is a second line that does not fit in the table: the enquiry that arrives at eleven at night. Nobody answers it until morning, and by then the customer has asked two other companies. It is harder to put a number on, and often worth more than the hours saved.
How a company assistant differs from a plugin
A generic chat plugin connects a language model to your website and little else. It has read the internet and never seen your price list. Asked about your lead times, it will produce a plausible sentence, and plausible is the problem. A wrong subsidy amount or a wrong delivery date, given under your name, costs more than no answer at all.
An assistant built on your company’s knowledge should do five things a plugin does not:
- answer from your offer, your prices, your service area and your own wording
- say plainly when it does not know, and pass the question to a person with the customer’s contact
- understand a question typed on a phone in a hurry, with typos and without Polish characters
- keep a log of what people asked, so you can see which questions come back and at what time of day
- stay current, because the rules in most industries change every year and a model’s training does not
The fourth point is the one owners end up valuing most. After a month the log shows what customers want to know, and it often differs from what the website was written to tell them.
The most important feature: “I don’t know”
If you test one thing before buying, test this. Ask the assistant something it cannot know: the date of your next free installation slot, the price of a part you do not sell, the name of the person on duty. A good assistant says it does not have that information and offers to pass the question on. A bad one invents an answer.
My own assistant answers seven of eight questions typed in by hand. The eighth, how long installation takes, gets an honest “I don’t know”, because that depends on a particular company’s calendar. At a client’s, that answer comes from the client’s own data added at setup. I treat that eighth answer as a feature, and I would be suspicious of any assistant that scores eight out of eight.
Customer data, GDPR and the AI Act
A chat conversation is a conversation with a customer, and customers type in their names, phone numbers and addresses whether you ask them to or not. Three questions to put to any supplier:
- Where are the conversations stored, and for how long? Expect a named place and a retention period in the answer.
- Is there a data processing agreement? Under Article 28 of the GDPR, a supplier who processes your customers’ data on your behalf needs one. If they have never heard of it, look elsewhere.
- Does the data end up in anyone’s marketing tools? The answer should be no. The conversations belong to your customers, and the supplier may use them only to run your assistant.
Since 2 August 2026 the EU AI Act also requires that a person talking to an AI system is told so. In practice that is one clear sentence at the top of the chat. It costs nothing and it builds trust, because nobody likes discovering halfway through that the helpful consultant was a program.
What it costs
Three price levels, very roughly. A ready-made plugin costs from a few dozen to a few hundred zloty a month and knows nothing about your company. A fully custom build, with integrations into your CRM and booking system, starts in the thousands and depends on the scope. In between sits an assistant built on existing sector knowledge and tuned to one company.
For a reference point, my version for solar companies costs PLN 3,900 net to set up and PLN 690 net a month, with the sector knowledge already in place and the company’s own details added at setup. For another sector the setup takes longer, because the knowledge has to be gathered first, and I price it after reading your questions.
- net, one-off setup of the solar version
- PLN 3,900
- net a month
- PLN 690
How to start without wasting money
- Write down the ten questions from the first section, with the answers your best person would give.
- Do the calculation on your own numbers. If it comes out negative, stop here.
- Test a live assistant, mine or anyone’s, with a question it cannot know. Watch what it does.
- Ask the three data questions and have the answers sent to you by email.
- Start with the website only. Integrations with the CRM or the calendar can come once you have seen a month of the log.
The first month tells you more than any presentation. I keep my own assistant public for the same reason: a company deciding about one should be able to talk to one first. What else I build with AI, and where a person keeps the last word, is on the AI implementation page.
Questions I hear most often
Is a chatbot worth it for a small company?
It comes down to how often the same questions come back. A two person firm answering thirty identical questions a week gains more than a large one answering five. The calculation in this article takes five minutes.
Will the chatbot make things up?
A generic plugin often will. An assistant built on your company’s knowledge should answer only from it and say “I don’t know” when the question goes beyond it. Test that before buying by asking something it cannot know.
Do you only build chatbots for solar companies?
No. Asystent PV is my ready version for that sector and my public demo. For any other company I build the same kind of assistant on its own offer, prices, area and customer questions.
What does the launch date depend on?
On the questions your office answers. The solar version starts from sector knowledge that already exists, so it needs less preparation; another sector needs more groundwork. I give a date in writing once I have read your questions.
