Artificial intelligence in customer service: AI chatbots that help rather than drive customers away

What is an AI chatbot and how does it differ from the old “button bot”?

AI chatbots are programs that understand a question asked in the customer's own words and reply in full sentences rather than with a choice from a menu. Old bots worked like a decision tree. The customer clicked “Order status”, then “Enter number”. An intelligent chatbot based on a language model (like those in ChatGPT or Gemini) can handle the question “I ordered shoes on Tuesday but only socks arrived – what now?”.

In English-language material you’ll come across the terms customer service chatbot, conversational AI chatbot if AI agent chatbot. The differences between the types are easiest to show in a table. Customer service is just one of the areas in which companies are turning to AI. The others are described in our guide to using AI in business.

How do AI chatbots built on company knowledge work?

A chatbot trained on company data (the search term is AI chatbot trained on my data) is usually not “trained” in the literal sense. It works in a set-up called RAG. First it searches the company's documents for passages matching the question, and only then does the language model compose an answer from them. We describe how this looks from the inside in the article on RAG, or AI built on company knowledge.

Hence the most important rule: a chatbot is only as good as the documents you give it. If the returns policy has three versions in three places, the bot will pick one of them. Not necessarily the current one. And if there is no answer at all, the model may “fill in” one. This phenomenon is called hallucination, and we have devoted a separate article to it on AI hallucinations.

Who is an AI chatbot for, and when does it pay off?

An AI chatbot for business makes sense where the same questions come up dozens of times a day. Think of an online shop that is flooded with delivery questions after every promotion, a service company whose customers keep asking about repair status, a recruitment team that endlessly explains the stages of the process to candidates, or a B2B company with an extensive range where customers are looking for the right product.

It does not pay off when there are few questions and each one is different. Nor when the conversation is sensitive: complaints from major customers, health matters, a customer's financial decisions. In such places AI can help the agent (suggest replies, summarise the history), but it should not handle the conversation alone.

Keep customer sentiment in mind. According to a 2024 Gartner survey, 64% of customers would prefer companies not to use AI in customer service, and their main concern is the difficulty of reaching a human. Gemius, in its “E-commerce in Poland 2025” report, also notes that a chatbot is not the preferred form of contact for Polish shoppers. The conclusion? Not “don't implement”. Rather, “always provide a way out to a human”.

Chatbots in Polish companies: examples

Here are the typical places where AI chatbots in customer service work best. The examples are illustrative. They show a pattern, not the story of a specific company.

• Website chatbot. It answers questions about the offer and prices, gives dates, collects contact details and passes the enquiry to a sales rep. On WordPress-based sites it can be added with a plugin or with code from the provider.
• AI chatbot for an online shop. It helps choose a size or model, checks order status and explains the returns policy. We write more about online retail in the article on AI in e-commerce.
• Chatbot in Messenger, Instagram and WhatsApp. It takes over questions from direct messages that would otherwise wait for an answer until Monday.
• B2B sales chatbot. It asks a few qualifying questions and passes a ready-made CRM note to the sales rep. We show how to connect this with sales when discussing CRM with AI.
• Recruitment chatbot and recruitment voicebot. It tells candidates what the working hours are and how to get there, and explains the stages. We explain what you must not do in AI-assisted recruitment in the article on AI in HR and recruitment.
• Helpline voicebot. It handles simple calls, such as confirming an appointment or checking a delivery status, and transfers the rest to an agent.

Implementing an AI chatbot in your company, step by step

The question “how do I create a chatbot?” has a simpler answer today than a few years ago, because many platforms let you launch a bot without a developer. The harder part is the one no platform will do for you. If you are thinking of a more autonomous solution that not only answers but also takes action in your systems, read our guide on how to create an AI agent.

1. STEP 01Collect the most frequent questions

   Go through emails, messages and helpline notes from the last three months. List the 20–30 questions that come up most often. These decide whether a chatbot makes sense.

2. STEP 02Organise the answers

   One current answer to each question, in one place. Remove old versions of terms and price lists, because the bot may come across them.

3. STEP 03Set boundaries and a way through to a person

   List the topics the bot does not handle (complaints above a certain amount, legal matters, grievances). Every conversation has a visible “talk to an agent” button, and the agent receives the conversation history so that the customer does not have to repeat everything from the beginning.

4. STEP 04Test with a small group

   Launch the bot on one page or in one channel. For two weeks, read the conversations and improve the knowledge base.

5. STEP 05Measure and expand

   Track the share of cases resolved without a human, response time, the customer's rating of the conversation and the number of handovers. Only with these figures should you decide on further channels.

How much does an AI chatbot cost for a business?

Platforms differ in their billing models. Some charge per agent seat, others for each case resolved by AI, and others still by the number of conversations, which with seasonal traffic can multiply the monthly bill several times over before anyone in the company has a chance to notice and react. Below are the items that make up the budget, with an example from a public price list.

Law and liability: what a company with a chatbot needs to know

Since 2 August 2026, Article 50 of the AI Act has applied: a person talking to an AI system must know that they are not talking to a human, unless this is obvious. The simplest way to deal with this is a notice in the bot's first message. We write more about the rules in the article on The AI Act in Poland.

The second issue is liability for content. In February 2024, a Canadian tribunal ruled in Moffatt v. Air Canada that the airline was liable for incorrect information about a discount given by its chatbot. The company argued that the chatbot was a “separate legal entity”. The tribunal did not accept this. It is a ruling from outside the EU, but the logic is universal, because customers do not distinguish whether a promise was made by a person or a program.

Which AI chatbot should you choose? Criteria instead of a ranking

“Best AI chatbot” round-ups usually compare general assistants (ChatGPT, Gemini, Claude), not customer service tools. For a business, other things matter. Instead of looking for the winner of a ranking, check six things:

1. whether the bot answers only on the basis of your documents and admits when it doesn’t know,
2. whether handover to an agent works smoothly and includes the conversation history,
3. whether it supports the channels your customers use (website, Messenger, Instagram, WhatsApp, phone),
4. whether it handles Polish well, including inflection and typos,
5. where it stores data and whether it uses it to train models,
6. how charges are calculated when traffic grows, e.g. in peak season.

When to implement AI in customer service, and when to hold off

The best moment is two or three months before the peak season. There is then time for testing and fixes, and the effect will arrive just when the team is under the most pressure. Launching a bot on the day a big promotion starts is asking for trouble.

Hold off if the company has no written service rules and each agent answers differently, if the terms and conditions are being changed, or if the brand is going through a difficult period for its reputation. In a crisis, an automated system that answers cheerfully and off-topic can only add fuel to the fire. We write about this in the article on AI in crisis communication. We have collected good principles for talking to customers, whatever the channel, in our guide on communication with customers.

How do you measure whether a chatbot is helping?

The most common trap? A company boasts that its bot has “handled” thousands of conversations. But so what, if half of the customers called the helpline after the conversation anyway. Count cases resolved, not conversations held. Add to that the customer's rating of the conversation, the rate of handovers to a human and the number of repeat contacts about the same issue.

Fewer emails to customer service? Check whether that is thanks to the bot or, for example, to lower traffic in the shop. The cleanest way is to switch the bot on for only part of the site and compare both groups in the same weeks. If you need support in designing such an implementation, visit the page Commplace® AI agency.

Mistakes that turn a chatbot into a machine for annoying customers

Read next