Customer support automation: AI chatbot, AI agent and voice agent

AI automation of customer support means that an AI chatbot, AI agent or voice agent handles the questions customers ask most often, while people take over everything that calls for judgement. It makes sense when the same questions keep coming up and there are accurate, written answers to them; to introduce it, you need a well-kept knowledge base, clear rules for handing over to a live agent, and compliance with the GDPR and the EU AI Act.

What is the difference between a chatbot, an AI agent and a voice agent?

They differ in what the system understands and what it is allowed to do.

• A scripted chatbot leads the conversation along branches and buttons drawn up in advance, and does not understand a question nobody anticipated.
• A knowledge-base AI chatbot understands questions asked in the customer’s own words and answers from your documents, but changes nothing in your systems.
• An AI agent acts as well as answers: it checks an order, opens a ticket or changes an appointment, within the permissions you give it.
• A voice agent is an AI agent that does the same job over the phone: it listens, answers by voice and, if it is connected to your systems, takes action.

Example: on a Sunday evening, a customer writes to an air-conditioning service company to say that their unit is leaking. A scripted chatbot will offer the buttons “Service” and “Contact”. An AI chatbot will explain what the customer can check themselves, following your instructions. An AI agent will also check the warranty by serial number, offer an available slot and open a work order. A voice agent will do the same over the phone.

The more a system is allowed to do, the more a mistake costs: at worst, a scripted chatbot disappoints a customer, an AI chatbot can get an answer wrong, and an AI agent can get an action wrong.

What to automate first, and what to leave to people

Start with enquiries that are frequent, have one correct answer and do not call for an exception to the rules: opening hours, delivery, returns, order status, instructions for use and booking appointments. A website AI chatbot is enough for this. Another good starting point is collecting details, such as the order number and a description of the problem, before the enquiry reaches a person.

Leave these to people: complaints in which the customer disputes a decision or asks for compensation, exceptions to the rules and refunds outside the usual terms, upset customers, sensitive topics such as health or debt, and negotiations with large customers.

For some of these decisions there is also a legal reason. Decisions that affect a person legally or in a similarly significant way, such as terminating a contract, may be based solely on automated processing only in the exceptions set out in Article 22 of the General Data Protection Regulation. The Croatian data protection authority, AZOP, stresses in its recommendation on automated decision-making (in Croatian) that human involvement in such a process must be real, not merely formal.

What a good handover to a live agent looks like

Handover to a live agent is the moment a person takes over the conversation. More than anything else, it decides whether customers will trust the system. It is done well when:

• the option to talk to a person is always visible, not hidden behind several failed attempts,
• the system hands over the conversation by itself when the customer asks for it, when it fails to understand them twice, when it has no source in the knowledge base for an answer, or when the topic is on the “people only” list,
• the person receives the full transcript, a short summary and the details already collected, so the customer does not have to repeat anything,
• outside working hours, the system opens a ticket and says honestly when a person will get back to the customer.

How to check that a chatbot works well in Croatian

Modern language models understand Croatian, but that does not yet mean they will give your customers accurate answers. Test the system on real questions from email and chat, with personal data removed, and check in particular:

• messages typed without diacritics (“mozete li poslati racun” for “možete li poslati račun”, “can you send the invoice”) and typos,
• case endings in the names of products, places and streets,
• regional expressions and words borrowed from English,
• consistent forms of address: decide whether the system addresses customers with the informal “ti” or the formal “Vi”, and stick to it,
• grammatical gender: the question “jesi li bio zadovoljan” (“were you satisfied”, in the masculine form) assumes the other person’s gender, so phrase it differently.

Have a native speaker from your team review a sample of conversations before launch and regularly afterwards. With a voice agent, also check how it pronounces names, numbers and addresses, how it copes with background noise and how it handles callers who interrupt.

How to prepare the knowledge base

A knowledge base is the set of documents the AI takes its answers from, so the answers cannot be better than the knowledge base itself.

• Gather the sources: FAQs, general terms and conditions, instructions, delivery, returns and warranty policies, and good answers from email.
• Keep one current document per topic, because contradictory and outdated documents lead straight to wrong answers.
• Write short sections with a clear heading and a single topic, and keep product data in a table.
• Give every document an owner and a date of last review.
• Remove personal data that is not needed for an answer, as AZOP advises in its recommendation on the safe use of large language models (in Croatian).

The best source of new answers is the conversations handed over to people, because each one shows what is missing from the knowledge base.

Connecting to your CRM and ticketing system

An AI chatbot that only answers questions does not need access to your systems. As soon as it has to say where an order is or open a ticket, it connects to your CRM, ERP or customer ticketing system (helpdesk). Three rules then apply:

• read-only access first, and writing data only once reading works reliably,
• an identity check before order or invoice details are shown, for example with the order number and email address,
• every action the agent takes is logged, and actions that create an obligation, such as a refund or a cancellation, stay with people or need their confirmation.

Every conversation should end up as a record in the ticketing system or CRM, with a summary and tags, so the team can see the customer’s whole history.

When a voice agent makes sense

A voice agent makes sense when your customers mostly phone, when calls pile up at certain times of day, or when you need to make many similar calls to customers. We work on this ourselves: CallVista, our voice agent platform, handles inbound and outbound calls in Croatian and other languages and answers from a knowledge base loaded from PDF, Word and CSV files. After each call it records a summary, the caller’s sentiment and tags; it can send an offer by email and schedule a callback; and it carries the client profile over from call to call, so the next conversation picks up from what has already been said.

The phone also brings additional rules. If calls are recorded, the caller must be told during the call about the purpose of the recording, who receives the recordings, and their right to object and to access the recordings, as the European Data Protection Board’s data protection guide for small business explains. The simplest way is to say so right in the greeting. For outbound calls, Article 50 of the Croatian Electronic Communications Act (in Croatian) also applies: automated calling systems without human intervention may be used for direct marketing and sales only with prior consent. A callback the customer asked for is a different situation, but record when and how they asked for it.

What to measure

Before launch, record your baseline values, because without them you will not know whether anything has changed; afterwards, track the same indicators for the same types of enquiry:

• the share of conversations resolved without a person and the share handed over to a live agent,
• time to first response and time to resolution,
• first-contact resolution and repeat enquiries from the same customer on the same topic,
• the satisfaction rating after the conversation,
• answer accuracy on a sample that a person reviews by hand every week,
• questions the knowledge base has no answer to,
• for the phone, also dropped calls and completed callbacks.

A high share of conversations handled without a person is not a success if the customer sends the same question by email the next day. That is why you should look at the indicators together.

GDPR and conversation transcripts

A transcript of a chat or call contains personal data as soon as a person can be identified from it, whether by name, email address, order number or simply the description of a complaint. The General Data Protection Regulation (GDPR) applies to it, which in practice means:

• state the purpose and legal basis of the processing, and the retention period, in your privacy notice, and link to it from the chat window (Articles 6 and 13),
• keep transcripts only as long as they are needed, then delete or anonymise them (Article 5(1)(e)),
• do not ask for more data than you need for the enquiry (Article 5(1)(c)),
• sign a data processing agreement with the supplier of the chatbot and of the language model (Article 28), and check where the data is processed, how long it is kept and whether it is used to train models, as AZOP also recommends,
• transfers of data outside the European Economic Area are subject to the specific conditions in Chapter V.

AZOP also recommends assessing, before you introduce such a tool, whether you need a data protection impact assessment.

What the EU AI Act requires from 2 August 2026

Since 2 August 2026, the transparency obligation in Article 50 of Regulation (EU) 2024/1689, better known as the AI Act, has applied. AI systems intended to interact directly with people must be designed so that the person finds out they are dealing with an AI system, unless this is obvious to a reasonably well-informed, observant and circumspect person. The notice must be clear and distinguishable, given at the latest at the time of the first interaction, and in line with accessibility requirements.

In practice it looks like this: “Hello, I’m a virtual assistant and I answer with the help of artificial intelligence. If you’d like to talk to someone from the team, just say so.” A voice agent says the same in its greeting. Do not rely on it being “obvious”: with a human name and a natural voice, it is anything but.

Formally, the obligation falls on the provider, and a company that had the system developed and then put it into service under its own name is also a provider (Article 3(3)). So write the notice into the specification, whoever builds the system. The Regulation provides for administrative fines for breaches of the obligations in Article 50 (Article 99).

Providers of systems that generate synthetic audio or text must also ensure that the output is marked in a machine-readable format (Article 50(2)). For systems placed on the market before 2 August 2026, the Digital Omnibus on AI set a deadline of 2 December 2026 for this. A customer support chatbot is not, as a rule, a high-risk system under Annex III, which covers areas such as employment and creditworthiness assessment.

An eight-step rollout plan

1. Pull out real enquiries from email, chat and phone over a period that includes both quiet months and your busiest ones, and sort them by topic.
2. Choose two or three topics that are frequent, simple and have a clear answer, and write down what the system must not do.
3. Prepare the knowledge base for those topics, with an owner and a review date for every document.
4. Set the rules for handover to a live agent, and decide who takes over the conversation during working hours and who outside them.
5. Connect only the systems you need, read-only at first.
6. Sort out the legal side: the notice that customers are dealing with AI, the privacy notice, the data processing agreement and the retention period for transcripts.
7. Test internally on anonymised customer questions, and have employees deliberately try to confuse the system.
8. Start a trial run on one channel, review the indicators and conversations every week, fill the gaps in the knowledge base, and only then add new topics, actions in your systems or the phone.

In our projects, the first working version is usually ready in 4–8 weeks, depending on the number of topics and systems to be connected.

If you need an AI chatbot for your website and customer support, the first conversation with us is free. In it we go through your most frequent enquiries and channels, the systems you use and the rules that apply to you, and then suggest where to start, what to leave to people and what a phased plan with a cost estimate would look like. The applications we build and maintain are hosted on servers in the EU.