What an AI agent is
An AI agent is a program that uses a language model to carry out a task in your systems on its own: it reads a request, fetches the data it needs, decides on the next step and performs the action, but only within the permissions you give it.
The difference from a chatbot is that a chatbot answers questions in a conversation, while an agent does work. It logs a customer request in the CRM, prepares a draft quote, sends a report or schedules a callback. If all you need is answers to customers' questions, an AI chatbot for customer support is enough. For companies in Croatia, we build agents that work in Croatian and almost any other language and connect to the ERP and CRM systems you already use. Agents can also work in other languages, which we test on real questions before launch.
Examples of tasks for an AI agent
- Sorting enquiries: the agent reads incoming email and messages from the web form, recognises the topic and urgency, opens a ticket in the CRM and assigns it to the right person.
- Replies and quotes from the CRM: based on the customer's history, the price list and terms, the agent prepares a draft reply or quote, which a salesperson reviews and sends.
- Reports from the ERP: the agent gathers sales and stock data, puts together a report and answers questions asked in plain language.
- Phone calls: an AI voice agent answers inbound calls and makes outbound ones, and after the conversation it sends a quote or schedules a callback.
An AI agent makes sense for a task that repeats, has clear rules and relies on data that is already in your systems. If a simple rule or a scheduled script can handle the task, AI is not needed, and we will tell you so. Decisions with serious consequences stay with people.
Permissions, approvals and a record of every action
An agent that is allowed to act in your systems must have clear limits. Every agent we build has these safeguards:
- Its own access to your systems, with the minimum permissions needed. Wherever possible, the agent only reads, and data it does not need is not visible to it at all.
- An approval step for actions with consequences. An email to a customer, a price change or an order waits for a person's confirmation at first, and once the results prove reliable, you decide what the agent may do on its own.
- A record of every action: what the agent received, which data it read, which tool it called and what it did. Every decision can be checked afterwards, and the agent can be switched off at any time.
We describe rules, roles and oversight of AI systems in more detail on our AI security and governance page.
The transparency obligation in the EU AI Act
Article 50 of Regulation (EU) 2024/1689, known as the Artificial Intelligence Act or EU AI Act, has applied since 2 August 2026. When an AI system interacts directly with people, they must be informed that they are interacting with an AI system, at the latest at the first interaction, unless this is obvious from the circumstances. That is why an agent that writes to customers says so in its message, and a voice agent says so in its greeting at the start of the call. What else the Act requires of your company is described on our EU AI Act compliance page.
Our experience: voice agents and AI on ERP data
We use AI in our own products, which run in production.
CallVista is our platform for voice agents in call centres. The agent takes inbound and makes outbound calls in Croatian and other languages, answers from a knowledge base built from PDF, Word and CSV documents, writes a summary with a sentiment score and tags after each call, sends a quote by email and schedules callbacks itself. The client profile is built up across all calls.
Business Booster is our business platform for communication, tasks, analytics and AI on ERP data. For Instar informatika, a Croatian retailer and distributor of IT equipment, we connected it to their ERP, Luceed (Tomsoft). Every night, a read-only copy of the ERP data is synchronised. The system answers questions asked in plain Croatian from the ERP data, with a short AI explanation: for example, which products sell best, which stores have the highest turnover, which goods have been sitting in the warehouse for a long time, or what is low in stock but still selling.
Every week, the same system emails sales and stock reports by brand, in each supplier's own Excel format. These reports are not written by AI: rules applied to the data are enough for them. Read more about the project in our references.
How we introduce an AI agent, step by step
- A free first call: we choose one task that repeats and has a clear result.
- A review of systems and data: the CRM, ERP, email, documents and APIs the agent works with.
- Scope, cost estimate and a phased plan, with a list of permissions and approvals.
- A first working version in 4–8 weeks, tested on real examples from your business.
- Work with human approval, until the results show where the agent can work on its own.
- Maintenance, monitoring and new tasks, with progress tracked in the client portal.
We build agents on Node.js and PostgreSQL, and use pgvector to search documents and data by meaning. We host the applications we build and maintain on servers in the EU. How AI connects to business systems is explained on our AI integration with CRM and ERP systems page.
What the price of an AI agent depends on
- the number of systems the agent works with, and whether they have an API;
- the number of tasks and exceptions the agent has to recognise;
- how many actions need approval, and who approves them;
- the volume of enquiries, calls or documents, which determines the cost of using the language model;
- the quality and availability of the data;
- requirements for security, logging and data retention.
How the estimate is put together is explained on our AI implementation cost page.
What you need to prepare
- a description of the task as you do it today, step by step;
- real examples of enquiries, quotes or reports, without personal data if necessary;
- access to the systems or the documentation of their APIs;
- a decision on who approves the agent's actions and what the agent must never do.