An AI agent is a software program that receives a goal, collects the information it needs from your tools and data, decides on a sequence of actions, carries them out and checks the result, within limits a person has set. A chatbot answers a question and waits. An agent finishes a task. That difference is why agents are the part of AI that actually changes how a company runs.
This guide is written for business owners and managers who hear the term every week and want a plain answer. It covers the definition, how an agent works under the hood, the difference from chatbots and classic automation, six real examples, what an agent costs in 2026, how long it takes to build one, and when you should not build one at all. We build agents for companies across Europe from Prague, so the prices and timelines below are the ones we actually quote, not estimates.
AI agent: the short definition
An AI agent combines a large language model (the reasoning part) with three things the model does not have on its own: access to your tools (CRM, inbox, ERP, databases, calendars), a memory of the task it is working on, and a loop in which it plans, acts, observes the result and plans again until the goal is met or a limit is reached.
Five facts that hold for every agent worth the name:
- It is given a goal, not a script. "Answer every order-status email using our order system" is a goal. A fixed if-then flowchart is a script.
- It can call tools. Read a database, send an email, create a ticket, update a CRM record. Without tools it is a chatbot.
- It works in steps. Each step is a model call: decide, act, read the result, decide again. A typical customer-service task takes 3 to 8 steps.
- It has limits. What it may read, what it may change, when it must hand over to a person. These limits are the difference between a useful agent and a liability.
- It is measured. Every run is logged: what it read, what it did, how long it took, where it stopped. If you cannot see that, you cannot trust it.
AI agent vs chatbot vs classic automation
The three are often sold under the same label. They solve different problems and cost different amounts.
| Question | Chatbot | Classic automation (RPA, Zapier, n8n flow) | AI agent |
|---|---|---|---|
| What it does | Answers questions in text | Runs a fixed sequence when triggered | Reaches a goal by choosing its own steps |
| Handles unexpected input | Partly, in conversation only | No, it breaks or stops | Yes, within its limits |
| Touches your systems | Rarely | Yes, exactly as scripted | Yes, as needed for the goal |
| Typical cost to build (2026) | EUR 500 to 3,000 | EUR 1,000 to 5,000 per flow | From EUR 5,000 for one agent, one job |
| Best for | FAQ, first contact | Stable, repetitive data moves | Work that needs reading, judgement and action |
A useful test: if you could write the whole process as a flowchart with no "it depends" boxes, you want classic automation, which is cheaper. If the process needs someone to read, interpret and decide, you want an agent.
How an AI agent works, step by step
Every agent we build has the same five parts. Knowing them helps you judge a vendor's offer.
- Instructions. A written description of the job, the rules, the tone and the hard limits. This is the most important document in the project and it is written in plain language, not code.
- Model. The reasoning engine. In 2026 that is usually Claude, Gemini, GPT or an open-source model running on your own servers. The choice depends on the job, your data rules and cost per request; we test two or three on real data before deciding.
- Tools. Connections to the systems the agent needs: an API to your CRM, read access to your order database, an email account, a calendar. Each tool has its own permissions.
- Memory. What the agent remembers inside one task (the conversation so far, what it already checked) and, when needed, across tasks (a customer's history). Most agents need only the first kind.
- Logging and handover. Every run is recorded. When the agent is unsure, hits a limit or meets a case outside its instructions, it stops and hands the task to a person with a summary.
A run looks like this in practice. A customer writes: "Where is my order 48213?" The agent reads the email (step 1), looks the order up in the shop system (step 2), sees it shipped yesterday and finds the tracking number (step 3), drafts a reply in the customer's language (step 4) and sends it or, if you prefer, puts it in a draft folder for a person to approve (step 5). Elapsed time: 20 to 40 seconds. Model cost for that run: well under one cent.
Six AI agent examples from real companies
These are the jobs companies most often hand to an agent first, with the numbers that typically apply.
1. Order and delivery questions
Reads incoming emails and chats, checks the order system, answers status, delivery and return questions. Usually covers 60 to 80 percent of a shop's support volume. Hard cases go to a person with a summary.
2. Lead qualification
A new form submission arrives at 11 p.m. The agent replies within a minute, asks two or three qualifying questions, scores the lead in the CRM and books a meeting for good ones. Sales starts the day with warm leads only.
3. Invoice and document processing
Reads supplier invoices from a mailbox, extracts the fields, matches them against purchase orders, flags mismatches and posts the clean ones to accounting. One client went from four hours a day of manual entry to twenty minutes of review.
4. Internal knowledge assistant
Answers employee questions from your own documents (policies, product specs, past tickets), with a link to the source every time. Different from a public chatbot because it runs on your data and respects who may see what.
5. Reporting
Every Monday at 7 a.m. the agent pulls last week's numbers from three systems, writes a one-page summary with the three things that changed, and sends it to the management group. No dashboard that nobody opens.
6. Data entry between tools
Takes what arrives in one system (a signed contract, a booking, a support ticket) and creates the matching records in the others. The most boring job in the company and the one with the fastest return.
What an AI agent costs in 2026
Three cost lines, always. Vendors who quote only the first one are hiding the other two.
- Building the agent. At Buinsoft, one agent for one clearly defined job is a fixed price from EUR 5,000, excluding VAT. Larger scope, more systems or several agents are quoted after a free meeting, or billed at EUR 25 to 49 per hour depending on the role.
- Model usage. Paid to the model provider (Anthropic, Google, OpenAI) based on how much text the agent reads and writes. For an agent handling a few hundred tasks a day this is typically EUR 20 to 150 a month. We estimate it before the project starts.
- Running and maintenance. Hosting, monitoring, small changes when your systems or rules change. Plan for a few hours a month.
For a detailed breakdown with Czech koruna amounts, see our guide How much does an AI agent cost a business?
How long it takes
Our standard timeline for a single agent is about 30 days: a free 30-minute meeting to pick the job, a written quote within 3 working days, a free prototype in the first week that runs on your own data, weekly sprints you can test, and go-live with monitoring and a handover to your team. Projects that connect many systems or need new integrations take longer and are planned in phases.
When you should not build an AI agent
- The process is fully predictable. Use classic automation; it is cheaper and easier to audit.
- The data is not accessible. If the information lives in people's heads or on paper, fix that first. An agent cannot read what is not there.
- A mistake is expensive and cannot be reviewed. Agents are excellent at drafting and preparing, and should hand over to a person before an irreversible action (a payment, a legal commitment) until trust is established.
- Nobody will own it. An agent is a colleague that needs a manager: someone who reads the logs, updates the instructions and answers its questions. Without that person it decays.
Security and data: the questions to ask any vendor
Where is the data processed and stored (EU hosting or on-premise is possible with every major model)? Which systems can the agent change, and which can it only read? Is every action logged? What happens when it is unsure? Can we switch models later without rebuilding? If the vendor cannot answer these in one sentence each, keep looking. We cover this in depth in our AI agent security guide for small businesses.
Where to start
Pick one job that a person does on a screen every day, that follows rules most of the time, and where the information already sits in a system. That is your first agent. Buinsoft builds AI agents and AI integrations for companies across Europe from Prague, at a fixed price, live in about 30 days. The first meeting is free, and so is the one-week prototype. Write to info@buinsoft.com or book a meeting.
Last updated 20 October 2026. Prices exclude VAT and reflect Buinsoft's published price list.
