What is an AI agent
An AI agent is a program built on a language model (such as the models behind ChatGPT) that does not stop at a text answer. It is given a goal, works out the next steps itself and uses tools: it checks data in your company systems, prepares documents and sends them for approval. After each step it assesses the result and decides what to do next.
Put simply: a chatbot talks, an agent acts.
AI agent vs chatbot
| Chatbot | AI agent | |
|---|---|---|
| What it does | Answers questions | Carries out tasks in several steps |
| Access to data | Usually a knowledge base or FAQ | Data from your shop, ERP, CRM or mailbox — within its permissions |
| Result | An answer for the user | A ready draft of an offer, order, report or reply |
| Control | You check the content of answers | You check permissions, steps and decisions |
One does not rule out the other: a chatbot on your website can pass tasks to an agent that completes them in the background.
Examples of use in a company
- Enquiries and orders. A customer asks about order status, availability or delivery date. The agent checks the data in your shop and ERP and prepares a reply.
- Offers and documents. Using ERP data, the agent drafts an offer, an order or a reply for an employee to approve.
- Orders from e-mails. The agent reads an order from the message body or an attachment, matches products to your catalogue and creates a document in the ERP — unclear cases go to a person.
- Shop data analysis. Ask in plain language, e.g. “which products will run out this week?”, and the agent gathers data from the shop and warehouse.
- Reports on demand. Sales, margin or stock summaries in seconds, without manually combining spreadsheets.
- Shop content. Product descriptions, translations and missing attributes based on ERP data — approved before publication.
How an agent connects to your systems
An agent is only as useful as the data it can access. We connect it to your systems through the shop and ERP APIs or through the MCP protocol (Model Context Protocol) — a standard that describes which tools and data a model may use. You can read more on our AI and automation page.
The key rule: an agent gets only the permissions it needs. An agent that prepares offers can read price lists and product records, but it does not change prices or issue invoices.
Security and control
- A person approves business decisions — the agent prepares a proposal, an employee accepts it.
- Limited permissions — a separate account for the agent, with access only to the data and operations it needs.
- Activity log — every step the agent takes is recorded and can be reviewed.
- Data — we agree where it is processed and what information reaches the model at all.
- Cost limits — limits on requests and budget, so model costs stay predictable.
How to implement an AI agent — stages
- Choosing the process. We start with one process that brings real savings: repetitive, frequent and with a clear result.
- Analysing data and systems. Where the agent gets its data from, what it may write, who approves the result.
- A prototype on real examples. We test on historical data before the agent starts working live.
- Rollout with human oversight. At first an employee approves every result; the agent’s autonomy grows over time.
- Measurement and development. We measure the effect and only then extend the agent to further tasks.
How much does an AI agent cost
The cost has two parts:
- one-off implementation — analysis, integration with your systems, prototype and testing; it depends mainly on the number of systems and the complexity of the process,
- running costs — fees for the AI model (depending on the number and size of tasks) and maintenance.
That is why we estimate the volume of tasks and set limits before we start.
How to measure the effect
- time to complete a task before and after implementation,
- share of tasks completed without corrections,
- number of errors and corrections,
- team time saved per week.
Where to start
Pick one process that takes up the most time today — for example answering questions about orders or retyping orders from e-mails. Describe it to us and we will suggest a first step and prepare a quote.
FAQ
Will an AI agent replace employees?
That is not the point. The agent takes over repetitive steps — looking up data, retyping, preparing drafts — while decisions stay with people.
Can an AI agent work with data from our ERP?
Yes. We connect it to your ERP, shop or application through an API or the MCP protocol and give it access only to the information and operations it needs.
Do we need our own AI model?
Usually not. We use existing language models; the advantage comes from connecting them to your data and processes.
At what company size does it make sense?
As soon as a task is repeated dozens or hundreds of times a month and takes up real team time.