AI agents
AI agent for business
An agent is not a chatbot with better answers. It is a system that reads the input, decides, and then acts in your tools — records, assigns, replies, escalates.
An AI agent is automation that makes its own decisions within a bounded task. It receives input in natural language — an email, a message, a document — understands it, pulls out what matters, and triggers concrete steps in your systems. A chatbot answers. An agent acts.
In practice: an enquiry arrives, the agent works out what it is about, records it in the CRM, sends a reply that makes sense, and hands anything outside its scope to a person. Deployment starts at €700 depending on the number of steps and integrations. An example from practice: the agent for the insurance site gopoistenie.sk picks a topic from data every week, checks the facts against official sources and prepares the article as a draft — a person decides whether it is published.
Agent, chatbot or plain automation?
This is the most common misunderstanding in enquiries, and it usually decides both the price and whether the solution will work at all:
| Automation | Chatbot | AI agent | |
|---|---|---|---|
| What it does | runs a predefined procedure | answers questions | decides and takes action |
| Input | structured (form, API) | text from a person | free text or a document |
| Decision-making | fixed rules | picks an answer | evaluates the situation |
| When it is the right choice | the procedure is always the same | repeated customer questions | input differs every time but the response follows a pattern |
| Risk | low, behaviour is predictable | low | higher — the model can be wrong, so checkpoints are needed |
If your process is always the same, you do not need an agent — it would be the more expensive and less reliable option.
What an agent handles in practice
Processing enquiries
Reads an email or message, determines the type, asks for missing details and creates the record.
Qualifying leads
Judges whether an enquiry matches what you sell and assigns priority by your criteria.
Working with documents
Extracts data from an invoice, order or contract and writes it where it belongs.
Sorting and routing
Splits incoming messages by topic and sends them to the right person or queue.
Analysing feedback
Goes through hundreds of comments or reviews and returns an overview of themes and sentiment.
Preparing material
Summarises a call, an email thread or a document into something a person reads in a minute.
An agent does not have to be visible. Some of this runs entirely in the background — neither the customer nor the employee knows it is there, they only see that the thing is handled.
Where the human stays in the loop
I consider this the most important part of the design, and I deal with it at the start rather than after something goes wrong. A language model can be wrong — the deployment has to assume it.
- Scope boundaries. The agent is explicitly told what it may and may not do. Outside that it does not improvise, it escalates.
- An approval point. For steps with consequences — a quote, a commitment, sending money, contact with an authority — a person confirms.
- A visible trail. Every decision the agent makes is recorded along with what it was based on. It can be checked afterwards.
- Safe failure. When the agent is unsure it does not guess. It flags the case and hands it over.
The aim is not to push the human out of the process, but to take away the part that is mechanical and repeated.
How a deployment runs
- Consultation. We go through what input arrives and how it is handled today. This is where it becomes clear whether it is work for an agent or whether simpler automation is enough.
- Defining the scope. We set out what the agent may do, what it escalates and where a person approves. This is the most important document of the whole project.
- Prototype. I build a first version and we test it on your real inputs — not on invented examples.
- Tuning. We go through the cases where the agent decided differently from what you wanted and adjust the instructions and boundaries.
- Deployment with oversight. The first period runs under closer supervision. The scope widens only once it holds.
When I advise against an agent
When the procedure is always the same. Plain automation is then cheaper, faster and more reliable. An agent would add cost and unpredictability for no gain.
When the result must be perfect and unchecked. Legal filings, tax returns, medical or financial decisions — there an agent can at most prepare the material.
When there is nothing to build on. If no record exists of how cases were handled so far, the agent has nothing to work from and the first weeks will be expensive.
When the volume is small. A handful of cases a week will not pay for itself — design, tuning and running costs exceed the saving.
What drives the price
- Number of input types. One channel and one kind of enquiry is different work from five channels and ten kinds.
- Number of integrations. Every system the agent writes into has to be connected and tested.
- Difficulty of the decision. Simple classification versus multi-step reasoning over a document.
- Level of control. Approval steps and an audit trail add work — and they are worth it.
- €700 – 1,200An agent for one bounded scenario — one channel, clear boundaries, connected to a calendar or spreadsheets.
- from €1,200An agent with several integrations, approval steps and work over documents.
On top of the deployment, budget for running costs — mainly the language-model API, which depends on the volume of text processed. You get an estimate in the proposal and it can be monitored as you go.
Try it yourself
Instead of figures you have no way to check, there are three working demo bots on Telegram. One click, no sign-up, and you can see how the solution actually behaves:
- AI Store Assistant — an assistant for an online shop: products, availability, customer questions
- LeadCRM Demo — receiving and qualifying an enquiry
- CoursePay Demo — selling and granting access to an online course
They are demos, not client deployments — built so you can see the real behaviour before deciding anything.
Frequently asked questions
What is the difference between an AI agent and a chatbot?
A chatbot answers questions. An agent takes action in your systems based on the input — creates a record, assigns responsibility, sends a reply, escalates a case. A chatbot is an interface; an agent is a worker.
How much does deploying an AI agent cost?
An agent for one bounded scenario falls in the €700 – 1,200 range. A solution with several integrations and approval steps starts at €1,200. On top of that, budget for language-model API costs by volume.
Can the agent be wrong?
Yes, which is why the deployment is designed so that nothing depends on it being right. The agent has defined boundaries, a person confirms steps with consequences, and every decision is recorded. When the agent is unsure it hands the case over rather than guessing.
Where does our data go?
It depends on the model. When an external language model is called (Claude, GPT), the processed text is sent to the provider. I tell you in advance what that data is and what can be excluded or anonymised. If that is unacceptable, part of the solution can be built so sensitive data never leaves your infrastructure.
What languages does the agent work in?
Current language models handle English, Slovak and Ukrainian. Quality in a specific language is always tested on your real inputs rather than sample sentences — differences between models show up in specialist terminology.
How long does deployment take?
A prototype for a simple scenario is usually within two weeks, followed by tuning on real data. More complex solutions take longer and you get the date in the proposal up front.
Do we need our own IT person?
No. The system is handed over with written documentation and the credentials are yours. If a change is needed it can be handled by me or by anyone familiar with the tool.
What if the agent does not work the way we expected?
That is exactly why a prototype is built first and tested on your inputs — so it shows before there is investment in a full deployment. If testing shows your inputs are not suited to an agent, I say so and propose something simpler.
Related articles
Related work
Related services
Describe one case you handle by hand
One typical enquiry or document is enough. From it I can tell you whether an agent can handle it, or whether simpler and cheaper automation would do.
Get a free consultation →