In recent months, we have witnessed models like GPT-4o or Claude 3.5 Sonnet exceeding our expectations in creative writing and programming. However, as expert Andrew Ng points out, the real breakthrough comes with the implementation of agentic workflows. The difference between a regular chatbot and agentic AI is similar to the difference between ordering food through an app (chatbot) and hiring a chef who buys ingredients, prepares the recipe, and serves the meal himself (agent).
What exactly is Agentic AI?
Traditional LLMs (Large Language Models) work on the principle of predicting the next word. They are excellent at generating text but lack the ability to "act" in the real world. Agentic AI is a system that uses these models as its "brain" but adds three key capabilities:
- Reasoning: The ability to break down a complex task into smaller, logical steps.
- Tool Use: The ability to interact with external software – a browser, calendar, database, or accounting system.
- Iteration and feedback: The agent can evaluate the outcome of its work. If it finds an error, it tries a different path without needing human instruction.
According to the definition provided in expert analyses by Grammarly, these systems range from simple automation to autonomous decision-making within defined parameters.
Comparison: Chatbot vs. AI Agent
To understand the scope of this change, let's look at a comparison of current top models and their capabilities in the context of agentic work:
| Feature | Standard Chatbot (e.g., GPT-4) | Agentic System (e.g., AutoGPT / CrewAI / Claude Computer Use) |
|---|---|---|
| Interaction | Reactive (waits for prompt) | Proactive (works towards a goal) |
| Workflow | One-time response | Iterative cycle (planning -> action -> review) |
| Tools | Text/data only in context | Direct access to API, web, and files |
In benchmarks, the evaluation is no longer just about how well a model answers questions (MMLU), but how successfully it can solve a task in a real environment, such as automatically processing a complaint or managing an e-shop. In this regard, systems using Claude 3.5 Sonnet excel due to its ability to "see" and control a computer much like a human.
Practical Impact: What does this mean for you?
This shift has profound implications for various segments of society. Let's look at specific examples:
1. For small businesses and entrepreneurs in the Czech Republic
Instead of paying expensive agencies for simple administrative tasks, you can use agentic tools. Imagine an agent you instruct: "Monitor competitor prices for our three main products, and if they drop below X CZK, adjust our prices in Shoptet and send me a report." The agent will do it themselves. For the Czech market, the availability of these tools via API is crucial. Most top models (OpenAI, Anthropic) are available to developers in the Czech Republic without restrictions, allowing local companies to build their own solutions.
2. For retail and e-commerce
As trends in Retail Technology Innovation Hub suggest, AI is integrating directly into logistics and customer services. Agentic AI can solve delivery problems in real-time or optimize inventory based on demand forecasts, leading to significant cost reductions.
3. For employees and the job market
The change is not that AI will replace humans, but that the human role will shift from "executor" to "supervisor." You will need to be able to define goals, monitor the ethics of agent decisions, and handle situations where an agent encounters limits. This requires new digital skills.
Pricing and Availability
If you want to start experimenting with agentic AI, here's an overview of costs from the biggest players:
- ChatGPT (OpenAI): The free tier is limited. For full potential (GPT-4o), a subscription costs 20 USD/month (approx. 460 CZK). For developing agents, it's necessary to use the API, where you pay for tokens (consumption depends on the complexity of the task).
- Claude (Anthropic): Similar to OpenAI, Claude Pro offers 20 USD/month. Their Claude 3.5 model is currently considered top-tier for agentic tasks due to its excellent logic.
- Gemini (Google): Integrated within Google Workspace. Prices vary by business plan, but basic access to Gemini 1.5 Pro models is available via Google AI Studio.
Regulation and Security: EU AI Act
For European users and businesses, it is essential to mention the EU AI Act. Agentic systems that have the ability for autonomous decision-making may be classified as high-risk systems, especially if they affect the rights of citizens (e.g., in HR or banking). This means that companies in the Czech Republic must ensure a high degree of transparency and human oversight (human-in-the-loop) to meet European standards for safety and ethics.
Are agentic systems safe when they can use tools themselves?
Safety depends on the "sandbox" (isolated environment) setup and permission limitations. Professional implementations always include limits that prevent the agent from, for example, making unlimited payments or deleting critical data without human confirmation.
Can agentic AI work in Czech?
Yes. Since agentic systems are built on models like GPT or Claude, which have excellent knowledge of Czech, they can perform tasks (writing emails, analyzing Czech documents) in our language without problems.
What is the difference between automation and agentic AI?
Classic automation (e.g., Zapier) does exactly what you tell it: "If an email arrives, save the attachment." Agentic AI asks: "This email contains an invoice; I need to check it against the order in the system, and if it doesn't match, I'll write an inquiry to the supplier."
