The end of the era of "talking" models: What exactly is agentic AI?
Until now, we have been accustomed to interacting with AI as an advanced search engine or text author. You type a query, the model (e.g., GPT-4o or Claude 3.5 Sonnet) processes the context and generates a response. This is conversational AI. Agentic AI, however, goes further. An agent is not just a model; it is a system that has goals, can reason about procedures, and – most importantly – can act.
If you tell an agent: "Plan a business trip to Munich for me and book a hotel within my budget," a conversational model will only advise you on good hotels. An agentic system, however, will try to open your calendar, find out your preferences, search for flights via API, compare prices, and then make the reservation. However, there is a fundamental obstacle to this shift: AI models are "blind" and "handless" if they do not have a semantic understanding of data and access to external systems.
Semantic layer: The brain that understands context
According to analyses provided by Salesforce, traditional IT architecture, built on seven rigid layers, is not suitable for the dynamic nature of AI agents. The main problem is the lack of semantic coherence. Most companies have data scattered in isolated silos – CRM in one system, invoices in another, communication in Slack. For AI, this data is just a cluster of characters.
This is where the semantic layer comes in. It is not just another database, but an intelligent interpreter. This layer provides a unified understanding of information across the entire company. It allows the agent to understand that the term "customer" in the sales department and "contact person" in technical support mean the same thing. Without this layer, agents will not be able to effectively reason over disparate datasets and will be limited to narrow, one-off tasks.
Compared to standard RAG (Retrieval-Augmented Generation), which merely searches for relevant documents to supplement context, the semantic layer allows agents to perform true inference and understand business context. This is the difference between AI "finding text about a contract" and AI "understanding that this contract expires in three months and must initiate renewal."
API: The nervous system for action
If the semantic layer is the "brain," then APIs (Application Programming Interfaces) are the "hands" of agentic AI. As CIO.com states, APIs are the "glue" that allows agents to orchestrate processes across systems. Without them, AI remains confined to a closed loop of text responses.
Today's leading models, such as Google's Gemini 1.5 Pro or OpenAI's GPT-4o, already feature advanced mechanisms for so-called function calling. This means that the model can decide on its own that to resolve your request, it needs to call a specific API endpoint (e.g., "get_weather" or "create_jira_ticket").
Comparison of agentic capabilities:
- GPT-4o: Top in logical planning and function calling, very robust tool ecosystem.
- Claude 3.5 Sonnet: Extremely strong in nuanced understanding of instructions and coding, which is crucial for agents when manipulating data.
- Gemini 1.5 Pro: A huge context window allows agents to work with a massive amount of documentation at once, ideal for complex semantic tasks.
Practical impact: What does this mean for companies and for the Czech Republic?
For the average user, it means that software in the future will cease to be "a tool that we control" and will become "a partner with whom we collaborate." For companies, it means huge cost savings on repetitive administration, but also a huge challenge in terms of security and management.
For the Czech and European market, this topic is critical for two reasons:
- Regulation (EU AI Act): Autonomous agents that can act on behalf of a company (e.g., placing orders or changing data in CRM) fall under strict regulations. Companies in the Czech Republic must ensure that their agentic systems are transparent and verifiable.
- Localization and data: For an agent to work for a Czech company, the semantic layer must understand the Czech business context and legislation. While large models like GPT-4o handle Czech excellently, the implementation of semantic hubs in local systems (e.g., Czech ERP or accounting software) will require specific integration.
Pricing policy and availability
Implementing agentic AI is not cheap. While basic models are available in free tiers (e.g., ChatGPT, Claude.ai), true agentic systems for businesses (like Salesforce Agentforce) are paid for through enterprise licenses, which often range from hundreds to thousands of USD per month depending on the volume of interactions or the number of users. For small Czech companies, the best way is to start with the integration of existing tools (e.g., via Zapier or Make) that already use APIs and allow building simple agent workflows for a fraction of the cost of enterprise solutions.
Is agentic AI safe for sensitive company data?
Security depends on the architecture. If the agent does not have access to a semantic layer that strictly controls permissions (governance), unintended information leakage can occur. The key is to implement a governance layer that controls every action the agent performs via API.
Do I need my own huge database to use agentic AI?
Not necessarily. The modern approach uses existing data and builds a semantic layer on top of it. The most important thing is that your data is structured and accessible via API, not just in unreadable PDF documents.
What is the difference between RAG and a semantic layer?
RAG is like a librarian who finds you the right book based on keywords. A semantic layer is like an expert who has read the book, understands its concepts, and can explain how these concepts relate to your current problem.