What exactly is Agentic AI and why does it matter?
Most people still think of artificial intelligence as a chat interface like ChatGPT or Claude, where you type a question and it responds. While this is useful, it has its limits — the model is trapped in a "bubble" of text. Agentic AI breaks through this bubble. An agent is not just a model; it is a system built on a model that has reasoning capabilities, memory, and most importantly, access to external tools.
While a standard LLM (Large Language Model) can write an email for you, Agentic AI can insert factual data from your CRM system into the email, check your calendar, negotiate a meeting time with a client, and then save it in your calendar. To do this, it uses so-called reasoning loops, where the system continuously checks whether it is approaching its goal or has made a mistake.
Market projections: Huge differences in the numbers
Data from various sources paint a fascinating, albeit somewhat inconsistent picture. According to a study by SAP, the Agentic AI market in India is estimated at $14.4 million, representing fivefold growth. However, this is just a small fraction of the global picture.
Taking a broader view, Gartner reports that total AI spending in 2026 will reach $2.5 trillion. A very significant portion of this will be agent capabilities integrated into enterprise software, estimated at a staggering $201.9 billion.
Here we encounter an important distinction in definitions:
- Standalone market: The market for purely independent agent platforms (estimated at $7–8.5 billion).
- Embedded Agentic AI: Agent capabilities built into tools that companies already use (Salesforce, Microsoft, SAP), which accounts for the massive figure of over $200 billion.
Technology comparison: Who leads in "agentic reasoning"?
For an agent to function, it needs an extremely powerful "brain." Not every model is capable of maintaining a complex plan without getting lost. As of now (June 2026), we're witnessing a tight race between the main players:
- OpenAI (GPT-4o / o1 series): Currently the leader in logical reasoning and mathematical precision, which is crucial for agents when planning steps.
- Anthropic (Claude 3.5 Sonnet / Opus): Considered by many developers as the best model for agentic tasks thanks to its ability to follow instructions and "hallucinate" less in complex contexts.
- Google (Gemini 1.5 Pro): Excels thanks to its massive context window, allowing agents to work with extremely long documents or entire code repositories at once.
Price is also crucial for developers. While model prices for standard chatting (e.g., via API) are very low, agentic tasks are token-intensive. An agent must "think" in loops, meaning a single task can consume ten times more tokens than a question and answer with a standard chatbot. For models like GPT-4o, the price is in the range of dollars per million tokens, which companies must carefully calculate when scaling.
Challenges: Why 40% of projects fail?
It's not all growth — there are shadows too. According to data from McKinsey, only 23% of organizations have actually deployed agentic systems at scale. And Gartner warns that by the end of 2027, as many as 40% of Agentic AI projects could end in failure.
Why is this the case? The main reasons are:
- Reliability: An agent can get stuck in an infinite loop or make an incorrect step that has a chain reaction.
- Security and control: How much can we let AI decide on money transfers or data deletion?
- Implementation complexity: Connecting AI to real enterprise systems (ERP, CRM) is technically and security-wise very demanding.
Impact on the Czech market and European regulation
