From chatbot to digital employee

AI agents represent a fundamental shift in how artificial intelligence works. While a classic chatbot like ChatGPT answers your question, an AI agent can act on its own — browse the internet, edit files, send emails, or conduct conversations on your behalf. It's not just a smart search engine, but more like a digital assistant to whom you delegate tasks and it carries them out from start to finish.

According to an extensive report by The New York Times from March 2026, these agents are finding their way into the daily lives of tech enthusiasts. They can search the web, write reports, edit files, and communicate via email or text messages — often completely autonomously while you sleep.

What do AI agents actually do?

Practical deployment of AI agents today covers several key areas:

Planning and logistics. The most recent example is Gemini Spark, which Google launched in early June 2026 as part of the AI Ultra plan for $99 per month. The Verge editor David Pierce tested it by planning a weekend trip to Hershey, Pennsylvania. Within minutes, Spark created a detailed itinerary — including routes from his home (he never shared his address), names and ages of his children, his wife's dietary preferences (she doesn't eat onions), nap times, and even the fact that concert tickets he had already purchased included parking. It mined all of this from emails, calendar, and search history, as Pierce described in his review.

Programming and development. The platform OpenClaw, an open-source tool for AI-assisted coding, has become a phenomenon. Its creator Peter Steinberger appeared at the Microsoft Build conference on June 2 and announced that OpenClaw now runs in secured Windows containers. "You can safely use OpenClaw even in a corporate environment," Steinberger said. OpenAI, meanwhile, reports 5 million weekly users of its Codex tool, which has long since stopped serving only programmers — companies use it for creating interactive websites and automating knowledge work.

Enterprise automation. At Build 2026, Microsoft introduced Scout — an "always-on personal agent" that can browse your inbox, participate in Teams conversations, check your calendar, and send daily summaries. It's part of a broader strategy of autonomous agents called Autopilots, targeted exclusively at enterprise customers. "They are autonomous, long-running agents with full enterprise security," emphasized CEO Satya Nadella on stage.

Microsoft's divorce from OpenAI: the agent war begins

The tech world's biggest power couple effectively split in April 2026 when Microsoft and OpenAI rewrote their partnership agreement. The result? Microsoft is no longer bound by exclusivity and can train its own models from scratch. At the Build conference, it unveiled seven new models including its first reasoning model MAI-Thinking-1, focused on mathematics and coding.

"As long as we had the exclusive deal with OpenAI, we couldn't train models at scale. Now we can," said Mustafa Suleyman, head of Microsoft's AI division. "We have to prove we can do everything ourselves from the ground up."

For the average user, this means one thing: the competitive battle for AI agents is in full swing. Google has Gemini Spark, Microsoft is building a Copilot super-app with Autopilots agents, OpenAI is preparing its own super-app connecting ChatGPT, Codex, and the Atlas browser. And Anthropic is entering the fray with Claude and the cybersecurity agent Mythos.

The price of privacy: how much does the agent know about you?

David Pierce's experience with Gemini Spark revealed perhaps the most fundamental dilemma of the entire agent era. Spark knew his home address, the names of his children and dog, his baby's age (for free park admission), his wife's allergies, and even correctly guessed nap times. All without a single query — mined from emails, photos, calendar, and history.

"There is a direct correlation between how much of yourself you reveal to an AI system and how useful it can be," Pierce wrote. "Google is in such a strong position precisely because it already has all this data. OpenAI and Anthropic are desperately trying to get it."

For Czech users, this has concrete implications. Most agent features are currently available only in English and in premium plans — Gemini Spark at $99 per month (roughly 2,200 CZK), and similarly premium enterprise plans from Microsoft. Czech localization for agents practically doesn't exist. However, if you use Gmail, Google Calendar, or Microsoft 365, the ecosystem is already building a profile about you that will one day allow an agent to function in as much detail as Spark.

What can go wrong: from bank account to legal disputes

The case of Sebastian Heyneman and the unwanted $31,000 debt is not unique. The New York Times describes other failures too: agents send emails to wrong addresses, access sensitive files without permission, or make up facts in conversations with real people. The problem is that agents get access to tools that can cause real damage — unlike a chatbot that "merely" hallucinates text.

Microsoft is trying to address this technically — at Build 2026 it introduced Microsoft Execution Containers, an isolated environment where agents run with limited permissions. "It prevents the agent from deleting all your files," Microsoft representatives explained. But the question remains: what if the agent makes a mistake that doesn't destroy your computer, but your reputation or bank account?

The European dimension: AI Act and agents

Thanks to the AI Act, the European Union has one of the strictest regulatory frameworks in the world. AI agents that act on behalf of a user and can cause harm fall into the category of high-risk systems. This means mandatory testing, transparency, and human oversight. For European companies that want to deploy AI agents, this brings higher costs — but also greater consumer protection.

For Czech businesses, this means that when choosing an agent solution, they should check whether it meets AI Act requirements. Most tools from Google, Microsoft, and OpenAI are adapting to European regulation, but local language support and integration with Czech systems still lag significantly behind.

How to choose an agent: a practical overview

For individuals:

• Gemini Spark (Google) — $99/month, deepest integration with the Google ecosystem, planning, emails, documents. Currently English only.
• ChatGPT + Codex (OpenAI) — from $20/month (Plus) to $200/month (Pro), coding, website creation, data analysis. Partially in Czech.
• Claude (Anthropic) — from $20/month, emphasis on safety and longer context. Partially in Czech.

For businesses:

• Microsoft Copilot + Autopilots — integration with Microsoft 365, Outlook, Teams; enterprise-grade security. Custom pricing, typically starting at $30/user/month.
• OpenClaw — open-source and free, suitable for technical teams; runs via API keys to models (OpenAI, Anthropic). Requires technical knowledge.

For the Czech market: none of the major agents offer full Czech localization. Communication in Czech is possible, but agents primarily operate in an English environment. For companies that need integration with Czech accounting systems or government administration, no ready-made solution exists yet.