When the bot takes the wheel, who holds the brakes?

A conversation about the “wild west” of AI agents, the risks of total system access, and the transition from manual dashboards to autonomous dialogue.

Host Péter Civin talks with György Fekecs, Tamás Leszkovár, and Péter Varga about the evolution of AI agents within DT-ITS. They explore the thin line between a helpful digital assistant and a security nightmare, and why the biggest hurdle to innovation isn’t the technology itself, it’s our own habits.

In the world of autonomous agents, “access” is everything. It is the difference between a tool that summarizes a meeting and one that can accidentally wipe a hard drive or empty a bank account. Moving from static bots to dynamic agents changes the stakes for every user and every organization.

The “Hacker’s Dream” in your terminal
A sobering reality check: for an AI agent to truly “work” on your behalf, it often demands the keys to the kingdom: file access, terminal control, and browser permissions. While the promise of a bot managing your monthly bank transfers sounds like the peak of convenience, the reality of running these tools locally without a “sandbox” is a massive security gamble.
The episode highlights why isolated environments aren’t just for developers anymore; they are the mandatory safety goggles for anyone experimenting with the next generation of automation.

Beyond the Dashboard
A significant shift is happening in how we handle information. Traditionally, data moves through a multi-step factory: developers code the logic, analysts build the Power BI or Tableau dashboards, and finally, the human interprets the result.
AI agents are shortening this loop. Instead of clicking through a pre-built chart, we are moving toward “dialogue-driven” data.

The QWERTY trap: Why habits outlast technology
Why do we still look at dashboards? Why do we still use the same keyboard layout designed in 1873 to stop typewriter keys from jamming? The “QWERTY effect” explains the lag in AI adoption.
Even when a better, faster, or more autonomous way exists, human habits are hard to upgrade. We trust what we can see, what we already know, even if a conversation with an agent would be objectively more efficient than staring at an Excel sheet.

Enterprise safety: Innovation within the “fence”
At DT-ITS, the “Agentic wild west” is being tamed through projects like the Stability Bot and the Chatbot Factory. These aren’t just tools; they are proof that AI can be both powerful and compliant. By keeping data within European borders and utilizing internal “Hyper-scalers,” the team demonstrates how to use agents for:
– Automated code reviews: Finding vulnerabilities before they become problems.
Instant documentation: Writing the “boring” parts of software – development automatically.
– Secure context gathering: Allowing AI to help without the data ever “leaving the house.”

AI agents are here to stay
The episode identifies the key levers for a successful AI transition:
– Sandboxed experimentation: Providing safe “playgrounds” where agents can’t cause real-world damage.
– Sovereign AI models: Ensuring data remains within the company’s own security perimeter.
– Habit-breaking design: Creating tools that are so much better than “legacy” ways that users are willing to learn a new “keyboard.”
– Cost-awareness: Balancing the speed of AI-generated insights with the long-term cloud costs of running them.

You don’t need to fear the agents. You just need to make sure they are working in a room where they can’t accidentally set the house on fire while trying to cook.

Listen to the full episode in Hungarian here: https://www.deutschetelekomitsolutions.hu/podcast/odaadnad-minden-adatod-egy-clawbotnak/