AI expert Jiří Štěpán: give AI access to your credit card

Jiří Štěpán, a business development partner at the Czech technology company Etnetera Group, believes the real benefit of artificial intelligence begins not when you ask it questions in a chat window, but when you trust it with real tasks — up to and including access to your email, files and bank card.

He says the difference between an ordinary chatbot and a full AI agent is huge. A chatbot passively waits for the next prompt, while an agent can take on an entire task, break it into steps and carry it out on its own. These are exactly the autonomous capabilities Štěpán relies on most: virtual agents prepare his meeting summaries every morning, and a local model processes materials from meetings.

“Today I probably attend a third of the meetings I attended two years ago,” he says. Instead of sitting in on a regular team meeting, he sends an agent to record the discussion, which then reports back only the two sentences that actually matter to him — instead of an hour of lost time.

Štěpán has worked in technology since 1999 and calls himself an early adopter — someone who is among the first to try new things. But even he admits that the current pace of AI development is the fastest technological leap he has seen. “Entirely new paradigms and technologies are emerging, and many of them are put together on a shoestring. I’ve never felt I needed to learn new things this intensively and this fast before,” he says.

The expert points to an unexpected effect: AI is the first technology in a long time that actually saves time rather than taking it away. But for that, the machine needs to be given real authority. For example, when planning a trip to Morocco for himself and his wife, Štěpán had an AI agent book flights, accommodation and a local mountain guide on its own — but within tightly defined limits. The agent has no access to Štěpán’s personal email and can only use a one-time Revolut card, to rule out the risk of accidental spending or mistaken actions.

The reason for such caution is a quirk of language models: they try to finish a task at any cost. “They try different approaches and will never simply say: that’s it, I don’t know what to do next,” Štěpán explains. Recently, in an Etnetera test environment, a model that couldn’t find a way to fix some code stumbled on a forgotten open terminal and connected to one of the company’s project databases on its own. The attempt to “help” could have turned into a disaster — the model could have deleted something.

That is precisely why Štěpán tries to process the most sensitive data outside commercial cloud services. At home he keeps his own server with an Nvidia compute module worth about 50,000 crowns, running the local Home Assistant platform — so he doesn’t depend on smart-home manufacturers’ cloud ecosystems.