Originally posted on LinkedIn on June 2, 2026
I follow AI news daily, and I’m impressed by the rate of progress. But it hits differently when you experience new capabilities first hand.
Over the last two months I’ve been “vibe coding” games, apps, and websites. I’m still surprised by what I can now create in a single day with the right prompts and a few usability iterations.
More recently, I’ve been experimenting with agentic tools like Claude Cowork and OpenAI’s Codex for more everyday work. Not just writing or coding, but tasks around files, email, calendars, and admin workflows.
What quickly became clear is that a lot of office work is becoming delegable. Not perfectly, but enough that the direction of travel is obvious.
And when AI tools cannot connect directly to your apps through plugins or APIs, they can increasingly operate them the way a person would: by looking at your screen, clicking, typing, and navigating interfaces. That “computer use” layer is a big deal.
Of course, this raises the important question: how much should we trust AI with our stuff?
Files, emails, calendars, client data, financial transactions, and private messages are not trivial things to hand over to a probabilistic system connected to the internet. The security and privacy risks are real.
What surprised me most, though, was how quickly my own guard dropped.
After just a few days, I found myself giving the AI more permissions than I initially intended. Why? Because it was useful. Accurate enough. And just too convenient.
That may be one of the biggest near-term risks: not evil AI, but useful AI quietly training us to lower our boundaries.
Curious: what office work are you already outsourcing to AI, and how are you managing the risks?
Back to AI experiments and musings