Who's responsible when AI fails?

Published Aug 19, 2025 · Around 3 minutes to read

I interned at an architectural firm 20 years ago. My first real job. I thought it was super cool to work on an actual building that people would live and work in.

My first task was to create a door schedule in AutoCAD. Basically, it’s a list of all the doors in the building (the door type, size, material, location etc.), that made it easy to coordinate buying and installing them. I was a pro at AutoCAD, so I knocked this out quickly.

My next task was to print the construction specifications for this project. Contractors used these specs to bid on the project. They were hundreds of pages long, and each contractor (there were about 15) got their own set. So I had to print thousands of pages.

I decided to be environmentally friendly and print them double-sided. It took me hours to print and assemble everything. When I handed the finished sets to the lead architect, he flipped through the first couple of pages and stopped: “These need to be single-sided.”

Crap.

I had to reprint all the specs again. In order to save a few hundred pages, I wasted thousands. Not to mention all the wasted ink.

—

AI agents are interns. Enthusiastic, but can veer way off course if they’re not directed properly. Interns need guidance: you onboard them, show them the ropes, supervise them, and verify their work.

Building AI agents is no different. You can’t just drop an AI agent into your tech stack, hook it up to a few APIs and let it run wild. You need carefully evaluate its every action, every step of the way.

“But I thought agents were autonomous. How is this different than traditional software development?”

Surprise: it isn’t.

90% of building agents is still traditional software development. But with a fancier name. The biggest difference is that observing the human process becomes even more crucial, because we’re enabling humans to do more with agents. The Agent Experience (AX) is really an intern experience: onboarding the agent, showing them the ropes, supervising them, and verifying their work.

Plus you wouldn’t give interns mission-critical work. No, you’d give them tedious and boring tasks, like getting a list of all the doors in the building. Computers (and interns) are best suited for repetitive, clerical tasks.

Agents should free up human brainpower to be used on actual thinking tasks.

"I want AI to do my laundry and dishes so that I can do art and writing, not for AI to do my art and writing so that I can do laundry and dishes.”

At Focused, we’re building AI agents that can handle the low-value work, freeing up humans for more meaningful work. This is where I believe the real value of agents lies.

—

P.S. What stuck with me after all these years is the utter disbelief in the lead architect’s face when I handed him the specs. His expression said: “How could this even be possible?” Single-sided specs were an industry standard. They needed to be back then because it was easier to flip through them for review, markup, and distributing specific sections. I think it’s all digital now.

P.P.S. If a human messes up, you can be somewhat confident that you’ve paid for the lesson they’ve learned, and that they’ll never make that mistake again. But there’s zero tolerance if an AI agent messes up because you can never be sure if the AI has learned. Which is why it’s even more critical to build robust evaluations alongside the agent.