What AI Can't Replace Yet

A couple of months back I was travelling home from a finished client engagement.

I had gone to the client to integrate their technology with a robot they had. Standard stuff so far. It just so happened that around that time Anthropic published their latest Opus model, and I'd had a chance to test it on the job.

So why am I telling you this? On the trip home I realised that over that engagement I had written no code. Not a single line of C++. And all of a sudden it hit me: if AI is writing all the code, what does the company need me for? What is my value proposition?

That instilled a real sense of fear. Because at first glance, this is what it looks like: AI doing the job of a programmer as well as someone with years of experience, only faster, with fewer bugs, and with more knowledge than any one person accumulates in a career. Unlike a person, limited by what they happen to know and the mistakes they happen to have made, the AI has all of that and more. So of course it writes better code.

In this despair I did the only thing that made sense, which in hindsight is a little ironic: I opened a new chat and told Claude. AI starts doing our jobs, and the solution is to go complain to the AI about it.

I told it the story, the observations, the whole spiral. I was expecting something like: "Yes, software development is dead, go become a plumber." I didn't get that.

What I landed on instead is that the AI's success on that engagement came down to a well designed framework: ROS. Topics, nodes, messages, the shape of the system is already decided before you write a line. Within that shape, Claude wrote good C++ with very little direction from me, because the judgment calls that define correctness had already been made by the framework itself.

That stopped being true on a different project, where I was designing a lightweight framework for robotics development in Rust, something in the spirit of ROS but leaner. Here nothing was decided yet. What should a node be. How should messages flow between them. What does correct even mean when you're the one setting the shape everyone else will build inside. On a task this open-ended, Claude's judgment calls were often wrong, and its code tended toward more complexity than the problem needed. I had to think about data flow, control flow, and what correctness meant before I could trust anything it wrote.

Claude was still useful there. It just wasn't a tool I could point at the problem and walk away from. On open-ended problems, where the shape of the thing being built isn't settled yet, AI still lacks the judgment to act fully autonomously. That's where I think the human value rests for now. Unfortunately, or maybe fortunately, it's probably only a matter of time before AI can do that too.