Problems with automation

Random thoughts on why AI automation is more complex and context-dependent than it looks.
  • I keep hearing people say AI is automating jobs across different industries.
  • But what is actually being automated?
  • Take law firms. What are they really automating?
  • Most of it seems like grunt work during discovery, searching and indexing documents, maybe even using AI to summarize or generate arguments for legal teams. That likely carries over into civil and family law too.
  • But zoom in. We already had software that helped with document search, and humans who did the analysis. AI just improves the speed or reach of those tasks. That’s not automation, not really. Lawyers aren’t being replaced, they’re getting better tools.
  • Automation means no human in the loop. If you have to review the output, verify the logic, check the citations, then you’re not automating anything. You’ve just got a very confident intern who never sleeps and often makes stuff up.
  • What happens when the model offers an insight that turns out to be completely wrong? Or cites sources that don’t exist? Do you now have to comb through the whole discovery yourself just to verify the suggestion?
  • I use AI tools daily for code, writing, and hybrid tasks. I’d call myself an advanced user. I don’t trust them. They hallucinate. They misunderstand. They confidently break things. If I don’t supervise them closely, they often make things worse, not better.
  • I recently read a post about why AI agents can’t reliably handle long tasks. The longer the timeline, the higher the odds of a single critical error that derails the whole thing. One misunderstanding compounds into others, fast. The numbers were something like 53 percent success on a one-hour task, 4 percent on a five-hour task, and 0.002 percent on a ten-hour task.
  • Honestly, those numbers seem optimistic. I’ve rarely had success even at the five-minute level without handholding.
  • I’ve seen LLMs behave in ways that are plainly stupid, but also surface ideas I hadn’t thought of. When that happens, it’s not novelty, it’s reflection. Some smart person solved the problem before, published the result, and that thinking ended up in the training data. The model just spit it back at me in the right moment.