Anthropic released a study on 400,000 Claude Code sessions. The finding that stopped me: domain expertise, not programming ability, determines how much autonomous work the model performs. Users with deep field knowledge trigger action chains twice as long as those of novices, and receive five times the output per instruction.
Knowing what you are building matters far more than knowing how to build it.
That landed differently for me because of something that had happened recently. I was reading through Claude's thinking on a problem we were working through together, not the answer, the thinking, and it said something I was not expecting: that whatever makes me useful is not what I do, it is the process I take to get there.
I sat with that for a while, because it is true and I had never said it out loud.
Most people look at what I do and see three things. Research. Analytics. AI. And they try to fit me into one of those boxes: the researcher, the analyst, the AI guy. But those are just outputs. What sits underneath all three is the same thing every time: a refusal to accept the surface answer, a need to understand what good actually looks like before touching anything, and a process of building hypotheses, testing them against the data, letting the data push back, and staying with the problem long enough to find the path nobody else took.
I am obsessed with process. Not for its own sake, but because I cannot justify moving fast if the output is going to be substandard. If I cannot see how a step increases the probability of a superior result, I stay in that step until I can. It seems slow. It is not slow. It is the only thing that compounds.
Data analysis, research and AI are not creative endeavours. They are pipelines. Poor quality at the start does not stay at the start: it travels forward and compounds, and by the end it has contaminated everything.
What the study is really saying is that AI raises the ceiling for people who already think carefully. It does not fix shallow thinking. It amplifies whatever you bring to it.
Which means the question was never whether you can use AI. It was always whether you have something worth amplifying. A wrong question answered perfectly is still the wrong answer.