A lot of AI conversation still treats the prompt as the main event. That made sense earlier, when the tools were novel and the easiest visible win was seeing a better response from a better instruction. But once a team is trying to use AI seriously, the bottleneck usually moves.

The higher-leverage questions tend to be workflow questions. Where does the context come from? What gets persisted? What should happen automatically versus manually? Who reviews output? What is the failure mode? What gets retried? What gets escalated? How much ambiguity can a user tolerate before the system stops feeling useful?

That is why I am interested in AI workflow design as a category of work. It sits between technical capability and actual usefulness. A strong workflow can make an imperfect model far more useful. A weak workflow can make a great model feel unreliable.

Claude Code and Codex matter here because they lower the cost of exploring possible systems. You can mock up an internal tool, test a research flow, or build a rough agent loop quickly enough that the conversation changes from abstract argument to concrete tradeoff.

The important thing is that workflow design is not just orchestration for its own sake. It is about deciding where intelligence should sit, how much structure the user needs, and what kind of human trust the system has to earn.

When I think about AI product opportunity, this is one of the main places I look. Not just “what can the model do?” but “what workflow becomes dramatically better if the model is placed in the right spot?”

For companies, that is often where the real leverage lives. For me, it is also where the most interesting work is.