Why workflow products usually matter more than AI wrappers

Every time I have built something with an AI model at its center, the product work turned out to be everything around the model, not the model itself.

Work Search is a good example. The resume parser, the ATS scoring engine, the scheduler that keeps listings fresh, the normalization layer across job feeds — each of those is a harder problem than swapping in a better embedding model.

The same pattern showed up at UnitedHealth Group. The agentic workflows I built with LangGraph were only as good as the memory, handoff logic, and confidence thresholds around them. Strip those out and you have a chat box.

AI wrappers feel fast to ship because you skip the surrounding system. But that system is where the actual user value lives: persistence, automation, explainability, and the ability to act on results rather than just display them.

I now start every AI product by designing the workflow first. Where does data come in, what decisions get made, what happens to the output? The model is one step in that chain, not the chain itself.