When it’s not about the content returned from an online service but instead the actual automation performed by the application, doing something instead of saying something, it becomes exponentially more challenging to replicate via LLM over a weekend. That’s the distinction I think you are looking for.

For example a self hosted application that provides real time service monitoring with performance metrics and can proxy that information between other data systems versus some full stack app that provides service information from the cloud.

On the other side of the coin I do write my own apps to avoid reliance on media streaming subscriptions.