I can give you a concrete example: this week at work, it occurred to me that the 16 channels of expected and measured binary on or off test data I need to collect could benefit from a visualization because matching expectations will have visual properties that failures will not. So I had my AI agent create a script that encodes 16 channels of expected and measured binary wave forms over time, as a 32 channel 1Hz sampling frequency wav file, which I can view with audacity, which also has the necessary controls to measure time between transitions in the waveformms.
From hindsight, one could argue that since all of that solution consisted of rudiments of perfectly normal software that didnt need AI to be written or integrated, it was equally possible to create without AI. But knowing that could do it with the greatest of ease, for the total price of naming it, converted this from a project that required the motivation to figure out all of the necessary steps to one that just needed a good description.
Another things I noticed with AI assisted programming is the one track thinking. Someone has an idea, generate a working sample and then it becomes like a sunk-cost fallacy where they don't envision any other implementation choice or design. It's about adding more feature without taking a step back and assessing the overall goal of the project and if that feature is really needed.
Antoine de Saint-Exupéry has said it best:
“Perfection is achieved, not when there is nothing more to add, but when there is nothing left to take away.”