Most of what I try with AI doesn't work

I try a lot of things with AI. Most of them do not work.

Not "needs more prompting" do not work. I mean I build the thing, use it for a few days, and quietly stop, because it did not actually make anything better. For every idea that earns a place in how I work, there are several that looked clever in the moment and turned out to be a worse version of what I already had.

This is not a complaint. It is the job. The hype makes it sound like every new release is a win you are too slow to adopt. The reality, if you are actually building and not just posting about it, is that the hit rate is low, and that is fine. The value was never in finding tools. Everyone has the same tools now. The value is in the judgment to tell the one thing that helps from the nine that do not, and to do it fast and cheaply.

So I have gotten ruthless about two things: making it cheap to try, and making it easy to kill.

Cheap to try means I can stand up a rough version of an idea in an afternoon, not a quarter. If testing an idea is expensive, you over-commit before you know whether it is any good, and then you keep it out of sunk cost. Make the test small enough that throwing it away costs nothing.

Easy to kill means I decide up front what "this worked" looks like, and I stay honest when it does not. The trap is the thing that is impressive but not useful: a demo that makes you go wow and then never fits into a real day. I have built plenty of those. The test that matters is boring. A week later, am I still using it without thinking about it? Almost always, no. Occasionally, yes, and that one is worth more than the other nine combined.

If you run a revenue team, this is the part to internalize before you buy another tool. The constraint is not access. You and your competitors can call the same models. The constraint is the judgment to run a lot of small, honest experiments and keep almost none of them. The teams that pull ahead will not be the ones that adopted the most AI. They will be the ones that got good at throwing most of it away.

Most of what I try does not work. That is not the failure. That is how you find the part that does.