Any ml model can be called AI now.

A $20 battery-powered camera that can detect animals, humans, vehicles, and trees, and store + send an alert in real time of which it detects (as well as a description). Or a mini robot that can navigate obstacles and perform tasks while offline. Or a tiny industrial sensor that can detect a motor degrading due to the sound of its hum deviating from trained examples. Or an RF detection algorithm that ignores common band patterns and isolates unusual spikes in usage among dense noise. Or a motion tracking sensor that uses radar and video to track an object in a room. I mean really there's thousands of use cases. And this little Flipper One has 10x-100x the power of the ESP32-S3

I can't define AI, but I can define "great"

Great is the not-sucking.

If it is great, then it does not suck. If it sucks, then it is not great. To be great is not merely to be good, it is to actually not suck. Then you are great, in the most minimal, barely clawed yourself over the line way.

One I've personally done is wake-word detection for a local voice assistant. Not sure what exactly "great" is as I don't benchmark unless someone's paying me, but for my personal use case, good enough precision/recall that it doesn't annoy me!