Probably accurate enough to learn how your body responds to different foods and exercise, to help prevent pre-diabetes, to help with general weight loss, and to help endurance athletes in their training.

But there is no universe it will be accurate enough to make insulin dosing decisions. Insulin dependent diabetics will require CGMs or finger pricks for another couple decades.

I manage my diabete with a "sensor" (Dexcom), an insulin pump, and a "loop device." I would NEVER use Apple Watch for therapy, and I don't think Apple wants to step into this minefield. BUT I guarantee you that having your glucose level as an (instantaneous/statistical) data point is a game-changer. There are "pre-diabetes" phases where even mild monitoring (and an alert) can be essential. And for those already in the tunnel, knowing you can have an additional "backup" alarm for hypo/hyper is very interesting (though, thinking about it, another alarm… no, better not :-D).

Pretty accurate since this works of averages - discarding outliners and norming get you close to the actual number pretty quickly - also with glucose your're interested in fluctuations more than absolutes

that's an interesting question.

there's sort of a usefulness threshold, and then there's a "can calculate insulin" threshold.

I think a LOT of people could benefit from plain high, medium, low with the understanding that you wouldn't make insulin decision based on it.