I definitely agree that it depends on how you use the data, although that seems to bleed into a judicial rather than an investigatory role.

In the case of this geofence kind of data request, though, it seems to self-enforce that a little bit, though, right? If you cast too broad a net, then you “catch” hundreds of people, and the jury rolls their eyes at you pointing the finger at any arbitrary one of them. If you construe the request narrowly, then you get a small number of leads, but those leads are in actual fact that much more incriminating by dint of how precise they are.

In some sense it reminds me of the way the squishier sciences dealt with p-hacking by normalizing preregistration: you kind of have to set the “power” of your request ahead of time, or the black box that spits out the results becomes less useful to you (and less convincing to the people you have to make agree with you).

Yes. And law enforcement seems to have consistent problems with p-hacking any new kind of evidence they are allowed to have. That's why the use of things like this tends to be denied.

That said, if some country manages to create a law enforcement organization with the right culture, it does indeed become much less of a problem. But it needs to guarantee the culture won't change either, and that the data won't become available for different organizations.