TMI isn't a thing if everything is digital. Machine learning classification techniques (go look up something as simple as maximum entropy) can do a great job of identifying classifications with high accuracy. What is being classified? Well, presumably whatever "they" think are threats to the nation, or at least to whoever has control of the system. One can analyze the behavior of targets deemed a threat and find common features shared between those targets. Even stuff a human would never, ever think to correlate could matter (the humidity, time of day, day of year, AND whether they are a certain religion). The beauty is that a human doesn't need to work out what correlates with a threat. The machine does it. You give it features, it gives you statistical probabilities that the entities in your data are a threat. It would take an enormous amount of computing power to do this with the amount of data the NSA apparently has. Something like this for example:
Then it is just a matter of drawing the line for the threshold of what constitutes a threat. I just described something someone could have done 10 years ago. Machine learning has come along pretty well since.
The state of affairs is so disturbing because all technical hurdles to a dystopia have been overcome. Someone with these resources won't suffer from information overload. There DO exist learning algorithms which can deal with this much data and they clearly have invested in the necessary hardware. Laws and morality don't appear to be slowing them down. What safeguards are left...?