You can intercept every phone call in the world but you're no closer to finding a call organizing a terrorist attack.

A natural language model could act like an agent listening to every single conversation intercepted in far more depth than current search engines.

You create a prompt like "conversation between two terrorists planning an attack" and then compare every phone call's text to how similar it is to the prompt's output space.

You could also go deeper and also include every conversation they've ever had to see if it's a one-off false positive or there are "terroristic" trends to their speech.

You could also potentially link accounts and phone numbers and recordings by creating a style profile of known language and then again comparing an anonymous sentence to "in the style of Terrorist John Doe" to find potential linked data.