Usually, the law has specific procedures about how this information is requested, what responsibilities are with which party, and how long the response time should be for suchs a request.
When starting (or already being an ISP). You already know what kind of system you need to build that matches all these requirements by law. Simply saying, we do not have the required information wouldn't work because the law has very specific details about the requested information.*
* this is in a european country, so no clue if this is applicable to the US.
As a follow-up the agency, with the right court order, could get all the raw connection records and try to figure it out themselves. But if you don't know the exact time and (source IP, port, destination IP, port) combination you're not going to figure it out in a network with large scale NAT.
(For those who haven't heard the reference https://www.youtube.com/watch?v=FKCmyiljKo0#t=0m40s )
A good CGNAT implementations have support for static blocks: the subscriber always ends up a a specific ipnumber+portblock combination. (Each subscriber is assigned a specific number of exit ports and this all just logged once during startup so you always know where each subscriber ends up).
Should they run out of their assigned portblock, there are pools which you can borrow from (these need then to be logged who borrowed at what time etc). So all in all there is less logging than when everything was dynamic.