I don’t think it needs too much manual input these days. When I was analysing data, there’s a ton of ways to remove the normal responses and highlight those that need further investigation. Even then a lot of stuff can be compared to acceptable responses. For example, if you want to check for delivery look for either stop starts around lunch time and through the evening. Also look for journeys that have pauses within a single journey for longer than you’d expect for traffic lights, that’d spot someone leaving the car running whilst taking the delivery to the door. For misuse by another person, you can analyse journeys that don’t include the award holder’s address and there you have a list of regular journeys where the rightful person isn’t in the car.
There will be a need for someone to intervene once the computer has done its thing but that can come out of the already announced investigation budget.
I'm Autistic, if I say something you find offensive, please let me know, I can guarantee it was unintentional.
I'll try to give my honest opinion but am always open to learning.
Mark