Thanks!
Yes I'd say it is a utility based AI at it's core. The AI always simulates a few diffferent moves from the ginormous number of options, this is done following some heuristics that try to cheaply guess which moves make sense; there's a bunch of auxiliary data structures that help the AI understand which tiles are important to defend or capture etc. After it has those candidate moves, it then runs a pretty expensive utility function to properly score the end result of each candidate move sequence and pick the best looking one.
Balancing defense and offense correctly was indeed a daunting part, but the solution I ended up is pretty simple. One of the candidate moves always includes commiting everything to offense (because sometimes that's the best defense :)), and the others each use different offense/defense splits. The utility function will decide which approach works out best for this turn.