I wouldn't call it smart with how the player can exploit this behavior when the AI try to seize them. They're too binary in term of committing the force, to the point of feeling too zergy. Of course, the opposite of AI always commit to a fight when it's not possible to win isn't smart either, but neither one where they only fight when victory is certain is a good system, despite how it sounds. Again, the reason why the player can cheese and exploit the AI because it's 100% predictable.
They should have it work based on a threshold. Like a 10% disadvantage is considered an acceptable risk, then have the AI roll a dice based on this probability to see whether it will accept the risk or not.