I know for certain it is not the limits of AI capabilities we have today because we have seen some incredible machine learning algorithms applied to games like Starcraft and Dota and even For Honor. Just a matter of priorities, resource management and design choices by TaleWorlds.
It is important not to confuse coloquial AI with "true" AI. What most people call AI in games is nothing but scripted behaviour. True AI is non-deterministic, specially when it comes from the realm of Machine Learning (a subset of Computer AI)
True AI is often not suited for Video Games for two reasons:
1- Machine Learning and Neural Network construction is as much an art as science. Good developers on these areas are scarce and expensive.
2- True AI is
alien. It is not great a mimicking human behaviour. Computer Neural Networks are not Neurons, and their input/outputs are not our 5 senses and muscles. When DeepMind made history by deafting a Go champion (a game thousands of years old and one of the more complex games ever created by humans), it was noted that it behaved in a way that no human had ever tried to play Go before. Something similar happened with the famous Dota loss to DeepMind's AI. Several big players noted how it behaved in ways they had never thought of playing before.
AI is hard to code. Be it the coloquial one or true one. Changes to small behaviours can lead to unpredicted fallouts in other areas. The scenarios are so vast that is often not feasible to create exhaustive automated testing.
Unless I am mistaken, Bannerlord uses
Agent-Oriented programming, which is different from the far more common Object-Oriented or Functional programming paradigms. I am only tangentially familiar with the concept, but I would imagine it is not as straight forward to automate testing as the two others.
Automated testing is the important part here, because without it software changes become costly and risky. So don't fault them too harshly for iterating "slowly" on AI issues.