Who is liable when the AI decides?

i think in the long run it would help having human AI trained managers for these agents, which continue to tweak instructions for the models running these agents, and similar to how there's performance reviews for agents, i think agents would have to do something similar