AI-driven approximations, particularly machine learning (ML) methods, have demonstrated remarkable accuracy, sometimes approaching or even matching that of high-level quantum mechanical methods like Coupled Cluster (CCSD(T)). However, the extent to which they can match or surpass CCSD(T) accuracy depends significantly on the specific context, the quality and quantity of the training data, and the targeted chemical systems.

Therefore, in my opinion, I consider that if your system was not included or is very different from the training set. The best thing would be to go by ab initio methods. But this not only happens in AI, but also in semi-empirical methods or DFT.