I'm more and more impressed with AI every week. However, it's still be unable to perform some data analytic tasks that I've thrown at it. Perhaps I'm using the wrong models, or need to enhance the prompts. For example, I struggle to get AI to recognize that one data set has Eddie Smith and another is Eddie M. Smith, and to consider them the same. And that the first name field may have the middle initial in it as well. (Note in my limited data, I don't have currently have anyone with duplicate first and last names.)
I also find that data hallucinations are less likely now. I've run through some training that demonstrates that you can apply steps to validate an answer perhaps even using a different model to validate. If Gemini tells you that Napoleon ruled New Zealand, you can ask OpenAI to validate the answer. If found to false or questionable, either model can be re-asked the question and validated again. The downside of answer validation is time/cost.