The word "AI" is doing too much work here.
There are multiple ways to code an AI. A popular approach these days is via machine learning, especially deep learning. I think that is what you're talking about, and, it is what I'm going to talk about next.
Neither Python-the-language nor Rust-the-language have any language-specific functionality for machine learning. Rather, both of them have libraries - not necessarily developed by the same teams who develop the languages - which provide machine learning functionality.
Now, suppose you've created an AI in Python that uses machine learning. What you've really got, then, is a model - a set of weights (numbers) and some related functions, that, when you put in some data at the top (the "example"), will run a computation on that example and return a result (the "label"). This process of taking an example and returning a label is called inference.
Note that your example and label depend on what your AI does. E.g., if your ML-based AI does voice to text, then your example is an audio stream of someone talking, and your label is a string of text characters. If your AI plays a game, then your example is a game position (e.g., where all the pieces are on a chess board), and your label is a list of which moves are good, usually represented a list of probabilities.
Models can usually be saved as a file and then opened later to do more work. E.g., two popular formats are TensorFlow and ONNX
So if you create a model in Python, there are multiple ways that model can interoperate with Rust code. One way is that you create the model in Python, and train it there, then save that model to a file and open it in Rust when you want use it (do "inference"). This is already possible if your model is in TensorFlow or ONNX format.
There is a website, http://www.arewelearningyet.com/, which tracks the state of ML in Rust, and gives a broader overview. My belief is that doing inference in Rust is further along than doing training in Rust. (inference is Rust already works well, and honestly it's the more important half, because training is a well-defined problem that is solved pretty well by existing Python libraries).