You would just use pixel values as inputs, see https://www.tensorflow.org/get_started/mnist/beginners
You can make it more performant by downsizing the documents, but obviously this is incredibly GPU-intensive to train. I second the suggestions to just use OCR.
https://cloud.google.com/products/machine-learning/
Look at the text API which parses text from an image and gives you it's pixel location.
If not AWS, how could I get started elsewhere? Every tutorial I've looked at deals with a CVS data file, not an image.
Could you point me to a resource that has an example of how to train AI with image documents?