Wait 1 year until we have something like chatgpt but with vision integrated in. Currently it's typically an ocr followed by some nlp. But in a year it can be as simple as give a few examples of what you want done (few shot prompting) in a single model hosted online somewhere

I'd wait a bit more.