To what degree does this product (or others) look at a collection of documents and offer various possible schema to choose from? This seems like not just a "hard for AI" problem but a "hard for humans" problem. In other words, even a high-quality AI with lots of "thinking" time isn't enough. It isn't just about reasoning through a problem -- there will be a lot of judgment calls. Judgment calls that require significant context and domain knowledge. This to me seems like an area where human-in-the-loop really matters.