The promise of automation in the legal sector faces a major structural hurdle: the tedious process of cleaning and annotating messy, unstructured legal data to make it machine-readable.