I like the idea by @Jonathan Goldberg about distinction based on 'actionability' and would like to refine it with measures of 'freetext contamination' where human intervention is needed when actions are needed. Flows which allow 'contamination' and still work are robust and when the contamination is detected in 'harmless' portions of the structured data, then automation is possible. This way even non-modular content is in better control.
e.g. footnote for a specific price in bid can prevent automation later on.
rgrds Paavo Muranen