A problem with summaries is that it is not possible to verify their correctness without creating a summary by a human from scratch. In other words validation of the task costs at least as much as the original task.

This is different from say code generation where, to first approximation at least, automated tools (compilers, unit tests) can do this cheaply.