How is an llm parser_spec to parser better than something like lex?

Because sometimes parsers can have weird inputs, like structured Excel files