It's probably domain/industry/company dependent, but the vast majority (>90%) of NLP work I do nowadays is sequence-to-sequence models.