Almost no job is one task wearing a trenchcoat
A lawyer's job is not 'draft documents.' It's drafting, plus reading a client's actual situation, plus judgment calls about strategy, plus negotiation, plus standing in front of a judge and being trusted, plus taking responsibility when something goes wrong. A customer support rep's job is not 'answer questions.' It's answering routine questions, plus de-escalating an angry customer, plus recognizing when a case needs to be escalated, plus representing the company's judgment in an ambiguous situation. Almost every real job is a bundle of dozens of distinct tasks, and they are not equally automatable.
This distinction — task automation versus job automation — is the single most important concept for reading AI-and-jobs headlines correctly. When a demo shows an AI system drafting a contract, summarizing a deposition, or writing working code, it is very likely automating one or several tasks inside a job, not the job itself. 'AI can write code' is true and has been true for a while; 'AI has eliminated software engineering as a profession' is not, because software engineering also includes understanding messy requirements, making architectural tradeoffs, debugging systems nobody fully understands, and being accountable when production breaks. The tasks that get automated first are usually the ones that are well-defined, text- or data-heavy, and don't require carrying real-world accountability.