1. The problem case is very well specified
2. There's a verification harness it can work with
3. You don't care about long-term maintainability or security
Producing the things that solve for these is 90% of the job.
Consider what goes into setting up automated verification: how do you write unit tests when the units aren't built yet? You need to understand and design them. That's the entire premise behind TDD. You design the code through writing the tests first.
I don't see AI improving on this area. Encoding business invariants into high value automated verification isn't itself a verifiable task. Neither are code quality, security, or knowing when something is under-specified.
For production work specifically I think the future may look like developers doing mostly specification, verification and review with the AI doing the plumbing. I'm not actually convinced this is a long term win as there are a lot of trade-offs, but we'll see.