Granted, more than thirty years after publication I'm not sure spoilers are still worth worrying about. But, you know, internal consistency is a thing, too...
So the thing mentioned above is not something you know from the get go?
Supply the AI planner with a set of tasks (objectives) and let it plan out what needs to be done. Provide it with a set of actuators (this could be people, software, physical systems) which are capable of, given some input, accomplishing the tasks needed for the final objective(s). Give the same planner or a separate one the authority to act within various constraints (time objectives, financial and other resource constraints).
EDIT: Honestly, I could see a lot of intermediary jobs being replaced or largely displaced by such a planner. As an example, the USAF has a role called "Item Manager". Their job is to track materiel (primarily) needed for various systems and customers, and make sure that they're available when they need to be, get serviced when they need to be, are replaced when they need to be.
A (very abstract) example might be a particular ground vehicle, let's say trucks. They have maintenance schedules, they have spare parts that are needed, and replacements need to be available should one irreparably breakdown. So an IM may be responsible for every truck of that kind for 100 customers (various squadrons around the world) each with maybe 10-20 trucks on average.
A lot of their work is already semi-automated, but with sufficient care the full scheduling of maintenance, receipt of damage reports, maintaining a supply of spare parts and trucks, can be fully automated. With a human perhaps only involved to assess the damage reports submitted from the field, and to periodically sanity check the numbers, and to actually authorize spending requests. The AI planner can be given all the information needed (including budget, shipping logistics, storage availability) to do a lot of the planning itself.
[1] https://www.wsj.com/articles/the-worlds-largest-hedge-fund-i...