The process compiler
This creates a second-order effect on AI demand. General intelligence may be needed to understand a messy process, infer rules, handle exceptions and design a better architecture. Once the recurring structure is known, execution can migrate to cheaper mechanisms.
General reasoning → discover recurring structure → encode/specialize → cheap execution
The endpoint need not be literal deterministic code. It may be a rules engine, state machine, database constraint, classifier, small model, cache, lookup, protocol, sensor or specialized controller. Computing has always moved expensive general work into cheaper specialized execution through compilation, caching, indexing and precomputation.
AI potentially expands the class of processes that can be 'compiled' in this economic sense. The paradox is that successful AI transformation can reduce the amount of frontier AI needed per unit of mature output.
Recurring AI demand may track unresolved novelty, not total automated activity
This does not imply that novelty stops. New tasks and environments continually arise. It does imply that a valuation based on perpetual frontier inference for every automated task must explain why optimization does not progressively route mature work downward.