This paper examines how the economic value of inference depends on the distribution of capacity and task information across stages of AI production. Controlled workflow experiments were conducted on externally verified software‑engineering tasks.
In two matched resource panels, direct execution achieved a constant success rate of 59.6% at logical‑token ceilings of 12,000 and 24,000, while success under information‑constrained planning rose from 36.2% to 51.2%. The planning disadvantage narrowed by 15.0 percentage points (95% task‑cluster bootstrap interval: 4.2 to 25.8).
A strict read‑only planning campaign compared whether the planner could see the task issue. At 12,000 tokens, issue access increased success by roughly 16 points. Compared with direct execution, task‑informed planning was about 10 points lower at 12,000 tokens but showed a 29.6‑point advantage at 24,000 tokens.
In the resource panels, direct execution used far fewer tokens than either ceiling, while the planning workflow’s binding rate fell from 46.2% to 0.8%, and downstream execution accounted for 89.9% of the total usage increase. Scale determines the capacity available to a system; workflow and information structure shape productive value.
Review