AI Economics: The Cost of Work You Can Actually Use

The finance team asks what the AI costs. Someone forwards the token bill. Everyone nods. Missing from the attachment: the afternoon spent correcting the output, the second attempt, and the customer interaction that never should have gone live.

The bill is real. The answer is incomplete. AI economics becomes useful when it measures the whole production process, including the human work needed to turn plausible output into something the organization accepts.

Why falling inference prices do not settle the question

The Stanford AI Index 2025 reports that the inference price for GPT-3.5-level performance fell from $20 to $0.07 per million tokens between November 2022 and October 2024. That is an important historical result about a specified performance level, not a current quote for the strongest available model.

It tells us that a capability can become dramatically cheaper. It does not tell us that every application’s unit cost falls by the same factor. When generation becomes cheap, teams may generate more candidates, use longer context, add more agent steps, or expand into tasks with a higher review burden. A cheaper ingredient can coexist with a more elaborate recipe.

Build the denominator before the dashboard

“Accepted” needs an operational definition. For a content workflow, it might mean fact-checked, original, internally linked, accessible, and approved for publication. For support, it might mean the issue was resolved without an inappropriate promise or later escalation. For code, it might mean the change passes tests and a real user journey without violating scope.

These definitions are not interchangeable. An output accepted because it looks polished is not the same as an output accepted after an independent check. If the denominator changes every week, the cost-per-outcome chart can improve while the service gets worse.

Start with a fixed task batch and a clear gate. Record every attempt, not only the successful ones. Include review effort on rejected outputs. Those minutes are still paid for, and they often explain why a cheaper model is not the cheaper system.

The invoice has at least four layers

Inference is the direct model workload. Separate input, output, caching, and extra usage where the provider exposes them. A subscription fee is an access expense, not a clean measure of underlying resource consumption; see the subscription bargain.

Orchestration is everything that makes the model useful: retrieval, storage, tool execution, monitoring, and the surrounding application. Calling these costs “not AI” does not make them disappear from the workflow.

Human attention covers scoping, review, correction, and exception handling. Use actual observed minutes multiplied by a defensible loaded labor rate. Do not claim savings by valuing the removed work at a senior rate and the new oversight at zero.

Failure exposure includes rework and incidents. Not every risk has a reliable expected-dollar estimate. When it does not, report the uncertainty separately and use hard constraints for unacceptable failures instead of manufacturing a precise average.

The most expensive output can be the one that looked finished.

What an honest comparison looks like

Suppose two systems complete the same batch of tasks. One has a lower model bill but a lower acceptance rate and more review. The other spends more per attempt but produces work that needs less correction. You cannot decide between them by comparing token prices; you need total cost per accepted unit and the distribution of important failures.

The workbench’s defaults make that distinction visible: 100 attempts at $0.50 each cost $50 in inference. Six review minutes per attempt at $60 per hour cost $600. If 80 outputs are accepted, the modeled cost is $8.125 per accepted output before overhead. These are illustrative assumptions, not an industry benchmark.

That example also explains why prompt polishing can be the wrong optimization target. If review dominates the bill, a clearer acceptance specification or better evidence retrieval may matter more than shaving a little output length. The relevant lever is whatever moves the accepted result, not whatever produces the most impressive token-saving percentage.

Route by uncertainty, not just difficulty

A small model may handle a narrow, well-specified classification reliably. A more capable model may be justified when the context is ambiguous or the cost of an error is high. Some tasks should remain human decisions regardless of model price. A routing policy should consider uncertainty, impact, and fallback—not simply attach “cheap” to routine work.

Measure the routing policy as a system. If a cheaper path escalates half its cases, the escalation expense belongs in its cost. If an expensive path eliminates a substantial review burden, that benefit belongs in its result. This is a more operational version of matching a model to the job.

The metric that should survive the meeting

Report the task, sample size, acceptance rule, total expense, cost per accepted outcome, review minutes, and material failure types. Keep a baseline. Repeat after any meaningful change to the model, workflow, or task mix. The aim is not a beautiful universal KPI; it is a comparison someone can audit.

The future of AI economics is less about finding the lowest token price than designing a process in which useful work is cheap to recognize and safe to accept. That connects directly to evaluations, operator workflows, and AI unit economics.