A model that answers a question in seconds can look like a breakthrough. The accounting starts when that answer is requested thousands of times, across multiple customers, with retries, long context windows, tool calls, human review and incident recovery. The visible token charge is only one line in the operating cost of intelligence.
Cost per answer is not the right denominator if answers differ in value and quality. Choose an economic unit tied to real work: a verified document, a reconciled transaction, a resolved support case, or an analyst hour saved with evidence. Measure marginal compute, engineering maintenance, integration, evaluation, security and exception handling per successful outcome.
Routing across models can be intelligent without being mysterious. Not every task requires the largest model. Classification, extraction and simple transformations may be handled by smaller systems, reserving expensive reasoning for ambiguity. But aggressive optimization can create unseen quality debt if evaluation is weak. The most expensive answer may be the incorrect cheap one that triggers an operational mistake.
A responsible AI margin model has three layers: what a unit of value is worth, what the system spends to deliver that unit reliably, and how costs vary as usage grows. The goal is not minimum token consumption at any price. It is maximum dependable value per unit of total cost.
Choose the right denominator
Cost per token is a supplier metric. A business needs cost per validated unit of useful work. For an invoice assistant, that might be an invoice correctly reconciled with no follow-up. For a data analyst, it might be an answer whose definitions and evidence can be reproduced. Different outcomes justify different model and review budgets.
Model total cost, not just inference
Add input preparation, search, tools, vector storage, egress, retries, manual review, evaluation, incident recovery and fraud controls. Include quality failure costs—not just average request time. A cheaper model that produces one critical error can make an apparently efficient system unprofitable.
Route by uncertainty
A tiered model strategy can use deterministic logic or smaller models for ordinary extraction while escalating ambiguous or high-impact work. But routing is only safe when the evaluation dataset reflects actual business failures. Optimize under a minimum quality threshold rather than minimizing raw cost.
The management question
Ask whether the system improves margin, customer experience or capacity at comparable quality. If not, the automation is a technical experiment, not yet a competitive advantage. A sustainable AI business model makes every expensive decision legible.