AI has made consumption measurable in extraordinary detail.
Companies can track tokens, inference calls, API requests, GPU time, agent actions and increasingly complex combinations of these measures. But as AI moves deeper into customer-facing products, simply knowing how much AI is being consumed is no longer enough.
The more important question is becoming:
Is that consumption profitable?
Boston Consulting Group recently argued that as AI moves from pilots into production, companies need to think differently about the economics of intelligence. Token costs can scale quickly as AI becomes embedded across workflows, products and customer interactions, making it increasingly important to understand not just what AI costs, but what economic return that consumption produces.
For companies building AI into commercial products, that creates another challenge: connecting consumption not only to cost, but also to revenue and margin.
The AI cost conversation has understandably focused on controlling infrastructure spending.
Organizations are asking which models should handle which workloads, where expensive frontier models are necessary, how context can be optimized, and how token consumption can be governed.
Those questions matter. But they address only one side of the equation.
Once AI is embedded in a customer-facing product or service, inference becomes a cost of delivering that service. BCG points out that unlike traditional SaaS economics, AI inference increases with customer interaction and can directly put pressure on gross margins. Its analysis suggests gross margins of approximately 65% to 80% for AI-enabled software and 50% to 65% for AI-native products.
At that point, Finance and Product teams need visibility into more than aggregate AI spend.
They need to understand what individual customers, products and workflows are consuming, what that consumption costs, what revenue it generates and what margin remains.
Consider two customers paying the same monthly price for an AI-enabled service.
One uses relatively simple interactions routed through lower-cost models. The other runs complex agentic workflows requiring longer contexts, repeated reasoning cycles and higher-cost models.
From a subscription revenue perspective, the customers may look identical.
From a cost-to-serve perspective, they may be completely different.
That gap matters because token costs are not uniform. BCG estimates that the cost difference between a relatively simple AI model and a frontier model can range from roughly five to 25 times per token.
As AI consumption becomes more variable, averages can conceal important differences in customer profitability.
A company can grow revenue while some customers become increasingly expensive to serve. If their AI usage rises faster than the revenue they generate, those customers can gradually reduce overall margins.
BCG proposes cost per outcome as a more useful way to assess AI economics than raw activity metrics. Instead of focusing only on tokens consumed, organizations should consider measures such as the cost of a resolved service ticket, qualified lead, completed analysis or successful customer interaction.
For companies selling AI-enabled products, the same idea should go one step further.
The commercial progression becomes:
Cost per token → Cost per outcome → Revenue per outcome → Margin per outcome
That shift changes the conversation.
If an AI agent completes a valuable workflow, knowing that it consumed 50,000 tokens is useful operationally. Knowing that the workflow cost $2 to deliver is better. But if the provider charges $1.50 for that outcome, the commercial problem becomes immediately visible.
Likewise, higher AI consumption is not inherently bad. A customer generating substantial token volume may also be generating highly profitable revenue.
The objective should not simply be to minimize consumption. It should be to understand the economics attached to that consumption.
Achieving that visibility requires companies to connect technical consumption data with commercial context.
A raw usage event alone does not reveal whether consumption was included in a subscription, deducted from prepaid credits, covered by a committed spend agreement, priced within a tier, treated as an overage or charged according to a customer-specific contract.
This is where AI unit economics becomes a monetization infrastructure problem.
Companies need to know:
What was consumed? What did it cost? How was it commercially rated? What revenue did it generate? What margin did it produce?
That visibility becomes increasingly important as AI companies experiment with combinations of subscriptions, usage charges, credits, commitments and outcome-based pricing.
Without it, teams risk managing AI costs on one side of the organization and customer monetization on another, without a clear view of the economics connecting the two.
BCG argues that companies should not aim for the smallest AI bill. They should aim for the highest return from the intelligence they consume.
The same principle applies to AI monetization.
The goal is not necessarily to reduce customer consumption or maximize the price of every token. It is to create sustainable economics where greater consumption can translate into greater customer value, greater revenue and healthy margins.
As AI products mature, margin per customer, workflow or unit of consumption may become one of the most important measures of whether an AI business model can scale.
The companies that can connect consumption, cost, pricing and revenue at that level will be in a much stronger position to make informed decisions about packaging, pricing, product investment and customer profitability.