Japjeev Kohli is Vice President of Technology at Transurban North America.
For the past two years, much of the enterprise AI conversation has centered on model intelligence. Which model is most capable? Which one reasons best? Which one writes better code, summarizes longer documents or handles multimodal inputs more effectively?
Those questions still matter, but as AI moves from experimentation to production, enterprise leaders face a different, more practical hurdle since they must now develop robust pricing intelligence.
The next phase of AI maturity will be about choosing the right model at the right price for the right task under the right operating conditions.
What Dynamic Road Networks Can Teach Us About AI Economics
A useful way to think about this shift is through the lens of dynamically priced managed lanes on major road networks. On Transurban’s 95, 395 and 495 Express Lanes facilities in the Washington, D.C., metro area, we manage roadway assets where capacity is finite, and demand fluctuates throughout the day. Pricing is one mechanism used to balance that system to maintain free flowing traffic even when congestion is high.
AI workloads are beginning to resemble this same dynamic. Model capacity is not infinite, demand varies across use cases and time and not every task requires the same level of performance. The triggers for routing AI workloads come down to three practical variables: complexity, urgency and risk. That distinction matters because AI is fundamentally shifting the economics of enterprise technology.
Traditional SaaS models are often priced around seats, licenses or predictable subscription tiers. AI usage, by contrast, is increasingly tied to consumption: input tokens, output tokens, cached tokens, context length, priority processing, batch workloads, tools, agents and multimodal usage. As adoption scales, this dynamic can quietly turn AI from a manageable innovation budget into a variable operating cost that touches every department.
From Cost Per Token To Cost Per Useful Outcome
For executives, it is clear that model strategy and cost strategy can no longer be separated. A raw token is not a sufficient unit of value. One million tokens from a frontier reasoning model are not economically equivalent to one million tokens from a lightweight summarization model. More importantly, a low-cost answer that requires human rework may ultimately be more expensive than a higher cost answer that is correct the first time.
Enterprises should stop asking only, “What does this model cost per token?” Instead, they should ask, “What does this model cost per useful outcome?” That shift requires the development of a new discipline in quality adjusted AI pricing.
Match Model Capability To Business Need
In infrastructure-heavy environments like ours, we are beginning to apply this principle in practice. Routine, high-volume tasks such as summarization, classification and internal productivity workflows can often be routed to more cost-efficient models that deliver acceptable quality at scale.
Higher-value or higher-risk tasks, such as compliance workflows, safety-related analysis or scenarios where precision and explainability are critical may justify more advanced models with stronger reasoning or multimodal capabilities.
The goal should not be blanket cost reduction. The goal should be intelligent allocation. This is where pricing intelligence becomes a strategic capability. Enterprises will need systems that can evaluate AI workloads dynamically and decide when to use premium reasoning, when to use lower-cost models, when to batch work asynchronously, when to use caching, when to route to a different provider and when to reserve capacity for mission critical use cases.
The Emerging Market For AI Inference Capacity
Over time, this may increasingly resemble a market for AI inference capacity. Today, most AI pricing is still based on posted rates from individual providers, but enterprise demand will likely become more dynamic. In that environment, the economic question becomes more sophisticated: What is the real-time value of a unit of AI inference, adjusted for quality, speed, reliability and compliance?
This does not mean every enterprise needs to build a trading desk for AI tokens, but it does mean that AI financial operations must become a core capability. Just as cloud adoption created the need for cloud cost management, AI adoption will create the need for AI cost intelligence. The companies that scale AI successfully will actively manage which workloads consume which models, under which service levels and at what effective cost.
The Questions Leaders Should Be Asking
Boards and executive teams should start asking a new set of questions.
• Are we using premium models only where premium intelligence is required?
• Do we understand our cost per AI assisted outcome?
• Are we measuring accuracy, error rates and human rework alongside token spend?
• Do we have visibility into which departments, applications and use cases are driving AI consumption?
These questions are not just technical. They are strategic.
The Bottom Line
The early enterprise AI race was about giving employees tools, experimenting with use cases and identifying where AI could improve productivity. The next race will be about disciplined scaling. Enterprises will need to move from enthusiasm to economics. That does not make model intelligence less important. It makes pricing intelligence equally important. The winners will not be the organizations that blindly adopt the most advanced models or the cheapest ones. They will be the organizations that understand the relationship between model capability, business value and cost.
The next advantage in enterprise AI will come from knowing when those models are worth the cost.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
