AI is getting cheaper to use, but that doesn’t necessarily mean it’s cheaper to operate.
As model and token prices fall, organizations are using more AI across more users, workflows, and applications. Agentic AI accelerates that trend. A single request can trigger multiple models, reasoning cycles, retrieval actions, and tool calls, expanding consumption well beyond the initial interaction.
That makes cost per token a poor proxy for the economics of AI at scale. It captures only one part of the investment and reveals little about whether that investment is delivering value. Two workloads can consume vastly different resources and still produce comparable business outcomes.
As AI moves from experimentation to enterprise-scale operations, the challenge shifts from pricing to management.
Leaders need visibility into what is consuming AI resources, who owns that consumption, and what the business gets in return, whether it’s a customer interaction completed, a workflow automated, or a decision improved.
AI economics now requires organizations to govern consumption, cost, and value together.
Scaling AI creates a consumption and governance problem
The gap between AI investment and measurable returns is widening. Gartner found that only 22% of organizations have successfully expanded AI across multiple business units or moved to an AI-first model. At the same time, 85% of functional leaders planned to increase AI spending in 2026, after allocating an average of 12% of their functional budgets to the technology in 2025.
McKinsey’s latest research points to the same tension. Respondents reported that enterprise-wide AI scaling increased from 38% to 44% over the past year. But only 37% of respondents reported that AI contributed to their organization’s EBIT, essentially unchanged from the previous year.
In other words, AI deployment is expanding faster than the financial impact organizations can attribute to it.
Part of the challenge is that AI introduces a different consumption model than traditional software. Usage can expand dynamically across models, applications, and agents, making it harder to trace spending back to the workload, owner, or business activity that generated it.
This is what I call “agentic amplification.” As agents take on more work autonomously, a single human request can multiply into multiple models, workflows, reasoning cycles, retrieval actions, and external tool calls, creating cost that is harder to predict and govern.
It’s a dynamic that will become increasingly important as agents scale. Among respondents at organizations with more than $1 billion in annual revenue, McKinsey found that 40% are now scaling AI agents in at least one function, compared with 27% a year earlier.
As autonomous activity grows, aggregate token or model spend reveals less about the economics of an individual workload. Leaders need to understand which agents and applications are consuming resources, who owns them, and what business activities that consumption supports.
Even when organizations can see their AI spending, they often cannot explain what is driving it. Many leaders have sized the overall AI bill without mapping that spend back to the decisions, behaviors, and workloads behind it. Cost pressures are starting to surface as a result.
Without workload-level attribution, efforts to reduce consumption can be just as misguided as allowing it to grow unchecked. A higher-cost workload may still be worth the investment if it creates proportionally greater business value. Organizations need to distinguish between productive growth in AI consumption and consumption that delivers no commensurate return.
Even organizations that achieve workload-level visibility face another obstacle, i.e., the lack of a common measurement framework for AI economics. Providers measure, price, and report consumption differently, complicating efforts to compare workloads, benchmark efficiency, or establish consistent governance.
As AI moves into production at scale, organizations will need more consistent ways to measure and govern consumption. Industry efforts such as the Tokenomics Foundation reflect the growing need for common frameworks, terminology, and best practices around AI economics.
But standardizing consumption measurement is only a starting point. Those measures ultimately need to connect AI usage with ownership, cost, and business outcomes so leaders can determine whether increased consumption represents productive adoption or resources spent without proportional value.
That is the accountability gap enterprises need to close. It’s no longer enough to know what AI costs. Leaders also need to understand what that spending produces.
Build an operating model that connects AI consumption to value
Closing the accountability gap requires more than better cost reporting. Organizations need an operating model that connects AI consumption to ownership and business outcomes, then uses that information to decide where to optimize, invest, and govern.
-
Attribute AI consumption to the workload and business outcome.
Start by looking beyond aggregate token and model spend. Organizations need workload-level visibility into where AI resources are going, who owns each workload and what business activity it supports. That means building a more complete cost baseline that includes not only the model, but also retrieval, routing, infrastructure, validation and human oversight.
From there, define what success looks like for each workload. Depending on the use case, it could be a customer issue resolved, a workflow completed, processing time reduced, or a decision improved. The objective is to create a direct line of sight from resources consumed to outcomes produced.
Consider an organization running multiple internal AI agents through a common governance and measurement layer. By budgeting and allocating consumption to specific business activities, the organization found that agents producing similar outcomes could vary significantly in token usage.
That visibility allowed leaders to distinguish workloads generating value efficiently from those consuming resources without a proportional business benefit.
-
Optimize for the economics of the outcome, not the lowest consumption.
Once cost and consumption are tied to outcomes, you can begin to optimize more effectively.
That could mean right-sizing models, reducing unnecessary context, improving retrieval strategies, or routing tasks to different models based on the capabilities and performance they require. Inefficiencies that barely register in a pilot can become expensive when repeated across thousands or millions of production interactions.
Reducing consumption should not become the default objective. A workload that uses more resources may deliver better economics if it resolves a customer issue without escalation, eliminates rework, or produces a materially better result. The relevant question is whether the additional cost produces enough value to justify it.
That distinction matters even more with agents. Managing agentic amplification requires organizations to evaluate the additional resources an agent consumes against the work it actually completes, rather than simply counting tokens.
-
Create shared accountability for cost and value.
No single function has enough information to manage AI economics on its own. Business teams understand the outcome a workload is supposed to produce. IT understands the architecture and technical requirements. Finance can evaluate cost and return, while security, legal, risk, and privacy teams establish the guardrails within which it operates.
FinOps and IT asset management (ITAM) can help connect those perspectives. FinOps already brings disciplines such as allocation, forecasting, showback, chargeback, and optimization, while ITAM contributes ownership, vendor management, lifecycle governance, and commercial accountability.
Applied to AI, those capabilities can help organizations answer not only what a workload costs, but who owns it, why it exists, and whether it is delivering measurable value.
However, it requires both disciplines to evolve beyond their traditional focus on managing technology spend. As AI becomes embedded across the enterprise, cost, consumption, ownership, and value increasingly need to be managed together.
-
Build governance that can evolve with AI economics.
Don’t wait for universal AI economics standards before establishing your own measurement and governance practices.
You can start by defining workloads and agent activity consistently, establishing workload-level telemetry and allocation practices, and benchmarking comparable workloads to understand how efficiency changes over time. Those internal practices can evolve alongside emerging industry standards for telemetry, agent activity, billing, allocation, and benchmarking.
The goal is to create a common language that connects consumption, cost, ownership, and outcomes so leaders can make better decisions about where AI warrants continued investment.
AI is getting cheaper. So why is it costing more?
The economics of AI change as organizations move from experimentation to production. What works for a handful of pilots becomes harder to sustain as AI spreads across users, workflows, and autonomous agents.
Organizations that can trace that growing consumption to ownership, cost, and measurable outcomes will be better positioned to determine where AI warrants greater investment and where it doesn’t.
The shift AI economics now requires is from measuring what AI consumes to governing the value it creates.
