For half a millennium, the fundamental cost structure of running a business was remarkably simple. When a company built a balance sheet, its operational expenditures were overwhelmingly dominated by two distinct categories. First, you paid your people, the Labour pillar, represented by wages, salaries, benefits, and payroll taxes. Second, you paid for the infrastructure that enabled those people to do their jobs, the Capital and Operations pillar, spanning office leases, physical machinery, and eventually, predictable SaaS software subscriptions.
But in 2026, this ancient, comfortable duopoly has permanently collapsed.
A volatile, metered, and liquid third pillar of corporate spend has emerged: Intelligence.
According to the June 2026 Ramp AI Index, a stark economic divide has opened up in the corporate world. While the median American business spends a modest $11.38 per employee each month on artificial intelligence, the top 1% of most aggressive adopters, the “AI-pilled” enterprises, are spending an astonishing $7,500 per employee every single month directly on AI platforms and API token consumption.
This is no longer a standard software line item. It is a parallel payroll. And it is completely breaking traditional corporate accounting.
The Rise of Cognitive Utility
To understand the scale of this disruption, we must first understand how we got here. When generative AI first entered the enterprise in late 2023, it was treated as an upgrade to the existing software stack. Companies purchased seats of ChatGPT Plus or Copilot at $20 to $30 per month. To a Chief Financial Officer, these looked exactly like Salesforce, Slack, or any other SaaS subscription: flat-rate, predictable, and easily capped.
But as AI-native companies began to rebuild workflows from the ground up, they realized that simple, one-shot conversational prompts were a bottleneck. To unlock true productivity, they turned to “agentic” systems, autonomous software agents that run in continuous, self-correcting loops.
These agents do not just answer questions; they perform deep cognitive labour. An agent evaluating a job candidate might conduct a screening, check public code repositories, run a background verification, draft an assessment, and cross-reference the results against millions of data points. A project management agent might write a piece of code, test it, read the error log, redesign the architecture, and try again, repeating the cycle dozens of times before presenting a finished product to its human supervisor.
This shift has transformed AI from a software tool into a variable utility, behaving far more like electricity or water than a Microsoft Office licence.
The financial implications of this transition are staggering. Speaking on the 20VC podcast in June 2026, Brendan Foody, the CEO of recruitment startup Mercor, made a startling admission:
“Right now, we’re spending more on tokens for our internal agents than we are on employee headcount.”
Mercor, which has conducted more than five million AI-assisted candidate interviews, uses autonomous agents across project management, candidate evaluation, accounting, and fraud detection. For Foody’s team, the traditional corporate pyramid has been inverted. Human employees are no longer the primary executors of work; they are the strategic editors of an autonomous digital workforce that consumes billions of API tokens every single hour.
The Token Cost Paradox
What makes this financial transition particularly challenging for CFOs is the high volatility and variable pricing of the underlying technology.
On paper, the raw unit economics of AI have never been cheaper. Over the past twenty-four months, competition among frontier model providers has driven the cost of processing a million tokens down by more than 98%. Yet, enterprise bills are tripling.
This is the classic Jevons Paradox in action: as a resource becomes more efficient and cheaper to consume, the overall consumption of that resource increases rather than decreases.
Because tokens are incredibly cheap, developers are building highly complex, multi-step agentic pipelines that run continuously. An interaction that would have cost a single prompt and five cents in 2023 now costs three dollars in 2026, because the agent is traversing thirty different loops of reasoning, search, and self-correction to ensure accuracy.
[2023: One-Shot Chatbot] User Prompt ──> LLM ──> Answer ($0.05)
[2026: Agentic Reasoning Loop] User Prompt ──> Agent ──> Code Generation ──> Execution Error ──> Self-Correction Loop (x10) ──> Final Output ($3.00)
This explosion in consumption is making flat-rate SaaS pricing models commercially unviable. On 1 June 2026, GitHub officially transitioned its popular Copilot tool to usage-based billing, citing the collapse of the flat-rate subscription model under the weight of resource-heavy agentic workloads. For an enterprise, a 50-person development team using Claude Pro or Copilot might look like a modest $1,000-a-month commitment on paper; under metered, agentic workloads, that actual bill can easily swell to between $15,000 and $40,000.
The cost of compute has grown so aggressively that it is beginning to dwarf even the most expensive human talent in the world. Bryan Catanzaro, the Vice President of Applied Deep Learning Research at Nvidia, revealed that for his team, “the cost of compute is far beyond the costs of the employees.” When a division populated by some of the world’s most highly compensated silicon valley engineers costs less to run than the GPU clusters they query, the traditional rules of corporate budgeting no longer apply.
Breaking the Legacy Ledger
The sudden appearance of this “token payroll” has caught traditional corporate finance teams entirely off guard.
Traditional enterprise resource planning (ERP) systems, expense management platforms, and procurement cycles were designed for a world of predictable, monthly billing. They are structurally incapable of managing the real-time, highly volatile spikes of metered cognitive infrastructure.
When a single runaway agent loop can burn through $10,000 of API credits over a weekend because of an unhandled exception in a coding script, standard accounting controls fail. Legacy procurement departments, accustomed to signing annual software contracts with structured discounts, have no framework for negotiating fluctuating, real-time token billing.
For modern finance leaders, surviving and thriving in this “AI-pilled” corporate landscape requires a fundamental evolution in strategy:
- Developing the “Cognitive COGS” Metric: CFOs must move away from treating AI as an indirect administrative expense (OpEx) and begin treating it as a direct Cost of Goods Sold (COGS). When AI is performing the core operational tasks of the business, such as recruiting, customer service, or software development, token spend must be mapped directly to the unit economics of the product or service being delivered.
- Implementing Real-Time Financial Circuit Breakers: Just as software engineers use rate-limiters to protect API endpoints, corporate treasurers must implement automated financial “circuit breakers.” If a department’s real-time token consumption spikes beyond a defined hourly or daily threshold, the system must trigger alerts or temporarily pause autonomous workflows to prevent catastrophic billing runaway.
- Optimizing via Multi-Model Portfolios: Ramp’s data reveals that the most advanced AI spenders do not rely on a single frontier model. Instead, they actively manage a portfolio. They route highly complex, strategic tasks to premium frontier models (such as GPT-4o or Claude 3.5 Sonnet), while shifting high-volume, repetitive agentic tasks to cheaper, open-source models (like Llama 3) hosted on their own private cloud infrastructure.
The Post-SaaS Balance Sheet
We are rapidly heading towards a corporate landscape where “headcount” will no longer be measured solely by the number of human beings on the payroll. Instead, it will be a blended metric of human capital and autonomous agent allocations.
In this new era, the role of the CFO changes completely. Historically, the finance leader’s job was to manage the cost of people and the allocation of physical assets. In the AI-pilled enterprise, the CFO’s primary responsibility will be the orchestration of cognitive efficiency: deciding when to buy a token, when to hire a human, and how to optimize the continuous flow of algorithmic capital.
The companies currently spending $7,500 per employee each month on AI are not outliers; they are a preview of the upcoming corporate standard. For the modern enterprise, the message is clear: the token payroll is here, and it is time to start balancing the cognitive ledger. An elegant short-form video featuring an interview with Mercor’s CEO on this precise dynamic is available for additional context on how fast token spend is eclipsing traditional payroll.
Mercor CEO on Token Spend vs. Salaries
This video is highly relevant as it features Brendan Foody, the CEO of Mercor, directly explaining the transition to spending more on raw AI tokens for internal agents than on their entire employee headcount.
