Your AI Agents Have a Salary * Your CFO Cannot See It

Your company pays a workforce it cannot see. Token prices have fallen 280-fold in two years [1]. Enterprise AI bills have tripled [2]. The math is not broken. Your accounting is.

Your financial system has a structural gap: a silicon workforce whose salary you pay by the millisecond but cannot see on any report. I will give you the Silicon Salary Model, a new operating expense framework that makes this invisible cost visible, measurable, and directly comparable to human payroll.

The Jevons Paradox of Tokens * Cheaper thoughts, bigger bills

The celebration over falling token prices is the most expensive misunderstanding in enterprise finance. The median price decline for large language model inference runs at 50x per year ^[3]. That number reads like salvation. It is an accelerant.

The mechanism is simple. Cheaper tokens do not shrink your AI bill. They enable more ambitious agent tasks. Agentic workflows have increased token consumption per task by 10x to 100x since December 2023 ^[4]. Every price drop invites a more complex delegation: longer reasoning chains, deeper research loops, multi-step planning across dozens of tool calls. Reasoning tokens, the internal thought process of models with extended thinking, are billed as output tokens, multiplying output costs by 10x to 30x on complex tasks ^[4].

The Jevons Paradox is the economic principle that efficiency gains in resource consumption increase total demand rather than reducing it. William Stanley Jevons documented this exact pattern in 1865 with coal. More efficient steam engines did not reduce coal consumption. They made coal useful for more applications, and total consumption exploded. Tokens are this century's coal. The per-thought cost is falling; the number of thoughts per job is detonating.

Enterprise generative AI spend surged to $ 37 billion in 2025, a 3,2x increase from $ 11,5 billion in 2024 ^[2]. Foundation model APIs alone consumed $ 12,5 billion of that total ^[2]. Your unit economics look better every quarter. Your total invoice looks worse. That is the Jevons Paradox of tokens, and every CFO celebrating cheaper AI is staring at the unit price while the aggregate bill compounds.

The Workforce Has Arrived * Inference now exceeds training

In 2026, inference spending will hit $ 20,6 billion, up from $ 9,2 billion in 2025 ^[5]. For the first time, running agents costs more than building models. Inference now captures 55% of all AI cloud infrastructure spend ^[5].

This crossover is the accounting signal that the agent workforce has reported for duty. Training a model is research and development. Running an agent is operational labor. When inference dominates, the primary AI cost on your books is no longer an R&D experiment. It is a workforce executing tasks, consuming resources, and generating invoices with every completed job.

The task-level numbers make this visceral. Adam Holter's task-adjusted cost model puts a single high-reasoning GPT-5 task at $823 and a Grok-4 task at $1,658 [4]. These figures account for the full token consumption of a complex agentic workflow with reasoning enabled, not a single API call.

These are wages, per task, for a silicon worker that finishes in seconds what a human analyst takes hours to complete. One enterprise deployed an AI predictive analytics tool at $ 200 per month during the pilot. At production scale, that bill ballooned to $ 10.000 per month ^[1]. The pilot looked like software. Production looks like payroll.

I run my own AI agent. A single task generates multiple calls across multiple models, and I watch the meter climb in real time. Multiply that across an enterprise deploying hundreds of agents, and you are staring at a workforce that submits its invoices by the millisecond, with no payroll department tracking the total.

The CFO's Blind Spot * No line item exists for variable cognitive labor

Your financial system was built for two categories: capital expenditures and operating expenditures. Software was a capital asset. People were a predictable salary. AI agents fit neither.

You cannot depreciate an agent over five years like software. You cannot forecast it in Q4 like a salary. An agent is a variable cognitive resource whose cost fluctuates with every task, every query, every recursive reasoning loop. I mapped this structural gap in my book AI Agents: They Act, You Orchestrate: your CFO does not have a line item for variable cognitive labor. You must create one.

The gap is not theoretical. Evan Goldstein, CFO of Seismic, put it plainly: "The era of buying AI for AI's sake is over. CFOs will remain willing to invest in AI but will require clarity on how it's tied to business outcomes." ^[6] Gina Mastantuono, CFO of ServiceNow, reinforced the mandate: enterprises expect "measurable gains in speed, resilience, and decision quality, not pilots and prototypes" ^[6]. Meanwhile, 88% of executives plan to increase AI budgets in the next 12 months ^[5], and 67% of finance leaders feel more positive about AI than the prior year ^[7].

The enthusiasm is real. The visibility is absent. Budgets are growing ahead of already-high expectations ^[2], with enterprises now apparently spending $ 590 to $ 1.400 per employee annually on AI tools ^[5]. AI coding tools alone represent $ 4 billion of the $ 7,3 billion departmental AI spend ^[2]. These numbers are buried inside existing cloud and IT line items where no one can isolate the signal from the noise. You are paying a workforce and calling it infrastructure.

The Silicon Salary Model *Three components to make the invisible visible

I built the Silicon Salary Model to bring this cost under financial discipline. The Silicon Salary, is a new operating expense category on your profit and loss statement that makes the cost of your silicon workforce directly comparable to the cost of your human talent. Three components compose it.

1. First, Cognitive Compensation: the agent's base pay. Every inference, every plan, every micro-calculation is metered. Claude Sonnet 4 charges $ 3 per million input tokens and $ 15 per million output tokens ^[4]. Claude Opus 4.1 charges $ 15 per million input tokens and $ 75 per million output tokens ^[4]. Unlike a salaried employee, the agent's labor bill scales with the complexity and volume of the work you assign. Leave the scope vague, and the meter spins. A single poorly framed research this command triggers recursive reasoning loops that compound cost from dollars into thousands.
2. Second, Tool and Equipment Costs. Your silicon employee requires its own arsenal. Every function call to a stock API, a proprietary database, or a booking system incurs a cost. These are rented capabilities. An agent that calls external tools without governance runs up a tab no one reviews.
3. Finally, Institutional Knowledge. The vector databases that hold an agent's contextual history are metered reservoirs of intelligence. Every retrieval-augmented generation call that pulls past experience into the present carries a price. This is the cost of keeping your silicon employee connected to what your organization knows. Starve this layer, and you degrade the mind of your Synthetic Labor force. Synthetic Labor is the class of AI Agent work that substitutes for human cognitive output at a fraction of the time cost.

By auditing your agents along these three axes, you calculate a total Silicon Salary per agent. This transforms a nebulous AI budget into a precise performance metric. You can place the return on investment of a silicon analyst directly alongside the ROI of a human analyst. Without this comparison, the conversation about where to deploy carbon versus silicon labor is impossible.

From Cost to Investment * The reframe your CFO needs

You might make the wrong turn at this point. The Silicon Salary is a workforce investment to optimize, with the same rigor you apply to human payroll.

Your CFO does not try to minimize human payroll. They try to maximize the return on every salary dollar. The same discipline applies to your silicon workforce. The only question worth asking: what is the productivity dividend of every token spent? This is the shift from the Economy of Intent as an abstract concept to the Economy of Intent as an operational accounting practice.

The Delegation Ladder, the framework I built for scoping agent tasks from vague intent to precise specification, becomes a financial discipline. Precision in delegation directly reduces Silicon Salary costs. A well-scoped task consumes 10x fewer tokens than a vague one running the same agent. Your Delegation Ladder is your cost control mechanism. Your Intelligent Circuit Breaker, the governance rule that halts agent execution when cost or risk thresholds are breached, is your spending cap. Your AgentOps function, the governance layer that provisions, monitors, and governs your silicon workforce, is your payroll department.

This is the architecture of a hybrid workforce managed with equal financial rigor on both sides: carbon and silicon. Every company that deploys agents without this framework is paying an invisible workforce with zero visibility into whether it earns its keep.

The Silicon Salary is a fiduciary obligation. Open your profit and loss statement. Find the line item for your silicon workforce. If it does not exist, you now know what to build. The meter is already running.

❗Note: The token prices and task costs cited here will age fast. AI inference pricing shifts by orders of magnitude within months. I wrote this article to establish the structural comparison between silicon and carbon labor costs, not to serve as a price sheet. The ratios hold even as the numbers move.

The Silicon Salary Model is one framework from one chapter of AI Agents: They Act, You Orchestrate by Peter van Hees. Across 18 chapters, the book maps the complete operating system for the Agent-First Era, from the AgentOps Trinity that governs your silicon workforce, to the Delegation Ladder that controls what agents spend, to the Human Premium Stack that defines where human labor remains irreplaceable. This article covers the cost; the book provides the architecture, the governance, and the strategy. Get your copy:*

🇺🇸 Amazon.com | 🇬🇧 Amazon.co.uk | 🇫🇷 Amazon.fr | 🇩🇪 Amazon.de | 🇳🇱 Amazon.nl | 🇧🇪 Amazon.com.be

References

^[1] Stanford 2025 AI Index Report (cited in PYMNTS), "AI Model Training vs Inference: Companies Face Surprise Bills," 2025. https://www.pymnts.com/artificial-intelligence-2/2025/ai-model-training-vs-inference-companies-face-surprise-ai-usage-bills/

^[2] Menlo Ventures, "2025: The State of Generative AI in the Enterprise," November 2025. https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/

^[3] Epoch AI, "LLM Inference Price Trends," March 2025. https://epoch.ai/data-insights/llm-inference-price-trends

^[4] Adam Holter, "AI Costs in 2025: Cheaper Tokens, Pricier Workflows," 2025. https://adam.holter.com/ai-costs-in-2025-cheaper-tokens-pricier-workflows-why-your-bill-is-still-rising/

^[5] Market Clarity, "Where is AI Spending Going in 2026?" 2025. https://mktclarity.com/blogs/news/where-ai-spending-is-going

^[6] Fortune, "AI in 2026: CFOs Predict Transformation," December 2025. https://fortune.com/2025/12/24/ai-in-2026-cfos-predict-transformation-not-just-efficiency-gains/

^[7] The CFO, "Five CFO Power Plays for 2026," November 2025. https://the-cfo.io/2025/11/24/five-cfo-power-plays-for-2026/