For most of modern economic history, the cost of cognition was indistinguishable from the cost of the person doing the thinking. You could buy capital, plant and machinery, software licenses, you could buy labor, salaries and benefits, but if the work was thinking, the line item was always payroll. The capital-to-labor ratio in a knowledge-work company was, in practice, the ratio of laptops to brains.
That has now changed in a way the accounting categories have not yet caught up to. Cognition has, for the first time, a competing supply at a measurable per-unit cost: the LLM token. The interesting consequence is not that AI is cheap. It sometimes is and sometimes isn't. The interesting consequence is that for the first time we can write down a ratio between two cleanly priced substitutes for the same kind of output, and ask what the ratio tells us.
I'm going to call them Token Capital and Human Capital. Token Capital is the total dollar spend a person, team, company, or economy commits to AI inference (prompted, retrieved, generated, and orchestrated). Human Capital is the total dollar spend committed to humans doing cognitive work (salaries, benefits, training, and the overhead that makes them productive). The ratio TC/HC, expressed as a percentage or a multiple, is the variable I want to interrogate.
two priceable substitutes for one kind of output
Knowledge work has always had three inputs: human time, software, and infrastructure. The cost of the last two has been falling for forty years. The cost of the first has mostly been rising. The aggregate cost of producing one analyzed paragraph, one rendered design, one passing test, or one investigated incident has stayed roughly proportional to the labor it required, because labor was the binding constraint.
An LLM token does not replace the human entirely, but it replaces a slice of what the human does: the synthesis, the first draft, the search, the lookup, the diagram, the recap, the test scaffold. Each of those slices used to be a tiny fraction of a salaried hour. Now each is a tiny fraction of a tokenized inference. The substitution is partial, but it is real, and it has a price.
What makes the ratio interesting, rather than just the absolute spend, is that it normalizes for size. A 10-person team with $50k in annual token spend and $3M in payroll has a TC/HC of about 1.7 percent. A 200-person company with $2M in token spend and $60M in payroll has the same ratio. The number says something about how the work gets done, independent of how big the entity is. That is exactly the kind of dimensionless quantity worth tracking over time.
defining the units honestly
The numerator looks simple but isn't. Token Capital should include: direct API spend, seat-based AI subscriptions (Claude, ChatGPT, Cursor, Codex, Gemini, Perplexity, and the rest), inference passed through orchestration layers (n8n, LangChain, custom agents), embedding and vector storage that exists only because of LLM workflows, fine-tuning runs, and any GPU compute amortized for model use. It should not include legacy software licenses, generic SaaS subscriptions, or the salary of the engineer using the tools. Those belong to Human Capital.
The denominator is messier than it looks. Human Capital is not just salary. It is fully loaded cost: salary, employer taxes, benefits, equipment, real estate, recruiting, training, and internal tools that exist mainly because people exist. A useful approximation is 1.4x to 1.6x the gross salary line. Without that grossing-up, the ratio flatters Token Capital, because the dollar spent on a contract worker and the dollar spent on a benefited employee buy very different amounts of work.
One important honesty constraint: contractor spend is Human Capital, not Token Capital, even when the contractor is offshore and inexpensive. The ratio is not a measure of cheap-versus-expensive labor. It is a measure of inference-versus-human labor. Confusing the two collapses the analytical value of the metric in a single move.
the substitution curve
For any given knowledge task, there is a curve. On one axis, the share of the task done by tokens. On the other, the total cost of completing the task at acceptable quality. For low-stakes synthesis, the curve drops sharply: you can reach most of the answer with mostly tokens and a thin layer of human review, at a tenth of the cost of doing it manually. For high-stakes work (debugging a production incident at scale, drafting a legal argument that has to survive a court, advising a customer through a real crisis) the curve flattens or even rises, because the marginal review effort grows faster than the marginal tokens save.
The TC/HC ratio is, in effect, an average across many of these curves, weighted by the mix of work an entity actually does. A company doing mostly first-draft synthesis will look like it should run a very high TC/HC. A company doing mostly high-stakes judgment will rationally run a very low TC/HC, and that is not a failure of adoption. It is a reflection of the curve.
This is the first thing the ratio reveals: it makes the curve legible at the entity level. A team with a low ratio in a domain where the curve is favorable to tokens is leaving productivity on the table. A team with a high ratio in a domain where the curve is hostile to tokens is shipping fragile work. Neither conclusion is automatic from the spend numbers, but the ratio is what makes them debatable in the first place.
the ratio at four scales
At the personal scale, the ratio is between an individual's AI tooling spend and a synthetic version of their own loaded cost. A senior engineer on a $300k loaded salary spending $400 a month on AI subscriptions and a few hundred more on API tokens runs a personal TC/HC of roughly 3 percent. A year earlier it was closer to 0.5 percent. The trajectory is more interesting than the level.
At the team scale, the ratio aggregates across people but stays narrowly observable. A 12-person engineering team with $40k of annual AI spend and $4M of loaded cost runs at 1 percent. Add an agentic workflow that consumes meaningful tokens, and the ratio can double in a quarter without any change in headcount.
At the company scale, the ratio is what finance teams will start surfacing in board decks within the next twelve months, because it is the most concise way to describe what is changing about operations. A SaaS company at 0.5 percent TC/HC last year and 3 percent this year has not just adopted AI. It has restructured how a unit of output is produced.
At the economy scale, the ratio is a leading indicator for labor substitution. National accounts do not yet break out token spending as a category, but the underlying flows are visible in the published revenues of OpenAI, Anthropic, Google, and the inference providers, set against payroll data. As that ratio rises across economies, the marginal worker in cognitive industries is competing more directly with inference than with another worker.
the trajectory
The thing that gives the ratio analytical bite is that its numerator and denominator are moving in different directions, on different time scales, and at different speeds. The denominator, human cost, is sticky and slowly rising. Salaries do not fall. Benefits compound. Real estate is reluctant. The denominator changes by a few percent a year in either direction, dominated by mix shifts and macro cycles.
The numerator, token cost, is collapsing. The per-token price of frontier inference has fallen by more than an order of magnitude in three years, and the trajectory has been remarkably consistent. Even when prices for a specific named model rise, the cost per equivalent unit of useful inference is dropping, because each generation of model produces more usable work per token. The price you should compare across years is not dollars per million tokens. It is dollars per unit of completed work, which has fallen faster than the headline price.
The result is that if usage stays constant, the ratio falls toward zero. The reason the ratio is in fact rising for most organizations is that usage is growing faster than per-unit prices are falling. People are buying more inference at lower prices, and the dollar total is climbing. This is the part of the ratio that requires intellectual honesty to interpret. A rising TC/HC could mean adoption is succeeding (more useful work being routed to inference) or it could mean spend discipline is failing (more wasted tokens chasing the same outcomes). The headline number cannot distinguish between the two.
what the ratio actually reveals
Used carefully, four things become visible.
First, the substitution surface. By breaking TC/HC down by function (engineering, support, sales, research, operations) you can see which functions have a curve that favors tokens, and which do not. The ratio per function diverges quickly, even within the same company. Engineering and support tend to lead; sales and research lag for different reasons; legal and finance lag for the same reason (high stakes per output, low tolerance for hallucination).
Second, the leverage you are buying with each marginal dollar. If TC grows by 50 percent and output grows by 50 percent, the leverage is one to one, and the new spend is paying for itself but nothing more. If output grows by 100 percent, the leverage is two to one, and the inference budget is doing real work. The ratio paired with an output measure is what tells you whether the spend is productive or merely fashionable.
Third, the durability of the operating model. An organization at 5 percent TC/HC with a deeply integrated agentic workflow has restructured how work is done. An organization at 5 percent TC/HC because everyone is paying personal subscriptions out of pocket and using them like search engines has not restructured anything. The ratio is the same; the structural meaning is opposite. The breakdown by procurement path (centralized API spend, departmental seats, personal subscriptions expensed) is what separates the two.
Fourth, the velocity of change. Quarter-over-quarter movement in TC/HC, especially when normalized for headcount, is one of the cleaner signals of whether an organization is actually changing how it produces work, or only saying it is. Slides claim transformation. The ratio reports on it.
what the ratio cannot measure
The honesty section, and the one easiest to skip.
The ratio does not measure quality. A team can spend more on tokens and produce worse work. The dollar number is silent on the truth-value of what the tokens generated. Without an output measure paired with it, TC/HC is closer to a vanity number than a metric.
The ratio does not measure complementarity. Most useful AI deployment is not substitution but amplification: the human does work they could not do alone, faster than they could have done it. The ratio treats every dollar of token spend as displacing a dollar of human spend, when the more common reality is that the token spend lifts the value of the human spend it sits next to.
The ratio does not measure decision rights. In any organization that takes accountability seriously, certain decisions cannot be delegated to inference, and the question of who signs the decision matters more than the cost of producing it. A 30 percent TC/HC in a business that requires human accountability for every customer outcome is a 30 percent leverage number with a 100 percent accountability number sitting next to it.
The ratio does not measure cycle time. An organization that spends one dollar on inference and gets the answer in six seconds is fundamentally different from one that spends one dollar on a human and gets the answer in six days. The compression of time is, for many decisions, the actual value, and the dollar ratio is silent on it.
Finally, the ratio's denominator is sticky in a way that hides what is changing. If an organization holds payroll flat and grows TC by a factor of five, the ratio rises by a factor of five. If the same organization quietly reduces headcount by 20 percent at the same time, the ratio rises faster, but the change in operating model is hiding in the denominator, not the numerator. The headline number obscures which side of the equation is moving, and that is exactly the thing a reader needs to know.
does this metric make sense
So is TC/HC worth tracking, or is it the kind of clever ratio that looks profound on a slide and dissolves under scrutiny?
The case for the metric is straightforward. It is dimensionless, comparable across entities, computable from data finance teams already have, and pointed at a transition that is unambiguously happening. The substitution of human cognitive labor with machine inference is the defining operational change of this decade in knowledge work, and any number that lets you watch the substitution happen at the team, company, or economy scale has prima facie value. A finance leader who cannot put a number on this transition is flying blind through it.
The case against the metric is also strong. It is a ratio of inputs, not outputs, which puts it in the same epistemic class as "marketing spend over revenue": informative on margin but easy to game and easy to misread. The numerator is moving so fast (token prices falling, products bundling, model generations compressing useful work into fewer tokens) that the unit is unstable across quarters. The denominator is sticky and politically loaded; changing the denominator usually means changing the headcount, which is rarely a metric question and almost always a humans question. And the metric is mute on the things that matter most: quality, complementarity, accountability, cycle time.
On the analysis above, the ratio makes sense under three specific conditions, and is misleading outside them.
It makes sense when it is paired with an output measure on the same time horizon. TC/HC alone is decorative. TC/HC alongside revenue per employee, story points shipped, tickets resolved, deals closed, or whatever the unit of work is for that organization, becomes a leverage measure. The pairing is what gives it analytic weight.
It makes sense when it is decomposed by function. Aggregate TC/HC across an organization averages out curves that are wildly different per function. A research team and a sales team and an engineering team have entirely different substitution surfaces, and the aggregate ratio collapses that variance into a single number that means very little. By-function ratios mean a lot.
It makes sense as a velocity measure rather than a level measure. The interesting question is not "is your TC/HC 1 percent or 5 percent today". The level depends on industry, on function mix, and on the rate at which token prices are falling under you. The interesting question is "how fast did your TC/HC change in the last four quarters, holding output flat or rising". That is the question that distinguishes a real operational shift from a procurement spike.
Outside those three conditions, the metric is the kind of number that flatters whoever made the slide. Inside them, it is one of the few clean handles we currently have on the most important shift in the cost structure of knowledge work since the introduction of the personal computer.
The shorter answer, for the people who want one: the ratio is worth computing, the level is not worth defending, the trajectory is worth fighting about, and the assumption that the ratio alone explains anything is worth resisting.