Token spend per employee at top firms
Humans are now cheaper than software: token spend per employee at top firms has overtaken the cost of the humans themselves.
For the first time in history, humans are cheaper than software — and AI is creating more jobs than it eliminates. George Sivulka argues that agent workforces and human workforces fail in the same way, and that seven parallels between them unlock the next trillion dollars of AI value creation: management.
The essay argues AI was supposed to replace human labor and did the opposite: we just gave every employee, even the worst ones, effectively unlimited headcount and budget — and managing AI is harder than managing people, because AI scales dysfunction instantly.
Technology has always solved one problem by creating another. Managing AI is harder than managing people, but we can learn from the past: agent workforces and human workforces fail in the same way. Tokens behave like a workforce — accurate only when prompted correctly, fast but meaningless across 100 retries, empire-building in token spend, dying between model releases, failing confidently with perfect formatting. The one place AI really beats humans is scalability, which is exactly why mismanaging tokens is so expensive — and why you must find and scale the 100X token. It's time to manage.
Humans are now cheaper than software: token spend per employee at top firms has overtaken the cost of the humans themselves.
AI is creating more jobs than it eliminates — headcount grows after AI adoption rather than shrinking.
The essay's historical anchor: the 1830s railroad buildout broke coordination before management fixed it.
In the 1830s, the railroad drove one of the largest infrastructure buildouts the world had ever seen — American track mileage 120Xed in a decade. Then the system broke.
On October 5, 1841, two trains fatally collided on the Western Railroad in Massachusetts due to a simple coordination failure — individual conductors were no longer enough to keep train travel safe.
Railroads hired managers per geography, defined roles in writing, and established hierarchies with reporting lines. Modern management was born — and rail became among the world's first billion-dollar industries, at its peak roughly 60% of the stock market.
Seven structural parallels between token workforces and human workforces — understanding them will unlock the next trillion dollars of AI value creation.
The tokenmaxxing hype cycle ran its full course in under a month, but token spend was never the real problem: people spend so much on tokens because they don't know how to use them. Maybe 1 in 100 employees knows how to give AI context; give an agent harness to the other 99 and they will produce loops.
In Claude Code/Cowork, Copilot, Karpathy's Autoresearch, or any harness, loops are a bandaid for the fact that almost nobody can successfully prompt. Agents call themselves to fix themselves only because a human never articulated the task cleanly — you are spending tokens on spending tokens.
Most companies are mismanaged: the vast majority of workers are cogs stamping approvals and hiring more cogs. Elon cut 80% of X's staff and the company performed better. Just like 80% of employees do nothing, 80% of tokens today do nothing. People create more people; tokens create more tokens. Looping is the new empire building.
Seen as employees, the promises of AI break down — but AI's real edge is scalability, which is why mismanaging tokens is so expensive. There exist tokens that give you 100X as much leverage. Humans are cheaper than tokens on average, but good tokens are cheaper at scale. Management converts one into the other.
Employees don't want to teach AI their secret sauce. At Meta, stock-owning employees are outraged that employee context is training data. Tribal knowledge has been job security for centuries — medieval guilds kept methods secret — and AI is the first technology that asks workers to hand it all over at once. Nobody trains their replacement for free.
Manage tokens the way you manage humans: define what good looks like. Coding escaped politics because it has built-in evals — code runs or it doesn't — hence 99% of AI revenue. Just as OKRs leverage a human workforce, evals leverage an infinitely scalable token workforce. A firm's eval suite becomes its most valuable resource.
Nobody has AI working reliably yet. Neofirms bet against incumbents for the $21T of services spend, but the greatest AI assets sit inside incumbents. AI transformation companies — selling evals, token minimization, and deep business programming as a net-new service — will be 10X larger than any neofirm, powered by a Jevons paradox: every use case adopted surfaces ten more. Palantir was never selling software; it was selling transformation.
Two verticals ground the essay's economics, modeled with the reused corpus Industry class.
Labor TAM: $21 trillion services spend across the knowledge economy.
Automation readiness: Low today — nobody has AI working reliably yet; unlocked by firm-specific evals.
NAICS: 5416
Peak scale: Roughly 60% of the stock market — among the world's first billion-dollar industries.
Historical precedent: Coordination crisis solved by inventing modern management.
NAICS: 482111
Token spend per employee at top firms has risen to the point where, for the first time in history, humans cost less than the AI software working alongside them — and AI is creating more jobs than it eliminates rather than replacing human labor.
As railroad complexity outgrew individual conductors, the fatal October 5, 1841 collision in Massachusetts forced railroads to invent modern management — geographic managers, written roles, reporting hierarchies. AI is breaking the coordination system again, and the same managerial response is required.
Tokenmaxxing — maximizing token spend — ran its hype cycle in under a month because token volume was never the problem. People spend heavily on tokens because they don't know how to use them; maybe 1 in 100 employees knows how to give AI context.
In any harness — Claude Code/Cowork, Copilot, Karpathy's Autoresearch — agents call themselves to fix themselves only because a human never articulated the task cleanly. Loops are brute-force compensation for failed prompting: spending tokens on spending tokens, just as meetings about meetings compensate for unclear human coordination.
The essay draws the parallel with human organizations: just as roughly 80% of employees do nothing that meaningfully impacts the business — Elon cut 80% of X's staff and the company performed better — 80% of tokens today do nothing. Looping is the new empire building.
For any given job there exists token context that cuts AI effort by orders of magnitude — 100X leverage — just as a handful of employees make others 10X as productive. The 10X engineer built the last era of companies; the 100X token will build the next.
Tokens are more accurate than humans — but only when prompted correctly. Faster — yet speed means nothing across 100 retries. Apolitical — but they build empires of token spend. They don't quit — but die between model releases and sessions. Trustworthy — yet they fail confidently with perfect formatting.
Employees withhold their secret sauce from AI because these systems aren't only there to 'help them' — nobody trains their replacement for free. Even at Meta, stock-owning employees are outraged that employee context becomes training data. Tribal knowledge has been job security for centuries, and AI asks workers to hand it all over at once.
Coding has built-in evals: code runs or it doesn't. That is why 99% of AI revenue today is coding. Broader cross-domain AI use cases will only come online when someone builds the requisite evals — turning fuzzy human processes into code.
Just as OKRs leverage a human workforce to optimal output, evals leverage an infinitely scalable token workforce. No two firms will have the same eval set; a firm's eval suite becomes its most valuable resource, and an organization running generic evals or generic agents has no edge.
Neofirms bet against incumbents to capture the 21 trillion dollars of services spend, but the greatest AI assets — differentiated processes that already work, with existing distribution — sit inside incumbents. Transformation companies sell a net-new service to those players, and a Jevons paradox makes transformation ongoing: every adopted use case surfaces ten more. The essay predicts they will be 10X larger than any neofirm.
On paper Palantir is the most Claude-disruptible company in software — hand-building bespoke enterprise applications — and by SaaS logic $PLTR should be a zero before $NOW. It isn't, because Palantir was never selling software; it was selling transformation. Today the real work has evolved to evals, token minimization, and understanding a business deeply enough to program it.
Ten terms wrapped in a schema:DefinedTermSet in the companion RDF, including the reused Agent Harness term from the sibling AI value-capture response.
Maximizing token spend as a substitute for knowing how to use tokens — the agent-workforce equivalent of throwing bodies at the problem. Its hype cycle ran its full course in under a month.
An agent calling itself to fix itself because a human never articulated the task cleanly — a brute-force bandaid for failed prompting; meetings about meetings for the token workforce.
For any given job, some token context cuts AI effort by orders of magnitude — the token-workforce successor to the 10X engineer. Good tokens are cheaper at scale; management converts humans into them.
Employees withholding their secret sauce from AI systems as a job-security tactic — the modern form of guild secrecy. Nobody trains their replacement for free.
A firm-specific, executable definition of what good looks like for an AI task — the mechanism behind coding's breakout and the path to running 100X tokens. A firm's eval suite will become its most valuable resource.
An 'AI Native Services' startup funded to capture the $21 trillion of knowledge-economy services spend, on the theory that incumbents mired in politics and processes will never manage the AI transition themselves.
A company selling ongoing AI transformation — evals, token minimization, encoding a firm's nuances into agents — as a net-new service to incumbents; predicted to be 10X larger than any neofirm.
The operating shell in which AI agents run — Claude Code/Cowork, Copilot, Karpathy's Autoresearch. Reused from the sibling AI value-capture response KG rather than re-minted.
Efficiency gains that increase rather than decrease total consumption — every AI use case an organization adopts surfaces ten more, so the more AI-enabled a firm becomes, the more transformation it consumes.
Unwritten know-how held by workers as job security for centuries — medieval guilds kept their methods secret. AI is the first technology that asks workers to hand all of it over at once.
Map where intelligence is being spent and wasted across the organization — the managerial survey the essay calls for now that infrastructure and services are sufficient.
Identify the token context that cuts AI effort by orders of magnitude for each job, the way great managers identify the employees who make others 10X as productive.
Treat agent loops like headcount bloat: often it's more efficient to cut the loop than to fund it. Record the loops that work; kill the ones that only spend tokens on spending tokens.
Close the 1-in-100 gap: cultivate people who can articulate a process clearly and empathize with a polluted context window, instead of handing harnesses to the other 99 and harvesting loops.
Define what good looks like by turning fuzzy human processes into code — expressing the qualitative as quantitative. Specific evals matter more than prompting lessons or chat harnesses.
Recognize that firms are wired — emotionally, structurally, politically — to reject the technology most important to their future, and that the holders of 100X tokens have the least incentive to surrender them.
Treat transformation as continuous, not a one-off project: a Jevons paradox means every adopted use case surfaces ten more, and encoding the firm's nuances into agents may become the largest economic task of the decade ahead.
Graph data is derived from the generated RDF entity and relationship model. Node and edge clicks use URIBurner resolver-backed IRIs.
Explore people, organizations, the seven parallels, industry verticals, FAQ, glossary, and HowTo structure.
Graph data embedded from companion RDF at generation time (54 nodes / 68 links). Controls tray is closed by default.