AI is beginning to create a new corporate resource alongside labor, software, computing infrastructure and capital: machine intelligence itself. As companies start measuring, budgeting and potentially rationing AI usage, the amount and quality of intelligence available to individual workers could become an increasingly important factor in productivity, competition and workplace structure.
THE SIGNAL
For most of modern business history, companies have allocated a familiar set of resources.
People receive salaries.
Departments receive budgets.
Employees receive computers, software licenses and expense accounts.
Technology teams receive servers, cloud capacity, storage and computing power.
Artificial intelligence is introducing something different.
Companies are beginning to allocate machine intelligence.
The earliest evidence is appearing in software development, where employees increasingly rely on AI coding systems that consume tokens every time they analyze code, reason through a problem, generate software or operate autonomously.
Uber provides one of the clearest examples.
After encouraging engineers to aggressively adopt AI coding tools, the company burned through its entire 2026 AI coding budget within the first four months of the year. Uber subsequently began reassessing both the cost of those tools and how much measurable value the additional usage was producing. (Fortune)
Meta is thinking even further ahead.
Instagram head Adam Mosseri has publicly suggested that companies may eventually place AI-token spending caps on individual engineers. He compared token allocation to familiar corporate decisions involving payroll, GPUs, storage and operating budgets and suggested that employees trusted to produce stronger returns could potentially justify larger allocations. (TechCrunch)
Meanwhile, Accenture has reportedly been confronting employees consuming significant AI resources on relatively basic tasks, while Google Cloud is explicitly teaching developers how to reduce unnecessary token consumption and warns that uncontrolled AI loops can consume entire token budgets. (TechCrunch)
Individually, these may look like cost-management stories.
Together, they point toward something larger.
Machine intelligence is starting to become a managed corporate resource.
And once intelligence becomes something that can be purchased, measured and distributed, companies will eventually have to decide who gets how much of it.
WHAT THE MARKET IS MISSING
The immediate conversation is focused on AI costs.
That may be too narrow.
Tokens are simply today’s measurement system.
Models will become more efficient. Token prices may decline. Companies may increasingly run models themselves. New pricing structures could replace today’s token-based billing altogether.
But the underlying economic issue does not disappear.
AI inference requires computing resources.
More capable models generally require greater resources than lightweight models. Long-running autonomous agents can consume more capacity than simple questions. Deep research, software development, scientific analysis and complex reasoning can require significantly more machine intelligence than routine tasks.
That means companies increasingly face a question they have rarely had to answer before:
How much intelligence should we allocate to each worker?
Consider two employees performing the same job.
One receives access to a lightweight AI assistant with restrictive usage limits.
The other receives continual access to the company’s most capable reasoning model, autonomous agents, enormous context windows and substantial computational resources.
They may technically hold the same position.
But they no longer possess the same productive capacity.
That distinction could become increasingly important.
The difference between workers may eventually be determined not only by education, experience and individual ability — but also by the amount and quality of machine intelligence placed behind them.
That would transform AI from another software application into something much closer to productive capital.
Factories allocate machinery.
Airlines allocate aircraft.
Financial firms allocate investment capital.
Cloud companies allocate computing capacity.
The AI-powered enterprise may increasingly allocate reasoning capacity.
FIRST-ORDER EFFECTS
The first effects are already becoming visible.
AI budgets become formal operating expenses
The experimental stage of giving employees broad access to AI is beginning to collide with financial reality.
Companies can no longer assume that more AI usage automatically produces greater productivity.
Uber’s experience illustrates the problem clearly: rising use of coding assistants did not automatically provide management with a simple measurement showing a corresponding increase in useful consumer features. (Fortune)
That means CFOs, CIOs and department heads increasingly need to ask:
What are we receiving for the AI resources we’re buying?
Employees may receive different AI allowances
A company does not give every department identical budgets today.
There is little reason to assume every employee will eventually receive identical AI capacity either.
A senior engineer working on a critical product could justify considerably greater AI spending than someone performing routine administrative work.
A researcher could require access to expensive reasoning models.
A customer-service representative might use a cheaper model optimized for repetitive interactions.
Companies may begin matching AI capacity to economic value, job complexity and expected return.
AI efficiency becomes an employee skill
Workers may eventually be evaluated not simply on whether they use AI but how efficiently they use it.
Google Cloud’s recent token-efficiency guidance is an early indication of this transition. Developers are being encouraged to choose models based on task complexity, reduce unnecessary context and prevent autonomous systems from entering costly loops. (Google Cloud)
The best AI worker may therefore not be the person consuming the most intelligence.
It could be the person generating the greatest value per unit of machine intelligence consumed.
SECOND-ORDER EFFECTS
This is where the signal becomes much larger.
Machine intelligence could become part of compensation
Companies already compete for talent using salaries, bonuses, equity, offices, technology and benefits.
Eventually, access to premium AI could join that list.
Imagine two competing firms recruiting the same engineer.
Company A provides limited access to standard AI models.
Company B provides effectively unrestricted access to frontier reasoning models, specialized agents, enormous context capacity and dedicated computing resources.
Even with identical salaries, those may no longer be equivalent jobs.
AI access could become part of the productive environment employers use to attract top talent.
Corporate inequality could become intelligence inequality
Organizations already contain unequal access to resources.
Executives receive larger discretionary budgets.
Top salespeople receive more valuable accounts.
High-priority projects receive greater capital.
AI could create another hierarchy.
Certain departments could receive frontier models.
Others receive inexpensive models.
Some workers could operate multiple autonomous agents simultaneously.
Others may be restricted to limited assistants.
The result could be an internal intelligence hierarchy in which access to machine reasoning becomes partly determined by organizational importance.
That isn’t necessarily unfair or inefficient. Companies routinely allocate scarce resources according to expected returns.
But it would represent a major change in how workplace capability is distributed.
Productivity measurements could change
Today’s productivity metrics largely measure human output.
AI complicates that equation.
If Employee A completes ten projects using $20,000 of AI resources while Employee B completes eight using $2,000, who is more productive?
The answer depends on value, quality, speed and cost.
Companies may eventually develop metrics resembling:
Revenue per AI dollar.
Output per million tokens.
Engineering productivity per unit of inference.
Cost per completed autonomous task.
This could create an entirely new category of corporate performance measurement.
Human availability may stop determining productive capacity
One of the strangest consequences appears when an employee still has working hours available but has exhausted the AI capacity needed for the workflow.
The person is still available.
The salary is still being paid.
But part of the worker’s productive system has temporarily disappeared.
This challenges a workplace model built around the assumption that available human hours equal available productive capacity.
As AI becomes more integrated into work, that assumption becomes weaker.
A future worker may effectively consist of:
Human judgment + skills + time + machine intelligence.
Remove one component and productivity may change substantially.
WINNERS
Workers who learn to leverage AI efficiently
People who can produce significant business value without unnecessarily consuming expensive AI resources could become particularly valuable.
Prompting alone will not be enough.
Workers may need to understand model selection, context management, agent orchestration, verification and when not to use an expensive AI system.
Companies that measure AI by outcomes
Organizations capable of connecting AI expenditures to revenue, productivity, quality or strategic value will have an advantage over companies simply chasing higher usage statistics.
The objective will shift from:
“Use more AI.”
to:
“Generate more value from AI.”
AI optimization platforms
A new software layer could emerge around managing corporate intelligence consumption.
These systems could route simple tasks to inexpensive models, reserve frontier models for difficult work, monitor departmental spending, identify waste and calculate return on AI expenditure.
In effect, companies may need an AI resource-management layer in the same way they eventually needed cloud-cost management.
Smaller and more efficient AI models
The most powerful model will not always be economically optimal.
Amazon CTO Werner Vogels has already described companies showing greater interest in cheaper open-source alternatives as AI expenses increase. (Fortune)
A sufficiently capable model that costs dramatically less could win enormous portions of enterprise workloads.
LOSERS
Companies that confuse usage with productivity
Early AI adoption has sometimes rewarded employees simply for using more AI.
That incentive becomes dangerous when usage itself carries substantial cost.
An employee can consume enormous numbers of tokens without generating equivalent economic value.
Token consumption is activity. It is not automatically productivity.
Workers who become completely dependent on unlimited AI
AI can increase productivity dramatically, but excessive dependence creates another risk.
If a worker becomes unable to perform meaningful work when a model is unavailable, usage limits are reached or systems go offline, the organization has created a new operational dependency.
Companies will need to decide how much human capability should remain independent of AI assistance.
Premium models without sufficient differentiation
If companies become sophisticated about routing tasks according to cost, expensive frontier models may need to continuously demonstrate why a particular workload requires them.
Routine corporate work could migrate toward cheaper models while premium models concentrate on problems where superior reasoning creates measurable value.
Traditional productivity management
Managers accustomed to measuring attendance, hours or visible activity could struggle.
A worker commanding powerful AI agents may accomplish in two hours what previously required an entire day.
Another worker may consume enormous AI resources during eight hours while creating comparatively little value.
AI makes time spent working an increasingly incomplete measure of productive contribution.
WHAT HAPPENS NEXT
The most important thing to watch is whether today’s scattered token restrictions become formal corporate policy.
The progression could look something like this:
Stage 1 — Unlimited experimentation
Companies encourage employees to use AI aggressively.
Stage 2 — Cost shock
Usage increases faster than expected and AI bills become material.
Stage 3 — Measurement
Organizations begin tracking employees, teams, models and workloads.
Stage 4 — Allocation
Departments and individuals receive AI budgets based on role, priority and expected return.
Stage 5 — Optimization
Software automatically routes work toward different models based on cost, capability, security and importance.
Stage 6 — Intelligence accounting
Companies begin measuring how effectively machine intelligence is converted into economic output.
Pieces of this progression are already visible.
Uber has experienced the cost shock.
Accenture has confronted usage management.
Meta is discussing eventual allocation.
Google is promoting optimization.
And Microsoft’s decision to pull back many internal Claude Code licenses while consolidating developers around its own tooling illustrates how quickly companies can change AI access when economics and strategic priorities shift. (TechCrunch)
The critical question is whether this remains concentrated in software development.
It probably won’t.
AI agents are moving into finance, marketing, research, law, medicine, customer service, engineering and operations.
As these systems become capable of performing longer and increasingly autonomous tasks, AI consumption could become material across large portions of the workforce.
That’s when token budgeting stops being an engineering issue.
It becomes a management issue.
BOTTOM LINE
The token is probably temporary.
The economic concept behind it is not.
Companies are beginning to discover that machine intelligence is neither unlimited nor free.
Once AI becomes deeply embedded in everyday work, organizations will have to decide how much reasoning capacity to purchase, which models to use, where that intelligence should be deployed and which employees or departments can generate the greatest return from it.
That could eventually make access to AI as strategically important as access to capital, technology and talent.
The workplace has traditionally been built around allocating human labor.
The next era may require companies to allocate something else alongside it:
machine intelligence.
And when that happens, one of the most important questions inside a company may no longer simply be:
How many employees do we have?
It may become:
How much intelligence can we put behind each one?
