When Labor Becomes Tokens: How AI Is Changing the Economics of Running a Company

Tokens are the new form of labor, a new way to calculate the opex of a company?

The other day, I was using Perplexity Computer, a new product feature that was rolled out. Basically, it’s Perplexity’s version of an AI agent, where it can draft and send emails,  run commands in parallel, create sub agents, and execute on tasks. Each task that was finished, my credits started to dwindle, and it made me think of something, which is that labor, in the future, is simply tokens.

For just this week, I’m enabling my readers to get access to Perplexity Enterprise Pro which contains Perplexity Computer by simply pre-ordering my book. Just send me a screenshot of order form by replying to this email.

Tokens as part of compensation?

It’s no wonder that Jensen Huang, in his recent GTC NVIDIA keynote talked about handing out tokens to engineers as a form of compensation. Jensen Huang is proposing to provide half an engineer’s salary to become tokens as an attractive hiring mechanism. So far, I don’t know of any company that has implemented this, but I’ll be watching and observing to see if we will see this in the near future.

Imagine that for a bit, a marketing person is faced with two offers from two different companies. One company says, we’ll pay you $150K + 1 billion tokens. The other company says we’ll pay you $180K with no tokens. Which job do you take? It’s kind of like saying to a marketing person, in one job we’ll pay you a little bit, then give you a massive marketing budget. Do what you want with it as long as it provides the largest ROI. In the other, it’s like saying we’ll pay you higher than market rate, with no budget. You’ll have to find your own way of creating output. Before you make a decision to pick one or another, I think there are other things to consider.

Because it’s also not as simple as that, there are a lot of other implications to consider. What model are they giving you? How much autonomy do YOU have with using them? Can you run autonomous agents? Can you build internal tools? Can you launch external products? Can you use tokens for side projects? Being given a lot of tokens to play with is essentially like saying we’re not just hiring you as an employee but you + whoever other “employee” you wish to bring in and manage. It looks very differently from the old way of hiring, which is hiring one person to do one job. If your role is very output driven, it might be a better idea to take the job that promises you with a certain number of tokens.

Compensation is moving from

Pay = salary

to:

Pay = salary + compute power

In the future, the best employees won’t just ask, how much do I get paid? They would ask, how much intelligence can I deploy? Because that determines how big they can play. Why this changes how we calculate cost in a company 

From fixed cost to variable cost

So if labor does become tokens, from a company’s financial perspective, it’s like changing a fixed cost to a meter, the more you use, the more you pay. A CFO could predict next year’s cost fairly easily because salaries are stable. AI shifts this toward something closer to cloud economics. So finance models start to look more like AWS billing models.

When a company hires an employee, it commits to a relatively stable annual expense that does not fluctuate with output. For example, an employee earning $120,000 in salary, with an additional $30,000 in benefits and taxes, represents a total cost of about $150,000 per year.

What makes this cost structure unique is that it remains largely unchanged regardless of productivity. Whether that employee writes ten lines of code or ten thousand, or handles twenty customer tickets or two thousand, the company still pays the same amount. As a result, the real economic question becomes one of utilization. Companies must ask how much output they are getting relative to that fixed cost.

This leads to a simple but important metric: cost per task. If a support agent handles 20,000 tickets per year, then the effective cost per ticket is calculated by dividing the total annual cost by the number of tickets handled. In this case, $150,000 divided by 20,000 results in a cost of $7.50 per ticket. This becomes the true unit economics of human labor.

AI operates on a fundamentally different model. Instead of paying for time, companies pay for usage. Costs are driven by tokens processed, compute consumed, and API calls made. Each task has a direct, measurable cost associated with it.

For instance, a typical AI-powered customer support interaction might involve 1,000 tokens to process the input and 500 tokens to generate a response, for a total of 1,500 tokens. If the model costs $0.005 per 1,000 tokens, the total cost of that interaction is approximately $0.0075, less than a cent. Compared to the $7.50 cost of a human-handled ticket, this represents a difference of several orders of magnitude.

The goal isn’t to replace all humans with AI

From a CFO’s perspective, “cost” might have transferred on the balance sheet. Whereas in the old world we pegged human labor and headcount as a fixed cost, in the new world, labor becomes a variable cost. I don’t think the CFO is trying to automate everything. Instead, they are trying to figure out what is the optimal balance between the two. Some tasks are 100% automatable, but some others, we will need the human to provide oversight. And even some other ones, we will need humans mostly to complete the task.

To think about this more clearly, a useful way to think about this is through a simple framework that maps work along two dimensions: cost structure and task complexity.

On one axis, you have cost structure, which ranges from fixed costs like salaries to variable costs like compute and tokens. On the other axis, you have task complexity and risk, ranging from low to high.

This can create four distinct zones that every CFO must now manage.

In the bottom left are low complexity tasks with fixed costs, what we traditionally think of as labor. These are repetitive, process-driven activities like basic customer support, data entry, and routine reporting. Historically, companies staffed these with large teams. The work is predictable, but it is not very efficient at scale because costs remain fixed regardless of output. This quadrant is shrinking and gradually giving way to its sister quadrant to the right.

In the bottom right are low complexity tasks with variable costs, which is where AI automation lives. These are the same types of repetitive tasks, but now handled by systems that operate on tokens and compute. The key advantage here is that costs scale with usage, and the marginal cost per task becomes extremely low. This is where companies can unlock significant efficiency gains, and it is typically the first place CFOs target for automation.

In the top left are high complexity tasks with fixed costs, which represent traditional human expertise. These are roles that require judgment, accountability, and specialized knowledge, such as strategy, legal work, and high-stakes decision-making. These are the hardest to automate.

Finally, in the top right is the most important quadrant: high complexity tasks with variable costs, where AI and humans work together. This is the domain of orchestration. Here, AI systems handle large portions of the work, while humans guide, supervise, and make critical decisions. These workflows combine scalability with judgment, making them the highest leverage part of the organization.

From a financial perspective, this shift changes how value is created. It is no longer about how many people you have or how many hours they work. It is about how effectively you allocate compute, human judgment, and workflow design across these four zones.

The companies that get this right will not just reduce costs. They will fundamentally increase their ability to scale, adapt, and produce value.

Measuring Value Per Token

As AI becomes integrated into operations, companies may begin tracking new kinds of performance metrics. I think everything will become more granular. CFOs and managers will go in depth in examining how much does it really cost for one task to be completed, and whether it’s a better to have AI do it, a human do it, or a blend of both.

Instead of focusing only on revenue per employee, businesses might examine metrics such as:

cost per automated task

tokens required per workflow

revenue generated per token consumed

These metrics treat intelligence as a measurable input into production. A company that generates thousands of dollars of value from a small number of tokens is operating efficiently. A company that consumes massive amounts of compute to produce little value is not. Over time, businesses will learn to optimize these ratios just as they optimize supply chains and manufacturing processes today.