The latest twist in the AI job replacement debate is not that machines are coming for everyone’s work. It is that, in a growing number of cases, the machines may not be cheaper. That is an awkward development for some. For the past two years, many executives have been encouraged to imagine digital labor as a form of near-frictionless substitution: fewer people, lower costs, faster output. Replace the call-center agent. Replace the analyst. Replace the junior engineer. Replace the back office. But the economics are becoming more complicated.
AI agents, coding assistants, and increasingly robotics do not simply remove labor costs. They introduce a new, highly variable, and sometimes explosive cost line: tokens, compute, orchestration, monitoring, tooling, risk controls, and infrastructure. In other words, digital labor is not free labor. In some cases, it is not even cheap labor. And that is precisely why leaders need to understand it more carefully. Their real value is not that they do the same work for less money. It is that they make entirely new kinds of work possible.
If this is news to you — and you were planning to pay for your AI strategy by firing half your workforce — you may be in for a rude shock.
The first concept leaders need to master is tokenomics. In the AI context, a token is a unit of information processed by a model. Every prompt, document, tool call, retrieved record, system instruction, chat history, intermediate reasoning step, and model response consumes tokens. In a simple chatbot interaction, that may not matter much. In an agentic workflow, where a system plans, retries, calls tools, reads files, checks work, invokes other agents, and loops through a task repeatedly, the economics change quickly.
What began as a technical metering device is now becoming a financial discipline.
Tokenomics is the practice of understanding, forecasting, governing, and optimizing the cost of AI work. It asks a deceptively simple question: how many units of machine cognition are we buying, and what business value are we getting in return?
Tokenomics is likely to become a core part of corporate FinOps. The pattern is familiar. SaaS began as a way for business teams to move faster without waiting for IT. Over time, as subscription sprawl, cloud consumption, and infrastructure bills rose, finance moved in.
Procurement teams scrutinized renewals. CFOs demanded usage data. FinOps emerged to create financial accountability across engineering, finance, product, and business teams. In some organizations, even technology leadership was pulled closer to finance, with CTOs and CIOs increasingly expected to justify architecture decisions not only in terms of performance and resilience, but also unit economics, utilization, and cost discipline.
AI agents will follow the same path, only faster. Token usage, model selection, inference costs, tool calls, data retrieval, latency requirements, and agentic retries will all become objects of financial control. AI spending is too technical to be left only to finance, too financially volatile to be left only to engineering, and too strategically important to be left to enthusiasm. But token governance is only the visible edge of the problem. The deeper issue is structural: companies are about to discover that digital labor is not a software feature they can simply turn on, but a new operating layer that changes infrastructure strategy, workflow design, management accountability, and the economics of work itself.
Deloitte has already warned that traditional total-cost-of-ownership models need to be refreshed for AI because tokens have become a primary unit of value and spend. Its recent work on AI token economics argues that CFOs need to connect token consumption to the P&L, model usage inflection points, and govern AI with the same rigor they apply to capital allocation. Deloitte’s survey data suggests that many companies are already generating more than 10 billion AI tokens per month, while the share expecting to exceed 100 billion tokens per month is projected to triple between 2025 and 2028. Goldman Sachs Research is even more dramatic: it expects agentic AI to drive a 24-fold increase in token consumption between 2026 and 2030, reaching 120 quadrillion tokens per month as consumer and enterprise agents scale.
Recent examples show how quickly AI costs can escalate. Microsoft has reportedly begun removing most internal Claude Code licenses and pushing many developers toward GitHub Copilot CLI instead. Claude Code had become popular inside Microsoft after the company opened access to thousands of employees, but the company is now winding down most usage in one major division by the end of June, with sources saying the decision is also financial.
Uber offers an even sharper warning. Uber CTO Praveen Neppalli Naga commented that the company’s surging use of Claude Code exhausted its expected 2026 AI coding budget only a few months into the year. Interestingly, Uber executives also started questioning whether higher token usage was translating into proportionally more useful consumer features, and that AI spending was creating trade-offs with hiring.
Then there is OpenClaw. Peter Steinberger, the creator of the open-source AI agent project, posted a dashboard showing more than $1.3 million in OpenAI API usage over 30 days. The bill covered 603 billion tokens across 7.6 million requests and roughly 100 coding agents. Luckily for Steinberger, OpenAI was picking up the bill. But the example still gives us a rare public glimpse into what large-scale autonomous coding can cost when budget constraints are removed.
The lesson is not that these tools do not work. Often, the problem is that they work well enough that people cannot stop using them.
That is the paradox. Cheap unit costs can still produce large bills when consumption explodes. A single token may be inexpensive. A fleet of agents running continuously across thousands of employees, tools, repositories, documents, and workflows is something else entirely.
This is why the naïve comparison between “human labor cost” and “AI labor cost” is so dangerous. It assumes that the machine is simply a cheaper substitute for the person. But digital labor behaves less like a salary and more like a utility. It scales with usage, ambition, workflow design, model choice, latency requirements, and governance discipline. A careless AI rollout can easily produce the worst of both worlds: the company still pays the salaries, adds a large new token bill, and gets only marginal productivity gains because no one redesigned the work.
So why would any rational executive use digital labor? Because the real selling point of a virtual workforce is not that it is cheaper. It is that it is different.
Before the rise of digital labor systems, Robotic Process Automation was typically sold as labor substitution. Take a repetitive process, map the steps, build a bot, reduce headcount or redeploy staff. The unit of value was usually cost reduction. AI agents are more interesting than that. They are not merely faster clerks. They are a different type of intelligence which is both autonomous and capable of handling ambiguous situations without breaking. Importantly, their value is highest not in isolation, but when they are combined with human intelligence in novel configurations.
In healthcare, that can mean moving from episodic patient support to continuous patient engagement. A pharmaceutical company introducing a new therapy can use AI-enabled systems to check in with patients daily, reinforce adherence, surface confusion, and flag side effects or anxiety before they become reasons for discontinuation.
In insurance, digital labor can absorb the administrative burden that overwhelms organizations during moments of crisis. Allianz’s Project Nemo, launched in Australia in 2025, uses seven specialized AI agents to manage simple claims and has reported an 80% reduction in claim processing and settlement time, while still keeping humans involved in final payout decisions. Neptune Flood offers a similar lesson from Hurricane Ian, when 30% to 35% of claims were submitted through its bot, allowing customers to act at any time of day while human teams focused on higher-complexity support.
In sales, the same logic shows up as a daily intelligence layer for frontline teams. AI-driven “next best action” systems can help thousands of representatives prioritize which customers to meet, why now, what to discuss, which objections to anticipate, and what content to share. Verizon’s experience points to the power of this augmentation model: an AI assistant used by 28,000 customer service representatives helped reduce call times and contributed to a nearly 40% increase in sales through that team.
These are not one-for-one replacement stories. They are examples of system redesign. Humans remain central, but their work shifts toward judgment, empathy, persuasion, exception handling, and relationship-building. Machines take on the work that is too repetitive, too granular, too constant, too data-heavy, or too uneconomic for people to perform at scale.
At Dell Technologies World in Las Vegas, I interviewed John Roese, Dell Technologies' CTO. He made a point that stayed with me. There is a whole class of work that has historically been uneconomic for people to do — not because it lacks value, but because the value is too small, too distributed, or too continuous to justify human effort.
Take data hygiene. Every executive knows that CRM data decays. Contacts change roles. Doctors switch specialties. Customers move. Account fields go stale. Most companies periodically pay people or third parties to clean this data retrospectively, which means the information is already wrong by the time it is fixed.
Roese has argued for autonomous “hygiene agents” — systems that continuously monitor, update, and clean databases in the background. In an earlier interview, he described these agents as a way to handle useful but neglected tasks, including CRM data cleaning, that humans often do not do because the return on manual effort is too poor.
The economics are fascinating. Suppose an up-to-date record is worth 50 cents, but it costs a dollar to fix through a cloud-based AI workflow. That makes no sense. But if you own a data center with GPUs sitting idle at night, the equation changes. Suddenly, a task that was uneconomic at 2 p.m. on rented infrastructure may be attractive at 2 a.m. on owned capacity.
This is why AI infrastructure strategy is becoming inseparable from AI operating strategy. The old cloud question — rent or own? — returns in a new form. For experimentation, renting is usually rational. For variable workloads, APIs are powerful. But for persistent, high-volume, strategically important digital labor, companies will increasingly ask whether they need their own AI factories, private inference capacity, or hybrid architectures.
The point is not that everyone should rush back on-prem. The point is that AI strategy now has an infrastructure P&L. A company that treats agentic AI as a software subscription may discover too late that it is actually running a variable-cost labor utility.
In my earlier piece on “tokenmaxxing,” I explored the strange new pressure on employees to demonstrate that they are being AI-powered at work. The phenomenon is understandable. Leaders have signed expensive enterprise AI agreements. Boards want proof of adoption. Managers want value. Employees want to demonstrate high performance.
But measuring AI transformation by token consumption is like measuring innovation by electricity usage. It is not only unwise; it is unfair. Employees should not be asked to carry the burden of corporate productivity simply because leadership signed a bulk licensing deal. In most companies, that approach will produce bad economics. Salaries remain. Token costs rise. People use the tools because they are told to use them. The visible activity goes up. The actual value may not.
The most dangerous version is when companies begin cutting people to pay for rising AI costs created by poorly governed AI adoption. That is not transformation. It is managerial arbitrage — and not a very good one. The right dashboard is not “tokens consumed.” It is “value created.” Did cycle time fall? Did error rates improve? Did conversion increase? Did customer satisfaction rise? Did engineers ship more reliable code? Did salespeople spend more time with the right customers? Did claims settle faster? Did risk decline? Did the organization create a capability that competitors cannot easily copy? Without those answers, token usage is just a vanity metric with an invoice attached.
There are people who should be worried about their jobs. But it shouldn’t be the average knowledge worker. Despite the hype, in the near term, the people most exposed are the leaders who do not understand what it takes to run a real AI transformation.
Real transformation does not come from marginal improvements to individual productivity. It comes from redesigning systems of work: where decisions happen, what data flows where, which tasks should be automated, which should be augmented, which should remain human, and how the economics of machine intelligence are governed.
This is why the frontier is already moving beyond prompt engineering. Anthropic’s engineering team describes context engineering as the natural progression of prompt engineering. The challenge is no longer simply how to phrase an instruction, but how to curate and maintain the entire context around a model: system instructions, tools, external data, message history, and the operating environment in which an agent acts.
That is a management lesson disguised as a technical one. The future of AI advantage will not belong to the companies with the most enthusiastic users. It will belong to the companies that know how to design context, workflows, incentives, governance, infrastructure, and human-machine teams around valuable outcomes.
For decades, too many managers have been rewarded for allocating tasks, counting activity, and reducing cost. Digital labor exposes the limits of that model. If intelligence becomes abundant but expensive, the scarce skill is not prompting. It is judgment. It is architecture. It is knowing where machine work creates leverage and where it merely creates a bigger bill.
The rise of digital labor should shift the debate beyond whether AI will replace jobs, or whether it will be cheaper than people. Both questions miss the point. Digital labor is a new class of enterprise capability: part workforce, part infrastructure, part capital investment. It has to be designed, governed, and combined with human judgment. The uncomfortable truth is that digital labor will not only test the adaptability of workers. It will test the imagination, competence, and relevance of the people managing them.