ON FUTURE OF WORK

The Future of Work Runs on Tokens. Attention Decides If It Pays.

One company spent $500M on AI in 30 days with no one watching the gauge. The future of work isn't about number of agents or headcount. It's about coordination.

Compute is cheap now. Attention isn't. The distance between them is where the next decade is won.

Whether or not you’re not an AI-industry follower, you probably heard the recent news.

One company ran up $500M dollars on a single AI vendor in only 30 days. That’s nearly $12,000 per minute for each of the 43,000 minutes in that single month. No one had set a usage limit, and the meter was running. The story recently hit and traveled across the media; and many leaders have echoed similar AI cost challenges within their own organizations.

Even well-known leaders selling AI tokens are uneasy. Sam Altman recently called their cost a huge issue, as overspending on AI has hardened into a serious concern from OpenAI customers. He added that six years ago, a top token spender would use 100,000 per month; today it's at 100 billion in that same period. Also, Microsoft began canceling most of its direct Claude Code licenses, partly over cost.

Strip away the anecdotes and one fact remains. Compute scales the moment you let it. It does not wait for a business case, and it never stops asking whether the work it just did was worth doing.

This is a future-of-work problem, and most of the debate aims at the wrong target. We argue about which jobs disappear and spend far less time on decisions, specific to which costs should run, or whether to run them at all. Attention is where the next decade separates its winners from its losers. What looks like a spending problem is really a challenge of measuring and informing decisions. In fact, we are scaling a cost faster than we built the instruments to see it.

Three fuels, and one that behaves differently

Picture the inputs a company runs on as a metabolism. For most of the last fifty years, the enterprise burned two fuels a CFO could see clearly: labor and capital, as well as people and money. Every instrument of the company, from the annual plan to the quarterly review, was built to measure and meter those two.

Then came the arrival of a third fuel that we call code. Agents and AI tokens now wired into daily work across people, teams, and whole enterprises. It behaves nothing like the first two. Overspend labor and people burn out, so you feel it fast. Drain too much capital and the balance sheet tightens, so the board feels it. Both push back, where tokens do not. They rapidly scale on their own, silent and elastic; the bill lands a quarter or two after the decision that caused it, long after anyone remembers making it.

It’s a major reason why business-related AI spending is not paying for itself yet. Technology usually does its job. What trips companies up is the choice of which work to hand it and the value derived from such. Axios reported that most teams default to automating the tasks people dislike rather than the ones most valuable to the business, and that AI still delivers most reliably in narrow areas like coding. Spread that across a few thousand seats and the metabolism runs red hot with no one watching the gauge.

Some companies made the gauge run backward on purpose. The idea and strategy is known as ‘tokenmaxxing,’ where leaders have espoused the idea that using more AI tokens equates to higher developer output and workplace automation. In fact, Uber set up an internal leaderboard ranking teams by how much AI tooling they used, then burned through its entire 2026 budget for AI coding tools in four months. Its president, Andrew Macdonald, later said he could not tie that spending to anything customers would feel. “That link is not there yet,” he said.

Even setting misuse aside, the economics bite. Gartner expects the per-unit cost of running advanced models to fall around 90% by 2030, and still projects enterprise AI bills will keep climbing, because agentic systems burn far more tokens per task than the tools they replace.

The mistake: dressed up as a strategy

While the reflex is understandable, it’s now plaguing organizations as a costly misstep. Automate the task and cut the headcount. Book the savings. A year on, the savings are real, and the run-rate that replaced them is more real. About a year in, a CFO and a CTO end up across a table from each other asking why the bill keeps climbing; each having assumed the other owned that number.  The CEO of CloudBees told Axios that for some firms, layoffs are now the only lever they can pull to offset their AI bills.

Read that slowly. The cost of not being able to have timely, accurate measurement of the value of the technology is forcing the cuts, and the cuts are being sold to the market as the plan.

This is the part of the future-of-work conversation that gets skipped. We talk about how many jobs AI will replace and almost never about what the replacement costs to run. Every decision to automate now carries a price that keeps ticking after the decision is signed, the way a lease charges long after you stop using the office. Compensation used to have two big lines, salary and benefits. It has a third now, and it is measured in tokens. Token cost has become a structural line in the P&L, recurring and growing, and most companies cannot yet tell you what theirs is. EBITDA was never built to hold it. The math of the enterprise changed, and the frameworks have not caught up.

Call this decision economics.

The number that matters is the full run-rate of the path you choose, priced before you commit rather than discovered two quarters later. The enterprises pulling ahead run a few questions before they automate anything. What will this cost to run once it is live, well past the flattering pilot number? Which calls still need a person in the seat, and what breaks downstream if the model is confidently wrong? Most business cases never get past the headcount math, which is why the bills keep surprising people who thought they had done it.

Tokenomics is the budget. Humanomics is the bound.

Here is the discipline almost no one has built. Tokenomics tells you what you can compute, and it scales without asking permission. Humanomics decides what is worth computing at all, and where the line falls between which work AI should carry and work that stays human run. The first is an accounting exercise. The second is a judgment call, and judgment does not come cheap.

A bound and a budget are different animals. A budget is a number you are allowed to spend. A bound is a position you take on what your company is for, and which of its decisions should never be handed to a system that optimizes for the wrong thing. You can have an enormous budget and no bound at all, which is more or less how you spend half a billion dollars in a month without noticing.

Drowning in data, starving for signal

The bound depends on something most enterprises are bad at: seeing the one thing that matters in time to act on it. We are awash in data and short on signal. In most organizations, the majority of the signals that should drive a decision, by some estimates as much as ninety-five percent, never reach the person who could act while it still counts. The information exists. It may well sit in a system three teams away, correctly logged, and no one connects it to the choice in front of them until the moment passes.

The decade we spent building dashboards deepened the problem. A dashboard renders, in beautiful resolution, what already happened. Worse, a wall of green dashboards reads as comfort while the single number that was about to break remains siloed and hidden. The skill the next decade rewards is pulling the rare signal out of the noise and routing it to the one leader who can act on it fast enough to matter. Collecting more data was never the hard part.

Attention is the input that runs out

Which points at the resource that actually decides the future of work: attention.

Compute is abundant and getting cheaper every year. Attention is fixed. A leadership team has the same hours it always had, and a flood of new noise competing for them, much of it from the systems meant to help. Every agent you deploy adds more alerts and summaries, more of everything that looks urgent. Add a thousand of them, and the room only gets louder. Clarity is the one thing more agents will not buy you.

So, the scarce input is human attention, the judgment that decides what the tokens were for. The tokens are not the real constraint. Coordinating that attention across leaders, teams, systems, and the agents now working beside them has become its own layer of the enterprise, sitting above labor, capital, and code. Add 50 to 100 agents per employee, multiple that over thousands or tens of thousands in an organization, and the coordination problem compounds to plague decisions, workflows, and business results.

What is now emerging is a new category of enterprise technology called Systems of Attention. A coordinating layer that ensures the most important signals from an organization surface and are brought to the attention of the right people, to inform the best decisions in a specific time window for action.

Signals decide what is worth seeing. Attention decides what is worth a person's time. And decision economics decides what is worth the compute. A company that gets all three right can spend aggressively on AI, gain greater benefit and value, while still knowing where the bound sits.

The stakes are not abstract

More than half the companies on the Fortune 500 back in 2000 have since fallen off it. Some were unlucky. Many simply kept optimizing the wrong things with great discipline, right up to the end. The current moment hands every enterprise a faster way to do that, a way to scale activity no one stopped to question.

Spend on AI. Spend heavily, only where it earns its place. The work that matters is finding your bound: the tasks worth the compute, the judgments that stay human no matter how cheap the alternative gets, and the signals you can no longer afford to miss. The institutions that win the next decade will not be the ones that spend the most tokens. They will be the ones that know their bounds and hold it.

The uncomfortable part is that no finance team can hand you that number. The bound is a string of human decisions about what your company is actually for. Which surfaces the question most automation roadmaps race right past. The machine can almost certainly do the work. The real question is whether you can still see what it is costing you.