The Bill Came Due

For two years, the story was simple: AI was coming for your job. A digital worker never sleeps, never takes vacation, and costs a fraction of a salary. Executives repeated it. Analysts modeled it. Workers braced for it.

Then, in May 2026, Microsoft quietly began canceling most of its internal Claude Code licenses. The tool had been rolled out to thousands of engineers in its Experiences and Devices division, the team behind Windows, Microsoft 365, Teams, and Surface, just six months earlier. Adoption was explosive. Engineers loved it. And that was exactly the problem. Token-based billing ate the division’s annual AI budget far ahead of schedule, and by June 30, engineers were told to switch to the cheaper in-house GitHub Copilot CLI.

Uber’s story was worse. The company gave Claude Code to roughly 5,000 engineers in late 2025. By April 2026, its CTO admitted the entire 2026 AI coding budget was gone. Four months. A full year of money.

The “Great AI Replacement” didn’t die because the technology failed. It died because the invoice arrived. Buried in that invoice is the best news workers have gotten in years, if they know how to read it.

The Costs Are Real

Start with the raw numbers, because they’re brutal.

At Uber, a typical engineer ran up $150 to $250 per month in AI usage. Heavy users hit $500 to $2,000 each. Per month. Per engineer. Multiplied across thousands of engineers. Uber’s own CTO burned roughly $1,200 in tokens during a single two-hour demo. The company’s COO called the figures “head-exploding,” and for all the tokens burned, he struggled to point to anything customers could actually feel.

The culprit isn’t chat. A chat prompt is a single, cheap exchange. The culprit is the agent: an AI that doesn’t answer a question but chases a goal. It reads a codebase, drafts a plan, writes edits, runs tests, and retries when something breaks. Every step in that loop bills tokens. A typical agentic task consumes 5 to 30 times the tokens of a simple chat exchange. In pathological cases, reports show spikes of over 1,000x.

The most expensive work is the work that fails. On SWE-bench, an industry benchmark built from real software bugs, the best agents still solve fewer than half the problems. A difficult bug can mean five to ten full attempts, each billed at full price, before the agent gives up or a human steps in. Full price for the dropped wrench, no guarantee the leak gets fixed.

Cheaper hardware won’t save the day either. Gartner projects that by 2030, running a top model could cost roughly 90% less than in 2025, and still predicts enterprise AI won’t get cheaper overall. Agentic workloads consume tokens faster than prices fall. Cheap everyday tokens and cheap access to frontier capability are two different products. The powerful one stays expensive.

But So Are the Gains

The costs are real. So is the productivity.

Across the major Q1 2026 datasets, the numbers land in a tight band: McKinsey’s Global AI Survey reports a median of 6.4 hours saved per knowledge worker per week. Salesforce says 6.7. Slack’s Workforce Index says 6.1. Nearly 80% of organizations report measurable productivity gains from AI tools. Uber’s own CEO acknowledged that AI agents now write about 10% of the company’s code. Real work, actually shipping.

Even Microsoft’s cancellation proves the point in reverse. Claude Code wasn’t cut because engineers disliked it or because it produced garbage. It was cut because it was, in the words of the original reporting, perhaps a little too popular. Engineers chose it over Microsoft’s own product at scale. The tool delivered enough value that usage, and therefore cost, blew past every budget projection.

Read that carefully, because it’s the opposite of “AI doesn’t work.” A tool that doesn’t work doesn’t blow up your budget. Nobody overuses a hammer that misses the nail. The productivity is real. What’s broken is the economics of how that productivity gets purchased.

That’s the paradox defining 2026: 79% of organizations report productivity gains, yet only about 5% achieve what IBM classifies as substantial ROI, meaning returns that clearly justify the total cost. The gains exist. They’re just being set on fire.

The Gap Is the Operator

So where do six saved hours per week go, if only 5% of companies see real returns?

Up in smoke. Burned by the way people use the tool.

Through 2025 and into 2026, Big Tech turned AI consumption into a sport. Amazon staff were pushed to “tokenmaxx,” slang for burning as many tokens as possible, as if usage itself were the goal. At Meta, an employee built a leaderboard called “Claudeonomics” ranking workers by AI spend. Heavy use became a status symbol. The engineer firing off the most prompts didn’t look wasteful; he looked like the future. Uber built leaderboards ranking teams by AI adoption, and usage climbed from 32% of engineers in February 2026 to 84% by March.

Every one of these incentive structures rewarded the same thing: activity instead of outcomes. To any individual engineer, one extra prompt cost a fraction of a cent. Nobody hesitated. Nobody asked whether the prompt was necessary. Multiply that across thousands of workers and millions of tasks, and you get a bill no one can defend. A textbook tragedy of the commons.

Meanwhile, the returns aren’t evenly distributed at all. Gartner’s 2026 pulse data found only 41% of agentic AI deployments reach positive ROI within a year, and 19% never reach payback. The failures were attributed overwhelmingly to governance gaps, unmeasured rework, and evaluation drift, not to the capability of the models themselves. The bottleneck isn’t the frontier model. It’s everything between the model and a measurable outcome.

Put plainly: two people can sit down with the identical tool, the identical model, the identical subscription, and one produces a multiple of the other’s value at a fraction of the cost. The tool doesn’t determine the return. The operator does.

The Formula: Domain Knowledge + Technical Fluency

What separates the operator who compounds value from the one who compounds costs? Two ingredients. You need both.

Domain expertise: knowing whether the output is actually right. An agent will confidently produce plausible-looking code, analysis, or copy whether it’s correct or not. The person who can’t evaluate the output has no way to catch the failure loop early. They accept confident garbage, or worse, they let the agent retry ten times at full price chasing a fix a knowledgeable human would have spotted on the first attempt. Domain knowledge turns a runaway process into a two-minute correction.

Technical fluency: knowing how to direct the tool efficiently. When to use a heavyweight agent versus a single cheap prompt. How to scope a task so the model doesn’t wander. How to structure context so the first attempt lands instead of the fifth. When not to use AI at all, because the task is faster done by hand. Prompt discipline is cost discipline.

One without the other fails predictably. The brilliant domain expert who can’t direct the tool burns tokens on vague, sprawling requests and drowns in retries. The technically fluent operator with no domain judgment produces polished, efficient, wrong answers. Both are expensive. Both look productive on a usage leaderboard. Neither shows up on the P&L.

The person who has both is a different economic species entirely. They make AI not wasteful, and in a world where the median deployment loses money, that’s a rare and monetizable skill. Companies that just watched Microsoft and Uber’s budgets evaporate aren’t going to hire “people who use AI.” They’re going to hunt for people who use it well, because those two profiles now sit on opposite sides of the ledger.

No Opting Out

Here’s the uncomfortable part: in a growing number of industries, refusing the tool won’t be an option.

Software engineering is already there. At Uber, adoption hit 84% of engineers in a matter of months, and AI now writes a meaningful share of production code. The same trajectory is underway in customer operations, marketing, analysis, legal review, and design. When the baseline worker in your field is saving six hours a week with AI, the worker who refuses it isn’t principled. They’re slow. The market won’t debate it. It will price them out.

But notice what that means for the skilled operator. If AI use is mandatory and the median operator burns money using it, then the person who combines domain mastery with efficient AI direction becomes structurally impossible to replace with automation alone. Why? Because replacing them means deploying the very agents that just torched Microsoft’s and Uber’s budgets. Agents that fail half of hard benchmark tasks, bill full price for every failed attempt, and require exactly this kind of human to supervise them profitably.

Competitor, Not Casualty

Follow the logic one more step, because this is where it gets interesting for workers.

The skilled operator, fluent in both the domain and the tools, disciplined with the spend, is now a one-person production unit of a kind that has never existed before. They carry the judgment of a senior professional and the output of a small team. Their employer knows it. Their employer’s cost spreadsheet knows it.

Historically, leaving your job to compete with your employer meant recreating an entire team’s worth of capacity. Now the same tools that companies deploy at ruinous scale are available to individuals at trivial scale. And at individual scale, the token economics that break a 5,000-engineer rollout work fine. A few hundred dollars a month of AI spend that bankrupts a leaderboard-driven engineering org is a rounding error for a skilled solo operator billing for outcomes.

One Caveat

None of this applies evenly. Some industries will face total automation, where the economics eventually favor the machine no matter how skilled the human. Others will prove automation-resilient, where judgment, accountability, and the cost of failure keep people at the center. Nobody can say with certainty which jobs land in which bucket. Only the future will