When AI costs more than people: the new unit economics of agents

For the first two years of the generative AI boom, cost was a rounding error. Inference prices for equivalent capability fell roughly tenfold per year, every deployment memo assumed the marginal cost of AI work was effectively zero, and the business case wrote itself.

2026 is the year that assumption met production agents. Fortune reported in May on Microsoft data showing that agent and token costs can exceed the cost of the human labor being replaced. Axios put it more bluntly in April: AI can cost more than human workers now. The anecdotes match the data; one widely reported example had a major tech company's engineering organization exhausting its annual AI coding-tool budget in four months.

How can inference get cheaper every year while AI bills explode? Because the two curves measure different things. The price of a given capability falls, but nobody deploys last year's capability. Teams migrate to frontier models, and agents multiply consumption structurally: an agent attempting a complex task does not use one prompt's worth of tokens, it uses hundreds of steps of reading, reasoning, retrying, and verifying. Cheap per token, expensive per outcome. Analyst data shows the same split, with frontier-tier pricing falling far faster than the budget tiers people actually run production loads on.

The recalibration is coming from the top

The cost reality is arriving alongside a performance reality, and it is being acknowledged in unusually candid terms. Mark Zuckerberg, whose company cut roughly 8,000 roles this year in an AI restructuring it later admitted was partly mishandled, told a Meta town hall this month that agent development "hasn't really accelerated in the way that we expected." Gartner has been forecasting for a year that over 40% of agentic AI projects will be canceled by end of 2027 for costs and unclear value. Bain's spring survey found only 7% of companies run fully autonomous agents in production.

Meanwhile the same companies tempering expectations are spending historic sums, roughly $725 billion in combined capex this year, on the infrastructure underneath it all, and operators like Cisco are rolling agents out to their entire workforce. Both things are true at once: the capability is real and heavily backed, and the economics at the task level are far less automatic than the 2024 narrative promised.

How to model agent unit economics

The good news is the math is not hard. It is just rarely done. Before deploying any agent workflow, put four numbers on one page:

• Cost per completed task. Not cost per token or per seat. Run a realistic pilot, divide total spend, including retries and failures, by tasks completed to an acceptable standard. Retries are where agent economics go to die: a workflow that succeeds on the third attempt costs three times the demo estimate.
• The human benchmark, fully loaded. What does the same task cost when a person does it, including their tools and overhead? This is the line the agent has to beat, and per Microsoft's own data, it sometimes does not.
• Supervision cost. Someone reviews agent output, handles escalations, and fixes the failures. That time is part of the agent's cost, not a separate line. The agent-to-human ratio you can support is set by this number.
• The right model for the job. The single biggest lever. Most production tasks do not need a frontier model, and routing routine steps to cheaper models while reserving frontier capability for the judgment steps can change task economics by an order of magnitude. Teams that never revisit their model choice are usually paying frontier prices for commodity work.

Then track one ratio monthly: value per task over cost per task. Falling model prices help the denominator every quarter. Improved workflows help both. If the ratio does not clear a sensible threshold, redesign the workflow or return the task to humans without embarrassment. An agent that loses money per task at scale is not innovation. It is just automated spending.

Do the math before your CFO does

Every technology wave ends its free-money phase, and that ending always sorts the field. When capital was free, discipline about return on capital was a competitive advantage. Now that AI is treated as free, discipline about cost per task is the equivalent edge. The companies that get burned in the next two years will not be the ones that deployed agents; they will be the ones that deployed agents without a meter running.

The winners of the cheap-inference decade will be decided now, in the expensive-agent years, by whoever builds the habit of knowing what their AI work actually costs. It is a spreadsheet, not a moonshot. Build it before your CFO builds it for you.