"AI costs more than the person it replaces": a headline to read carefully

A headline has been circulating in recent weeks that sounds counterintuitive, which is exactly why it has been shared so widely: AI is costing more than the workers it replaces. Forbes put it bluntly: AI Costs More Than The People It Replaced. Axios quoted an Nvidia vice president saying that, on his team, computing costs now exceed employee costs. Fortune reported that Uber used up its entire annual AI budget for programming in four months.

These figures are real. The problem is not the data, but the generalisations people draw from it.

What is really happening

The cases behind the headline share a specific pattern: companies using frontier models without spending controls for highly agentic software development tasks, where a single user request can trigger ten to twenty chained calls to the model. Amazon and Meta built internal rankings to measure AI use among engineering teams, and some employees ended up “burning” tokens on pointless tasks just to climb the leaderboard. One Anthropic employee spent as much as $150,000 in a month using Claude Code. For that spending to pay off, that engineer would have had to produce the equivalent of eleven ordinary engineers’ work.

This is not a problem of AI being inherently expensive. It is a spending governance problem: frontier models are applied indiscriminately, with no routing to cheaper models, and internal incentives reward consumption instead of results. An MIT study cited in Forbes’s coverage found that AI automation is economically viable for about 23% of the jobs analysed. For the other 77%, a person is still cheaper, simply because those jobs do not fit the kinds of tasks AI handles well.

What the headline leaves out: where AI really is much cheaper

Alongside those cases of overspending, there is equally strong data showing the opposite for well-defined tasks. A 2026 analysis of customer service workflows found that AI handles routine interactions for $0.25 to $0.50 per contact, compared with $3 to $6 for a human agent: a reduction of 85% to 92%, with a return on investment in four to six months. Researchers at Stanford and Carnegie Mellon found that AI agents completed some tasks 88% faster and at up to 96% lower cost than people, for the structured, high-volume work these systems are built for.

The difference between these two worlds is not the AI model. It is the type of task. Replacing a call centre agent who handles password resets is a structured, high-volume problem that can tolerate errors. Replacing a financial analyst who writes a credit report or a lawyer who reviews a contract clause is a contextual problem, with few cases per person, that demands a level of accuracy today’s agents cannot sustain without constant human oversight.

The underlying mistake: the big-model fallacy

A 2026 analysis of the enterprise inference cost crisis has a name for this mistake: the “big-model fallacy”, the assumption that every task requires a frontier model. Companies that fall for it see their AI bill soar even as the cost per token continues to fall: inference prices have dropped by a factor of 9 to 900 per year, depending on the task, and Gartner expects a further 90% decline by 2030. What is rising is not the price per token, but the uncontrolled number of tokens spent when there is no routing system to send each task to the right model.

The conclusion the data supports

As a general claim, “AI costs more than a human worker” is as imprecise as saying the opposite. The 2026 data points to a more nuanced conclusion: AI is consistently cheaper for structured, repetitive and well-bounded tasks. It can cost more than a person when used without control for complex tasks with frontier models and no cost-routing mechanism. Alarmist and triumphalist headlines describe two different scenarios for the same technology, not two contradictory truths about it.