The pitch has been consistent: replace expensive human labour with AI, cut costs, improve margins. But a series of reports this week is complicating that narrative in ways worth taking seriously. Uber's chief technology officer revealed that the company burned through its entire 2026 AI budget on coding tools alone, specifically Anthropic's Claude, well before the year is out. "I went back to the drawing board because the budget I thought I would need has already evaporated," he said. He was not, he added, considering slowing engineering hiring as a result.
An Nvidia vice president for applied deep learning offered a parallel observation: for his team, compute costs run well ahead of the cost of the employees the AI was nominally deployed to augment. These are not companies on the margins of AI adoption; they are among the most technically sophisticated users in the industry. Their experience suggests that the expected economic logic of AI deployment is not unfolding as cleanly as the efficiency narrative implies.
An MIT study adds a systematic dimension. Researchers examined whether deploying AI for computer vision tasks was actually cheaper than hiring human workers to do the same work. In only 23 per cent of tasks studied was AI the more economical option. For the remaining 77 per cent, when implementation, hardware, maintenance, and ongoing adaptation costs were included, human labour was cheaper. The study notes that the case for replacement is most compelling for highly repetitive, well-defined tasks at scale; it weakens significantly when the tasks require variation, judgment, or adaptation to context.
Part of what makes AI cost structures difficult to manage is their variability. A human employee's salary is a largely predictable fixed cost. Every AI interaction generates a charge: tokens consumed, compute used, API calls made. For companies running millions of customer interactions or generating large volumes of code, those costs accumulate in ways that are genuinely hard to forecast at the outset. Add to this the ongoing requirement for human oversight: because current AI systems produce errors with a frequency that varies by task, many organisations find themselves running a parallel layer of human review rather than eliminating human involvement. The result is sometimes higher total cost than either approach alone.
Global IT spending is projected to reach $6.31 trillion in 2026, up 13.5 per cent from 2025, driven in significant part by AI infrastructure, software licences, and cloud services. Investors and board members are beginning to ask with increasing seriousness how the productivity gains are supposed to justify those outlays, especially as the gains are easier to demonstrate in narrow task contexts than at the organisational level.
Gartner analyst Caitlyn McDonough identifies a category of costs that typically go unmeasured: what she calls talent ripple effects. As AI takes on routine tasks and junior roles shrink, the talent pipeline that would have grown into future capability quietly dries up. Skills that are no longer practised atrophy. Career pathways narrow. The human capacity the organisation will need in three years is being reduced by decisions that show only as cost savings in this year's spreadsheet. Those are real costs, deferred rather than avoided, and they are not appearing in most AI investment assessments.
None of this undermines the case that AI is genuinely useful and that productivity gains are real in specific, well-suited applications. The question is whether the overall economics of deployment are as straightforward as the narrative suggests, and whether the efficiency gains being cited to justify current headcount reductions accurately reflect what is actually happening to total cost. For shareholders and boards asking for proof, the honest answer is often that the evidence is partial, the timeline is short, and the full accounting has not yet been done.