The AI Spending Surge With Nothing to Show For It

Enterprise AI capex is accelerating like a Series B that just closed its largest round. Up 40% year-over-year, with no sign of deceleration. Boards approved it. CFOs green-lit it. Consultants sold it. Workers deployed it.

Then nothing happened.

Labor productivity in services sectors remains flat or has actually declined. The numbers are there in the quarterly data, stubborn as a subordinate who won't read the memo. Workers are now spending roughly 30% of their day interfacing with AI tools—ChatGPT sessions, internal LLM deployments, "AI-assisted" workflows that mostly assist in generating another meeting about workflow optimization. Output hasn't budged. Revenue per employee static. Quality metrics unchanged. The efficiency gains that were supposed to compound like crypto in 2021? Nowhere.

This is the productivity paradox boards finally can't ignore. And the question they're beginning to ask in closed sessions is the one that keeps CFOs awake: if the inputs are up 40% and the outputs are flat, what exactly are we paying for?

The answers arriving in this earnings season will be illuminating. Some CFOs will double down—cite "early-stage adoption" and "organizational learning curves," the same justifications deployed when any technology fails to deliver but the vendor relationship is too expensive to exit. Others will quietly redirect capex. A few will admit they don't know what they're measuring.

The pattern is recognizable to anyone who's lived through a digital transformation. Peak enthusiasm masks peak confusion. The executive who championed the AI initiative is now head of something else. The team that implemented it is tired and has moved on. The tool is still running, consuming budget and employee time in roughly equal measure.

What makes this cycle different is the scale of the bet and the speed at which reality contradicted the thesis. AI wasn't supposed to be like previous software cycles—slower adoption, longer payoff, organizational friction. AI was supposed to be different. Faster. More transformative. The ROI was supposed to be obvious within quarters.

It isn't.

Boards, which are increasingly serious about AI governance (another way of saying "increasingly nervous"), are asking for data that doesn't exist. Productivity metrics by tool. User adoption correlations. Actual time savings validated through... well, through something other than sentiment surveys filled out by the same people who approved the tool in the first place.

The quiet cuts have already begun. Transformation budgets that claimed AI as their centerpiece are being repriced. One global services firm, speaking anonymously because internal politics remain internal, cut its AI capex guidance by 15% last quarter while describing it publicly as "optimization." Another reported they'd pause new LLM licensing pending "integration analysis." Translation: the current tools aren't working and we're not buying more until we understand why.

What gets cut next depends on which CFO's nerve cracks first and admits what the spreadsheets already show: that enterprise AI adoption rates and actual productivity improvements have decoupled entirely.

The conversation won't happen in earnings calls, naturally. Those are for optimism. It will happen in closed board sessions where a single director—probably the one who actually worked in tech operations rather than finance—will ask the question that kills the narrative: "What would happen if we stopped spending on this and nobody noticed?"

The answer will determine whether AI capex continues its 40% annual climb or whether this quarter marks the inflection point where boards began to distinguish between technological adoption and actual business impact.

Based on historical precedent with transformational technology, the answer is probably: they'll keep spending, slower, quieter, with reduced expectations and revised metrics that somehow capture the value they can't quite measure.

But the productivity paradox won't solve itself. And CFOs have learned, finally, that pointing at a technology isn't the same as proving it works.