← ALL POSTS
AIHypeProductivityEngineeringReality

The Skeptic's Reality Check: What AI Is Actually Delivering in 2026

Goldman Sachs found no economy-wide productivity impact from AI. MIT Media Lab says 95% of organizations see no measurable returns. $650B in capex is meeting a very short list of demonstrated results. Here is what the numbers say.

April 3, 20265 min read

The Skeptic's Reality Check: What AI Is Actually Delivering in 2026

The AI industry spent 2023 and 2024 promising transformation. In 2026, the invoice arrived.

Goldman Sachs published a finding in March that should have been bigger news than it was: at the economy-wide level, there is no meaningful relationship between AI adoption and productivity. Not a weak relationship. Not a delayed one. No relationship.

That is the honest starting point for any serious evaluation of where AI actually stands.

The Productivity Data Is Not What You Were Sold

AI Productivity Reality: Where Gains Are Actually Showing Up

The Goldman finding comes with a nuance worth understanding: the economy-wide number hides real gains in two specific, narrow domains — software development and customer support, where median productivity improvements of around 30% are showing up. That is meaningful. It is also the entire list.

Outside those two use cases, the picture is bleak. MIT Media Lab research found that 95% of organizations report no measurable AI returns. Only 10% of S&P 500 management teams have quantified AI's impact on specific use cases. Only 1% have quantified its impact on earnings.

The gap between what the industry claimed and what organizations can actually measure is not a rounding error. It is the story.

Note: The Goldman finding does not mean AI has no future productivity impact — it means the timeline and distribution of that impact look nothing like what was promised in 2023. Broad economic transformation takes longer than hype cycles. That is not news; it is always true.

The counterpoint the optimists cite: employees at companies with ChatGPT enterprise accounts save 40–60 minutes per day on average, and 75% say they can complete tasks they previously couldn't. Those are real gains. The question is whether they show up in business outcomes — and at the 1% earnings quantification rate, the answer is: rarely, and rarely cleanly.

The $650B Question

The $650B Capex Reckoning: AI Spend vs. Measured Returns

The hyperscalers committed roughly $650 billion in AI-related capex. In February 2026, Nasdaq tumbled when investors did the math: that spend contributes an estimated 0.1 to 0.2 percentage points to GDP growth. That is not a return on $650B. That is a very expensive infrastructure bet on a future that has not arrived yet.

"The Great Pivot of 2026" is not a metaphor — investors are actively rotating out of AI-premium equities and into what analysts are calling the "Old Economy." The AI premium that the market granted based on infrastructure growth and grand promises is being repriced against demonstrated results. The results are not there at scale.

Agentic AI compounds this. The claims are largest ("AI agents will automate most white-collar work") and the evidence is narrowest. Real-world agentic deployments remain brittle outside controlled tasks, hallucinate at critical moments, require constant human supervision, and deliver marginal ROI in most enterprise contexts. The benchmark numbers are impressive; the production reliability is not.

What the Skeptic Should Actually Do

The skeptic's error is the same as the believer's error: treating AI as a single thing with a single trajectory.

The Goldman data makes the picture precise: 30% productivity gains in coding and customer support are real and repeatable. If your work is in those domains, the tool is worth using aggressively. If your work is outside those domains, you are in the 95% — and the honest move is to design a small, measurable pilot before committing infrastructure spend.

The questions worth asking before any AI deployment:

The organizations that will get value from AI in 2026 are the ones asking those questions. The ones running AI-for-AI's-sake initiatives are producing the MIT Media Lab statistic.

Key Takeaways

Sources

← BACK TO ALL POSTS