AI Productivity Paradox: Why Individual Gains Don't Scale Across Organizations

The AI spending wave is real. According to Gartner, worldwide AI investment hit $2.59 trillion in 2026, a 47% jump over last year. But something strange is happening: US total factor productivity grew just 0.07% over the four quarters ending in Q1 2026. That is near standstill. And 95% of enterprise generative AI pilots have produced no measurable bottom-line effect.

Writing in Computerworld, tech journalist Mike Elgan explores why.

The Doom Loop Theory (and Why He Doesn't Buy It)

A working paper by University of Pittsburgh professor Mark Ma and colleagues argues that employee resistance drives the productivity gap. Workers fear AI will replace them, so they quietly resist adoption, which kills the gains companies are counting on. Elgan acknowledges the data but pushes back on the causal claim: the researchers never establish causation, only correlation.

He notes that more than half of US workers already use AI on the job. If employee foot-dragging were the main problem, productivity numbers would look different by now.

A Better Explanation: AI Overload

Elgan's own thesis centers on what he calls "AI overload." When one person uses AI to triple their output, that output lands in someone else's inbox. At scale, the reading burden far outpaces any writing efficiency gained.

Key takeaways:

• LinkedIn data shows US job applications per posting have roughly doubled since spring 2022, burdening hiring managers while making it harder to get hired.
• A 2026 MIT and USC study of 4.5 million federal civil cases found AI-generated legal text rose from 1% in 2023 to 18% in early 2026, slowing courts.
• "AI can make an organization extraordinarily busy without necessarily making it more productive," said Justin Greis, CEO of Acceligence.

Elgan argues that the fix is not more AI adoption, but a wholesale redesign of workplace AI, building tools that raise organizational output rather than dumping individual productivity onto everyone else.

Read the full article on Computerworld