NBER: 6,000 Execs See Zero AI Productivity Gains

Six thousand senior executives across four countries. Three years of AI investment. 89% saw no change in productivity. The National Bureau of Economic Research just published the most comprehensive study of AI's actual business impact, and the number that should dominate every board meeting this quarter is not the billions being spent. It's the fraction of that spending producing measurable results: close to zero for the overwhelming majority of companies deploying AI today.

What Actually Happened

Researchers Ivan Yotzov and Jose Maria Barrero surveyed nearly 6,000 senior business executives at firms in the United States, United Kingdom, Germany, and Australia for NBER Working Paper 34836, titled "Firm Data on AI." The survey targeted decision-makers with direct visibility into their organizations' AI deployment, productivity outcomes, and employment changes over the prior three years.

The findings are specific and unambiguous. More than 90% of managers reported no AI impact on employment at their organizations over the past three years. 89% reported no change in productivity, measured as volume of sales per employee. These are not marginal majorities. They are nine-in-ten surveyed executives at companies that were, by definition, engaged enough in AI to participate in the survey. The companies not tracking AI at all are not in this dataset.

Usage data from the same survey reveals why. More than two-thirds of executives regularly use AI tools. But their usage averages just 1.5 hours per week. The modal pattern at surveyed firms is leadership paying for AI subscriptions and occasionally prompting a chatbot, not integrating AI into core production workflows. Adoption without integration produces adoption metrics. It doesn't produce productivity.

The same executives who report zero past impact are optimistic about the near future. They predict AI will boost productivity at their firms by an average of 1.4%, raise output by 0.8%, and cut employment by 0.7% over the next three years. The gap between reported past impact and expected future impact is the clearest signal in the study: executives have internalized the narrative that AI will transform their businesses but haven't yet done the organizational work to make that transformation real.

Why This Matters More Than People Think

The NBER findings land in a market that has allocated capital as if AI productivity gains were already guaranteed. Nvidia's valuation has been premised on AI compute demand. SaaS companies embedding AI features have repriced their products upward. Enterprises have signed multi-year AI platform contracts justified by productivity ROI projections. The 89% zero-impact finding doesn't just challenge those projections. It questions whether the organizational conditions for AI productivity gains have been created at all in the vast majority of companies deploying the technology.

The parallel to the Solow Paradox of the 1980s is now unavoidable. Nobel laureate Robert Solow observed in 1987 that computers were visible everywhere except in the productivity statistics. For roughly a decade after mass corporate computer adoption, productivity data stayed flat. Then, in the mid-1990s, productivity growth accelerated as firms reorganized workflows, retrained workers, and built complementary software ecosystems around the hardware. AI may be following the same curve: the productivity gains are real but delayed by 5 to 10 years of organizational adaptation.

But there's a distributional twist that Solow's era didn't feature. PwC's 2026 AI Performance Study finds that 20% of companies are capturing 74% of all AI economic gains. The beneficiary firms share a profile: they're using AI to grow revenue, not just cut costs, and they've restructured workflows around AI capabilities rather than adding AI tools to existing processes. The other 80% aren't in a holding pattern. They're running up AI subscription costs, training expenses, and integration complexity without realizing offsetting gains. That gap is widening with every quarter.

The Competitive Landscape

The productivity divide has direct competitive consequences. Firms in the 20% bracket, those capturing 74% of AI economic gains, are compounding advantages the remaining 80% cannot close through incremental adoption. When a financial services firm uses AI to process loan applications at 10x the volume with the same headcount, it doesn't just lower costs. It prices competitors out of the market. When a software development team using AI coding assistants ships features at twice the velocity, it captures market share that the slower team won't recover.

The employment data embedded in the NBER findings contains a paradox that the headline number obscures. Executives reported no employment impact from AI. But US employers cut 40,000 jobs in Q1 2026 attributable to AI automation, according to separate labor market data. The gap between what executives report and what labor markets show suggests one of two dynamics: either the layoffs are concentrated in the 20% of high-productivity AI adopters who aren't in the zero-impact majority, or executives are systematically attributing AI-driven restructuring to other causes in survey responses. Both explanations are uncomfortable for different reasons.

The skeptics' view deserves space beyond the questions section at the end. The bear case is that AI productivity gains are structurally limited by the nature of business processes themselves. AI tools excel at discrete, well-defined tasks. Most productivity at large companies is locked in coordination, judgment, and context management between tasks, the connective tissue of business operations that AI cannot yet automate. The risk is that the 1.5 hours per week usage average is not a failure of adoption but an accurate reflection of how much AI can help given the current state of enterprise process design. If that's true, the 80% of companies seeing no gains aren't failing to implement AI. They're accurately perceiving its current limits.

Hidden Insight: The 1.5-Hour Number Is the Real Diagnosis

The NBER paper's choice of productivity metric, sales volume per employee, is both practical and revealing. It measures output per worker, the right question for aggregate economic productivity. But it's the wrong metric for the AI use cases where gains are actually accumulating. AI is reducing hours per task, not necessarily tasks per employee. If a marketer using AI cuts their copywriting time by 60% but uses that time to run more campaigns, sales per employee may not change even as unit economics improve sharply.

Goldman Sachs Research estimates that generative AI will raise labor productivity in developed markets by 15% when fully adopted into regular production workflows. The qualifier "when fully adopted" is carrying most of the weight in that sentence. Goldman's baseline adoption timeline is approximately 10 years from now. The NBER survey's three-year review window catches AI at the beginning of that timeline, before the workflow reorganization that produces the productivity signal has occurred at scale.

The 1.5-hours-per-week usage average is the most diagnostic number in the study. A professional using AI for 1.5 hours weekly, spread across 40 working hours, is integrating AI into 3.75% of their working time. A 15% productivity gain assumes full workflow integration. Getting from 3.75% to full integration is an organizational change management problem, not an AI capability problem. The technology has outpaced the institutions deploying it, which is exactly the dynamic that preceded the productivity surge of the 1990s computing revolution.

The uncomfortable conclusion: the 89% zero-impact finding is accurate and temporary at the same time. It's accurate because most organizations haven't changed their workflows, incentive structures, or measurement systems to capture AI gains. It's temporary because competitive pressure from the 20% of firms that have restructured will force the rest to adapt or exit their markets. The paradox resolves not when AI improves but when organizational behavior changes around AI. That transition is underway at the high end. It hasn't reached the median firm yet.

What to Watch Next

Track the NBER researchers' follow-up survey, expected in Q4 2026, which will extend the review window to four years and incorporate data from the Q1 2026 AI agent adoption surge. If the 89% no-impact figure doesn't decline meaningfully, it suggests the productivity paradox may extend further than Goldman's 10-year adoption timeline assumes. A decline to 70% or below would signal that organizational adaptation has begun at scale across the median firm.

Watch the PwC AI performance cohort split quarterly. If the 20/74 ratio shifts to 15% of companies capturing 80% of gains, it signals that AI productivity advantage is becoming winner-take-most at the industry level, not just within firms. At that point, companies in the 80% aren't just leaving productivity on the table. They're ceding market position that may not be recoverable once the leaders have locked in compounding advantages.

The Goldman Sachs estimate of 6,000 AI-attributed job losses per month provides the labor market baseline to watch. If Q2 and Q3 2026 AI-attributed layoffs exceed that estimate while the NBER's productivity figure stays flat, it confirms the most troubling scenario: displacement is accelerating faster than productivity gains are materializing, which is a distributional outcome with political consequences that markets aren't pricing.

Finally, watch Anthropic's Claude Code and OpenAI's Codex for enterprise adoption metrics in software development. Coding is the sector where AI productivity gains are most measurable and most established. If enterprise development productivity, measured as features shipped per developer per quarter, becomes broadly visible in public company earnings calls in H2 2026, software will be the first sector to exit the 89% no-gain cohort at scale. The template that emerges there will determine how quickly every other sector follows over the next three years.

The productivity gains are real. The organizational change required to capture them isn't happening at 1.5 hours per week.

---