Why do 89% of executives say AI has had no impact on productivity?

The NBER 6,000-exec study: why AI is invisible at the top

A July 2026 NBER working paper surveying 6,000 executives found 89% of firms report no measurable impact of AI on labor productivity over three years, despite 69% of firms actively using AI. Paul Daugherty and H. James Wilson's Radically Human explains why: the 10% that see gains redesigned their architecture around human judgment (IDEAS framework), not the other way around.

I have spent the last two weeks reading through the same NBER working paper my founder-clients keep sending me, and the number in the headline is not the interesting part. The interesting part is why one in ten firms is seeing real gains while the other nine in ten are not — because the pattern that separates them is repeatable, and it is not the pattern most executives are chasing this quarter.

Here are the numbers, dated. The 2026 NBER working paper surveyed nearly 6,000 executives across the US, UK, Germany, and Australia. It found that while 69% of firms are actively using AI, 89% of executives report no measurable impact on labor productivity (sales per employee) over the past three years, and more than 90% report no measurable impact on employment. Deloitte's parallel July 2026 data corroborates it: only 10% of executives are currently seeing an ROI on their AI spend. Meanwhile, the 2026 State of AI Agents report from Kore.ai found that 88% of AI-agent pilots fail to reach production due to evaluation gaps and governance friction. This is happening against a backdrop of global AI spending projected to surpass $2.5 trillion in 2026. The scale of the gap between capital deployed and results captured is the largest I have seen in fifteen years of watching enterprise technology adoption.

What Radically Human predicted (and the 89% missed)

Paul Daugherty and H. James Wilson's Radically Human laid this out in 2022, and it is uncanny how well it maps the current numbers. Their thesis: companies that treat AI as an extension of the old legacy stack — bolt-on chatbots, back-office automation, "efficiency gains" — will see nothing at the top line, because AI does not scale that way. Their IDEAS framework (Intelligence, Data, Expertise, Architecture, Strategy) argues that productivity only shows up when you invert conventional assumptions: intelligence becomes more human-like (not less), expertise flows from humans to machines through teaching (not just data ingestion), and architecture becomes living and boundaryless (not layered on top of ERP). The 18% they call "leapfroggers" grew revenue at roughly four times the rate of laggards — a suspiciously similar fraction to today's 10-18% who are actually capturing AI ROI.

The three failure modes I keep seeing in founder sessions

Across the last two months I have run around fifteen sessions with founders and one C-level operator whose companies are in the 89%, and the pattern is consistent. First, AI is treated as a productivity tool, not a re-architecture. Ninety-three per cent of enterprise AI teams exceed budget on "response refinement" — burning cycles polishing agent outputs that were pointed at the wrong problem in the first place. Second, machine teaching does not happen. In Daugherty and Wilson's language, the frontline expertise stays locked in people; the AI stays generic; the specific competitive edge never gets built. Third, governance is bolted on last, so pilots die at the production gate — which is exactly why 88% of agent pilots fail to graduate.

What the 10% actually did — a leadership decision, not a tools decision

The firms capturing AI ROI in July 2026 are not the ones with the biggest OpenAI or Anthropic bill. They are the ones where the CEO or founder personally owned the AI architecture decision for six months — the machine teaching, the workflows that were redesigned, the metrics that were changed. NBER data shows 72% of CEOs now identify as their company's main AI decision-maker, a doubling from last year, and this correlates strongly with the leapfrog cohort. Sixty-two percent of enterprises now employ a dedicated agentic-ops lead — and that role correlates directly with successful production deployment. This is not a coincidence. It is what Daugherty and Wilson meant by strategy merging with execution: the leadership work and the implementation work stopped being sequential and started being the same work.

The three-question audit I now run before any AI investment recommendation

These are the three questions I ask a founder before we spend another dollar on AI tooling — they are the practitioner version of Daugherty and Wilson's IDEAS framework compressed into a decision filter. First: where is the expert knowledge in this company right now, and what is the plan for teaching it to a model? If the answer is "we will use ChatGPT," the productivity gain will be zero at the firm level, exactly as the NBER data predicts. Second: which existing workflow will we deliberately break and rebuild around AI, and who owns that architecture decision at C-level? No accountable architect equals no leapfrog. Third: how will we measure this in six months in a number that shows up on the P&L? Not hours saved. Not messages generated. Revenue per employee, gross margin, or cycle time on a decision that actually moves the business.

The honest limit: the audit above is easier to write than to run. Most founders discover on question one that their expert knowledge is in three people's heads and has never been documented, let alone taught to a model. That is the actual work. The tool choice is the last 10% of the decision, not the first. The NBER 89% number is not really about AI — it is about the number of firms willing to do the slow, unglamorous re-architecture work that AI value has always required. That has not changed since 2022 when Daugherty and Wilson wrote the book. What has changed is that the capital is now committed and the results are due. The reckoning is here.

Sources

NBER working paper via CEPR VoxEU column "AI, productivity, and work: Evidence from US firms" (2026); Saner.ai blog, "AI at Work Statistics 2026" and "AI Assistant Statistics 2026" (July 2026); AI Business Weekly, "AI Productivity Statistics 2026" (July 2026); Business Insider coverage of the AllianceBernstein CEO comments on the NBER paper (June-July 2026); Kore.ai, "AI Agents in 2026: From Hype to Enterprise Reality" (July 2026); MarketScale on Ninestone Research (2026); Paul R. Daugherty & H. James Wilson, Radically Human (HBR Press, 2022).