A working paper from the Federal Reserve Bank of Atlanta, co-authored with researchers at the University of Pittsburgh and later amplified by Fortune in August 2026, puts a number on something most business leaders already sense: roughly 90% of executives say AI has not yet produced meaningful productivity gains at their organisations.
The paper draws on two linked surveys. The first polled around 750 CFOs and senior executives across the United States in late 2025. The second ran a parallel international survey across four countries with nearly 6,000 respondents. Both found the same pattern: widespread AI investment, very little measurable return so far.
That matters enormously given what else is happening in parallel. AI-cited layoffs now account for roughly one in four announced job cuts in the United States in 2026. Companies are trimming headcount with one hand while reporting no productivity benefit with the other.
What the Data Actually Shows
The Atlanta Fed paper is careful not to say AI is failing. It says the gains haven’t arrived yet, and the reasons are identifiable.
More than 80% of firms reported no impact on employment or productivity over the previous three years. Yet 60% of responding companies have already invested in AI, with adoption reaching 80% at large firms. Large companies are spending an average of 55% of their AI budget on operational costs (subscriptions, services, training), 31% on internal development of customised systems, and 13% on hardware and infrastructure.
The paper does find positive signals. Productivity improvements, where they exist, are concentrated in high-skill services and finance. These are sectors where employees already have the analytical foundations to put AI tools to work. The gains in those sectors are projected to grow through 2026 and into 2027.
Employment effects are smaller than the headlines suggest. The researchers estimate AI will reduce aggregate employment by less than 0.4% in 2026. What they do find is a shift in the composition of work: fewer routine clerical roles, more skilled technical roles.
The Real Problem Is Underneath the Headlines
Here is the uncomfortable truth buried inside the data: the companies seeing productivity gains from AI are not the ones spending the most. They are the ones where employees understand data, can evaluate AI outputs critically, and know how to redesign workflows rather than just attach an AI tool on top of old ones.
The Fortune coverage that surfaced these findings again in late August 2026 also pointed to research from Wharton and Boston University showing that AI-related layoffs are actively damaging the conditions needed for AI to deliver returns. When employees feel their jobs are threatened by the tools they are supposed to use, adoption stalls and productivity falls further.
This creates a loop that many organisations have not recognised yet. Cut the people who know how the data flows, and you have also cut the people who would have made AI work.
Why the Gap Exists
There are a few structural reasons why most organisations are not seeing returns.
First, AI tools require good data to function well. If the underlying data is messy, inconsistent, or siloed across systems, AI amplifies those problems rather than solving them. Most organisations have not done the foundational data work.
Second, deploying an AI tool is not the same as changing how work gets done. Giving a team access to a language model without training them to evaluate its outputs, redesign their processes, or audit for errors produces very little. The tool becomes expensive automation that produces plausible-looking but unreliable results.
Third, and perhaps most important, the productivity gains require a workforce that can work with AI rather than alongside it. That means data literacy, prompt engineering literacy, and the judgment to know when to override an AI output. These are skills that are built, not assumed.
What This Means for Business Leaders
If you are a business leader looking at this data and wondering what your organisation is missing, the answer in most cases is not a different AI vendor. It is the capability layer that makes AI investments pay off.
The organisations outperforming in the Atlanta Fed data share one trait: they treat AI as a capability-building exercise, not a cost-cutting exercise. They invest in training alongside tooling. They build data infrastructure before layering models on top. They upskill their existing workforce rather than replacing it with automation that nobody fully understands.
The 90% finding is not a verdict on AI. It is a diagnosis of implementation gaps. And the treatment is well understood.
Businesses that have built the analytical foundations, trained their teams on data skills, and engaged the right advisory support are seeing compounding returns. Those that cut corners on capability development are the ones reporting no gains.
The gap between those two groups is widening every quarter. The companies that close it in the next twelve months will have a structural advantage heading into 2027.
Enterprise DNA’s learning platform helps organisations build the data and AI capability needed to turn AI investment into measurable returns. For businesses ready to go further, Omni Advisory provides fractional AI strategy support for leaders navigating exactly this challenge.