The NBER Just Checked the Receipts on Enterprise AI. The Math Is Bad.

The National Bureau of Economic Research just released a working paper that should make every C-suite executive in the country pretty uncomfortable. For the last two years, the corporate world has been operating on a diet of espresso, FOMO, and generative AI promises. We've seen billion dollar valuations for startups with nothing but a wrapper and a dream. We've seen legacy enterprises announce massive AI pivots to satisfy shareholders who are terrified of being left behind. But the NBER just checked the receipts, and the math doesn't look good for the hype cycle.

"Firm Data on AI" is the first representative international survey of its kind. Researchers from the Atlanta Fed, the Bank of England, the Bundesbank, Stanford, and Macquarie University surveyed nearly 6,000 CEOs, CFOs, and senior executives across the US, UK, Germany, and Australia. This isn't a vibes-based McKinsey deck. It's the closest thing we have to ground truth.

What the data actually says

69% of firms are actively using AI. Adoption climbed from roughly 61% to 71% in the twelve months between early 2025 and early 2026. That's a remarkable diffusion curve for a technology still finding its shape.

But over 80% of those same firms report no measurable impact on either employment or productivity over the past three years. Adoption is sprinting while impact is sitting on the couch. We're spending billions of dollars to achieve a result that effectively rounds down to zero.

The forecast isn't the rescue some headlines have made it out to be. Executives expect AI to boost productivity by 1.4% and output by 0.8% over the next three years. That is not a boom; it's a rounding error dressed up in a suit. CEPR's own write-up of the paper ran under the headline "Firms predict an AI productivity boom is coming," which tells you everything you need to know about how the optimist class is grading on a curve. Fortune covered the same data with the opposite spin: "Thousands of CEOs admit AI had no impact on employment or productivity." Both framings miss the point.

The two dominant narratives are both wrong

The skeptics treat this as proof that the AI bubble is about to burst. They look at the lack of immediate ROI and return a verdict that the technology is a parlor trick that can't survive the harsh light of a balance sheet.

The optimists do the opposite. They cling to the three-year forecast as if 1.4% is some kind of moonshot, and they treat the current zero as a minor scheduling conflict rather than a systemic failure.

Both interpretations fail because they treat AI as a standalone product that you simply plug into a company to get results. They assume the software is the bottleneck. The data suggests something much more damning. The bottleneck isn't the model. The bottleneck is the organization itself.

Leadership is leading from '90 minutes a week' away

There are pockets where this matters less.  Engineering teams running coding agents against well-structured repos, security teams pointing models at SIEM data, analysts working over clean warehouses. These groups see real lift because they're feeding the AI artifacts that already have governance, structure, and version control baked in. That's not most of the enterprise. Most of the enterprise runs on knowledge work, and knowledge work runs on email threads, Word docs, and Confluence pages nobody has updated since the pandemic.

Consider how leadership actually interacts with that side of the business. The paper finds that more than two-thirds of executives use AI in a typical week, but their average use is just 1.5 hours. One in four executives doesn't use it at all.

These are the same leaders who tell boards they're transforming their companies into AI-first organizations. You can't lead a transformation of this magnitude on 90 minutes of weekly tinkering, and you certainly can't lead it from the sidelines. AI is being treated as a novelty or a high-level research assistant rather than a fundamental change to the way work gets done. It's a peripheral activity for the people who are supposedly driving the strategy.

This usage gap explains why the productivity gains are stalled. If the people at the top aren't living in the tools, they can't see why the tools are failing. They don't see the friction. They don't see the hallucinations. Most importantly, they don't see the content debt that's poisoning every prompt their employees write.

The companion NBER paper released alongside it formalizes what's happening here. The authors document an explicit "productivity paradox" in which executives' perceived gains run well ahead of their measured gains. Leaders feel more productive because they can summarize a long email thread in 10 seconds. That feeling doesn't translate to the bottom line because the core work of the business is still buried under a mountain of unstructured, unverified, and/or disconnected content.

We have seen this movie before

This isn't new economic territory. Back in 1987, the economist Robert Solow famously quipped that you can see the computer age everywhere except in the productivity statistics. It took roughly fifteen years for that paradox to resolve itself, and when it did, it wasn't because the computers got better. It was because companies finally did the slow, expensive, deeply unsexy work of changing how they actually operated around the technology.

The NBER authors nod at this lineage directly when they note that transformative technologies are widely viewed as important well before their effects are fully reflected in measured productivity. The optimists aren't wrong that gains are coming. They're wrong about who gets them. The firms that eventually capture the upside won't be the ones with the best prompts. They'll be the ones that did the plumbing while everyone else was buying licenses.

The independent evidence is worse

The data from outside the NBER stack tells the same story, only more starkly. MIT's Project NANDA released "The GenAI Divide: State of AI in Business 2025" in July of last year. Their finding: despite $30 billion to $40 billion in enterprise GenAI investment, 95% of organizations are seeing zero measurable financial return. NANDA explicitly attributes the failure not to model quality but to a "learning gap": brittle workflows, weak contextual learning, and misalignment with day-to-day operations. NBER says 80%. MIT says 95%. The numbers converge on the same diagnosis.

The UK government's three-month trial of Microsoft 365 Copilot at the Department for Business and Trade is even more damning because it isn't a survey. It's a controlled deployment. 1,000 licenses. 300 consenting participants. The evaluation found that users were satisfied, time savings appeared on individual text-based tasks, and yet the report's own conclusion was that they didn't find robust evidence to suggest those time savings were leading to improved productivity. Worse, Copilot users produced less accurate Excel analysis and less accurate PowerPoint contents than non-users. Faster work, lower quality, no productivity gain. That's the implementation gap in laboratory conditions.

The MIT data also surfaces a finding nobody in the press is talking about: 90% of workers report regularly using personal AI tools at work, while only 40% of companies have official enterprise subscriptions. That's a shadow AI economy running parallel to the sanctioned one, and it's happening for a reason. The sanctioned tools are connected to the company's content. The company's content is a disaster. So employees route around it, paste sensitive material into ChatGPT, and call it productivity. This isn't a training problem. It's a content governance crisis disguised as user behavior.

Content debt is the hidden tax on every AI initiative

We talk about AI readiness as if it's a binary state you achieve by buying enough licenses. The reality is simpler and harder. Business AI requires business context, and business context is business content. 

Most companies are currently disqualified from AI success because their content is a mess. If your organization has 20 years of legacy documents scattered across five platforms with no consistent metadata and zero security governance, your AI is essentially a genius trapped in a library where the books are written in crayon and half the pages are missing. And if you've consolidated to one platform with some governance bolted on, you're not exempt. You're just earlier in the same library, with better lighting, and the books are still mislabeled. You can ask it a question, but you shouldn't bet your business on the answer. 

This is the penalty of content debt. It's the hidden tax on every AI initiative. When we at TEAM talk about the path to the Zero Click Enterprise, we're talking about paying that debt down. You'll never eliminate it entirely. The point is to work it down to a level where AI can actually start creating value instead of fabricating it. The vision of a system that anticipates goals and executes workflows without manual intervention is impossible in the current state of the enterprise. You can't have a zero click outcome if a human has to spend minutes verifying that the AI used the correct version of a contract.

The transition from a passive chat interface to a proactive execution engine (the state we're careening towards) requires a foundational overhaul that most companies are trying to skip. They want the intelligence without the plumbing. They want the agentic enterprise but they're still struggling with basic version control. The NBER data shows that the honeymoon phase is over. The novelty has worn off, and the reality of the implementation gap is setting in.

We see this as an infrastructure crisis rather than a technology failure. The reason 80% of firms see no ROI is that they're trying to use 21st century intelligence to navigate 20th century file shares. They're layering sophisticated models on top of a fragmented mess and wondering why the results are inconsistent. If you want to move the needle on productivity, you have to stop thinking about AI as a tool and start thinking about your content as an asset that needs to be prepared for automation.

What "AI ready" actually looks like

The media likes to frame the productivity paradox as a mystery of modern economics. It isn't a mystery. It's a predictable result of trying to automate chaos. You can't automate a process you don't understand, and you can't gain insights from data that's inaccessible or incorrect. The firms that eventually see real gains won't be the ones with the best prompts. They'll be the ones that had the discipline to fix their content layer before they started calling everything an agent.

The NBER report should be a wake-up call for any leader who thinks they can buy their way into the future. The data shows that the current approach isn't working. High adoption rates mean nothing if the technology is sitting on the periphery of the business, and they mean even less if executives are spending 90 minutes a week with the tools they're betting the company on.

There are no shortcuts. There are no magic models that bypass the need for a solid technical foundation. If you're a leader reading this, the next move isn't another pilot. It's an honest content audit. It's metadata strategy. It's governance for the orphaned SharePoint sites and abandoned Teams channels that nobody owns and everybody searches. It's putting policy around the shadow AI your employees are already using whether you sanctioned it or not. It's putting the plumbing in before you call yourself an AI-first organization.

The NBER has provided the diagnosis. The problem isn't the AI. The problem is the content debt that keeps the AI from being useful. Until that's addressed, the Zero Click Enterprise will remain a theoretical concept for the companies that refuse to attend to the plumbing first. The future is coming, but it isn't going to wait for you to find your orphaned SharePoint sites. It's time to stop playing with the tools and start building the engine.