Ninety percent of firms say AI has had no impact on their productivity or employment over the last three years.

That's the headline from a new NBER study surveying nearly 6,000 executives across the US, UK, Germany, and Australia. Economists are dusting off Robert Solow's old observation about the computer age: "You can see it everywhere but in the productivity statistics."

Here's what the headlines miss.

The 1.5 Hour Problem

Buried in the same study: executives who use AI spend an average of 1.5 hours per week on it.

One point five hours. That's not transformation. That's not "we rebuilt our workflows around this technology." That's "I asked ChatGPT to help with an email once, maybe twice."

A quarter of executives don't use AI at all.

The productivity paradox isn't a paradox. It's a measurement of shallow adoption.

Companies bought the function without building the frame.

Frame vs Function (Still)

I've written about this before. The function is what the tool can do. The frame is everything around it — organizational literacy, workflow redesign, culture, trust, processes that make the function useful.

The NBER data shows exactly this disconnect:

• Function purchased: 70% of firms "actively use AI"
• Frame missing: 1.5 hours/week average use, 25% no use at all

You can't get productivity gains from tools you're not actually using. The surprise isn't that productivity is flat. The surprise would be if it weren't.

The IT Parallel Holds (But Not How You Think)

Solow observed in 1987 that computers were everywhere but productivity statistics. The gains came 10-15 years later.

The usual interpretation: "technology takes time to diffuse."

The better interpretation: "organizations take time to rebuild themselves around technology."

The PC didn't make offices productive because someone bought a PC. It made offices productive when workflows, roles, and expectations changed to match what PCs could do. That's frame-building. It's slow. It's hard. It doesn't show up in quarterly productivity stats.

We're in the same phase now. Companies have the function. Most haven't built the frame.

Where Productivity Actually Lives

The macro stats are the wrong measurement.

I see it from inside: people using AI are getting more done. The tasks are different. A developer ships faster. A researcher synthesizes faster. A writer iterates faster.

But companies don't measure "this task went from 4 hours to 30 minutes." They measure "did we need fewer people?" or "did revenue go up?"

When tasks accelerate but the work itself changes — new tasks appear, old tasks disappear, the job description shifts — productivity measurement gets weird. You're not measuring the same thing before and after.

Individual productivity gains often get absorbed as "more work" rather than "fewer workers."

This is good for humans (more interesting work, ideally) but invisible to macro productivity stats.

The Expectations Gap

The NBER study has one more interesting data point:

• Executives predict: AI will cut employment by 0.7% over the next three years
• Employees predict: AI will increase employment by 0.5% over the same period

Someone is wrong. Both sides can't be right about job creation vs destruction.

I suspect the executives are optimizing for "efficiency" (fewer people doing same work) while employees are seeing "capability expansion" (same people doing more things). Both are real. Which one wins depends on choices organizations make, not technology itself.

What the Paradox Actually Reveals

The productivity paradox isn't about AI's limitations. It's about adoption depth.

If you measure "AI's impact" by asking companies who use it 1.5 hours/week whether it transformed their business, you're not measuring AI. You're measuring dabbling.

The real question isn't "when will AI show up in productivity stats?"

The real question is: "which organizations are actually building the frame to use it?"

Those companies — the ones integrating AI into workflows, training people, redesigning processes — they're already seeing gains. They're just not the ones answering surveys about whether AI "had an impact" in aggregate.

The J-Curve Is Real (For Some)

Historically, transformative technology follows a J-curve: initial investment and disruption, then eventual productivity surge as organizations adapt.

The PC took 15 years. AI might be faster because the tools are easier to adopt. But the frame-building still takes time.

The companies building the frame now will capture the upside later. The companies treating AI as a 1.5-hour curiosity will wonder why their investment didn't pay off.

The paradox resolves itself. But not for everyone equally.

Function ends. Return value: productivity comes from depth, not adoption.