The Wrong Paradox | Eliran Keren

In 1987, economist Robert Solow wrote: “You can see the computer age everywhere but in the productivity statistics.”

That quip — now called the Solow paradox — is back in circulation.

Last week, the San Francisco Federal Reserve published a research brief arguing that we may be living through it again. Individual workers are measurably faster with AI. But total factor productivity — the broad measure of how efficiently the economy converts inputs to output — has barely budged. The Fed researchers say give it time. The 1990s had the same lag. The boom came later.

That’s a reasonable argument. It’s also missing something.

What the 90s actually solved

The productivity boom that followed the Solow paradox didn’t happen because workers kept getting better at using their computers. It happened because the Internet came and rewired the underlying structure of how the economy coordinates.

Search made finding information essentially free. Email collapsed the cost of reaching people. E-commerce removed the friction from transactions. The entire cost of moving information, decisions, and agreements between people dropped by orders of magnitude.

The Internet was a coordination technology. That’s what made it productivity-relevant at scale.

When you lower the cost of coordination, you reduce the overhead that sits above every individual task. Less time spent figuring out who has the right document. Less friction finding the person who made the decision. Fewer meetings to re-establish context everyone once had. The Internet didn’t make individual workers faster at their specific tasks. It made the space between tasks cheaper.

That’s the mechanism the SF Fed’s analogy implies will repeat.

But it won’t. Not for the same reason.

What AI actually speeds up

AI tools make individual tasks faster. That part is real and well-documented.

Harvard Business Review published a study in February of 200 employees at a U.S. technology firm who used AI tools on their work. The result: workers completed tasks faster. Time saved was measurable and genuine.

Then the researchers kept watching.

The time savings didn’t become rest, or creative work, or strategic thinking. The time immediately refilled with more tasks. Workers who adopted AI tools reported working more hours by the end of the year, not fewer. Fewer breaks overall. Higher risk of burnout. The HBR summary: “AI doesn’t reduce work. It intensifies it.”

Read that slowly. AI makes tasks faster. The number of tasks doesn’t decrease. It increases.

Separately, a BCG study from March of 1,488 workers found that extensive AI use is generating what researchers are calling “brain fry” — cognitive overload from monitoring, evaluating, and managing AI outputs. 14% of AI users report it. Among marketing teams, 26%. The burden isn’t the task completion. It’s the overhead of working alongside AI: reviewing its outputs, catching its errors, deciding which outputs to use, integrating results into workflows that weren’t designed for machine-generated inputs.

The pattern: individual execution got faster. Everything above execution got heavier.

The overhead that isn’t in the data

There’s a number the productivity statistics don’t capture.

The SF Fed measures output per labor hour. What it can’t measure is cognitive overhead per decision — the mental cost of figuring out which tasks matter, what’s still open, what changed since yesterday, what someone committed to last week that hasn’t been tracked anywhere.

This is the coordination layer. It sits above execution. It’s the work-about-work that everyone complains about and nobody has solved.

In the 1990s, the Internet made this coordination layer cheaper. You could find the document. You could reach the person. You could check the status without scheduling a meeting. Coordination friction fell.

AI does nothing to this layer. In some ways, it adds to it.

More AI tools generating more outputs means more things to integrate, more decisions to make, more places to check. The individual tasks complete faster. The question of which tasks to do, in what order, based on what’s actually open — that question is structurally unchanged. It might be more demanding than it was before. You have more outputs to synthesize, not fewer.

This is why the 40% of workers who report no time saved at all from AI tools aren’t confused or wrong about their experience. The time saved at the task level is being absorbed by the overhead above it. Not because they’re using AI badly. Because the layer above execution is the one that’s carrying the load.

Where the data points

Three separate data points from the past 90 days, from three separate institutions:

The SF Fed says AI productivity is real but hasn’t shown up at scale — the lag mirrors the 1990s Internet adoption curve.

HBR says workers using AI are working more, not less — the time savings get consumed.

BCG says cognitive load from AI monitoring is rising, not falling — 33% more decision fatigue among affected workers.

These aren’t contradictions. They’re the same thing measured at different layers.

Individual task velocity is up. Total cognitive burden is up. Economy-wide productivity is flat.

The Solow paradox was resolved when a new technology reduced coordination costs. The current paradox will be resolved when a new layer reduces coordination costs again. Not by making individual tasks faster. By knowing which tasks need to happen at all.

What the analogy actually predicts

Here’s the part the SF Fed got right without knowing it.

If you take the Solow analogy seriously — not just the lag but the mechanism — it predicts something specific. The Internet productivity boom wasn’t powered by faster word processing or better spreadsheets. It was powered by the emergence of new infrastructure that made coordination fundamentally cheaper. The tools that did that — search, email, e-commerce — weren’t extensions of what people already had. They were a new layer.

The AI productivity boom, if it comes, won’t be powered by faster task execution either.

It will come when someone builds the layer that knows what’s actually open. What decisions are pending. What’s been resolved. What needs the human and what doesn’t. A system that doesn’t require direction on demand — because it already has enough context to provide it.

That layer doesn’t exist yet at scale. Individual workers and teams are building versions of it for themselves. The enterprise software industry is beginning to name it — Asana just acquired a company to reposition around it, announcing they are becoming “the coordination layer that makes human-agent collaboration work at enterprise scale.”

The vocabulary is being found. The infrastructure isn’t built.

The Solow paradox was solved by the Internet. The AI productivity paradox will be solved by something that does for cognitive coordination what the Internet did for information access.

That’s not a prediction about timeline. It’s a description of the structural gap that’s producing the data the SF Fed is trying to explain.

Eliran Keren — Founder of Deeplica, building the coordination layer for knowledge work.

Sources: SF Fed — Have We Entered an Era of High Productivity Growth? · Fortune — Employees using AI are working faster, but the economy isn’t more efficient · HBR — AI Doesn’t Reduce Work, It Intensifies It · BCG — When Using AI Leads to Brain Fry · Fortune — Asana bets future on human-agent coordination