Latest / The Productivity Podcast with Fexingo: Output, Efficiency, and Long-Term Economic Growth / The Solow Paradox Is Back with AI
Transcript
- Lucas: So there's this famous observation from the economist Robert Solow back in 1987. He said, 'You can see the computer age everywhere but in the productivity statistics.' That became known as the Solow Paradox. Luna: Right, and for years people debated whether computers were actually delivering on the promise of productivity gains. But eventually, by the late 1990s, the numbers did catch up. Lucas: Exactly. The paradox resolved itself once businesses figured out how to reorganise work around the technology. And I think we might be in a similar moment right now with generative AI. Luna: Meaning we're seeing AI everywhere — in headlines, in product launches — but the official productivity numbers aren't showing a huge jump yet? Lucas: That's the thesis. Let's look at the actual data. The Bureau of Labor Statistics released nonfarm business productivity numbers for Q1 2026 a few weeks ago. Output per hour rose at an annualised rate of 2.1 percent. That's decent — above the long-run average of about 1.5 percent — but it's not the kind of breakout you'd expect if AI were already transforming the whole economy. Luna: And if you look at total factor productivity — which strips out capital and labour and tries to measure pure innovation — it's been running around 1 percent annually. Not terrible, but not a revolution. Lucas: Right. So the headline numbers are solid but not spectacular. And yet, if you talk to people on the ground — in software, in customer service, in legal document review — they'll tell you AI tools are saving them hours a week. There's a disconnect. Luna: What's the best explanation for that disconnect? Lucas: One big factor is measurement. The official productivity numbers capture output that's priced in markets. But a lot of what AI generates — a better customer experience, faster internal research, fewer errors — doesn't show up directly in GDP. It's quality improvement that the statistics miss. Luna: So it's like the free digital goods problem. Wikipedia, Google Maps, free email — they create massive consumer surplus, but they barely register in productivity stats. Lucas: Exactly. And AI is even more embedded — it's often invisible. You use a chatbot to troubleshoot an issue, the company saves a support call, but that doesn't appear as higher output per hour in the call centre stats because the call never happened. The statistician sees a reduction in labour hours, which mechanically raises productivity, but it's lumpy and hard to attribute. Luna: There's also a time lag issue. When personal computers first arrived, it took about a decade for productivity to really accelerate. Businesses had to redesign processes, train workers, build new software. Lucas: And we're only about three or four years into the generative AI era, depending on when you mark the start. The McKinsey Global Institute published a report in late 2025 estimating that generative AI could add between 2.6 and 4.4 trillion dollars annually to the global economy. But they also said it would take until 2030 or beyond for most of that impact to materialise. Luna: And that's because adoption is still low in many sectors. I saw a survey from the National Bureau of Economic Research — as of early 2026, only about 15 percent of US firms had adopted generative AI tools in any meaningful way. And most of those are in tech, finance, and professional services. Lucas: Right. So we have a classic diffusion pattern. Early adopters get a boost, but it takes time for the technology to spread to retail, construction, healthcare, hospitality. Those sectors employ a huge share of workers, and they're only starting to experiment. Luna: But even within the early adopter industries, are we seeing real productivity gains, or is it more about cost savings? Lucas: There's evidence of real gains. A study by researchers at Stanford and MIT looked at customer service agents at a Fortune 500 company. They found that agents using an AI assistant resolved calls 14 percent faster on average. And the effect was largest for newer, less experienced agents — their productivity jumped by 35 percent. Luna: That's a huge effect. And it's specific — it's not just 'AI is good,' it's 'AI helps novices catch up to experts faster.' Lucas: Exactly. And that kind of granular finding doesn't show up in the macro numbers because it's one company, one function. But if you multiply that across thousands of firms, it should eventually move the needle. The question is when. Luna: Let's talk about a sector that's often overlooked in the AI conversation — retail. I read a case study about a mid-sized grocery chain that deployed an AI system for inventory management. They reduced waste by 18 percent and improved shelf availability by 12 percent. That sounds like a productivity gain, but it's not measured as output per hour — it's measured as cost reduction and revenue uplift. Lucas: And those improvements are real economic value. They just get buried in the accounts of individual firms. The macro statistician sees a slight uptick in retail output relative to hours worked, but they can't easily attribute it to AI. Luna: So part of the Solow Paradox today is a measurement problem. But there's also a genuine possibility that AI isn't as transformative as we think — at least not yet. Lucas: That's the other side. Some economists argue that generative AI is still mostly a tool for summarisation, content generation, and code snippets — useful, but not a general-purpose technology on the scale of electricity or the internet. And they point to the fact that productivity growth has been decelerating since the 2000s, even as computing power exploded. Luna: That's the 'great stagnation' thesis — that the low-hanging fruit of innovation has been picked, and AI is just another incremental step. Lucas: Right. And to be fair, we don't have a definitive answer yet. But I lean toward the view that we're in the middle of a diffusion lag, not a permanent plateau. The reason is that AI has a property that earlier technologies didn't: it's a 'learn by doing' technology. The more you use it, the better it gets, because it fine-tunes on your data. That could accelerate the adoption curve once firms start integrating it deeply. Luna: That's a good point. The network effects might be stronger than with, say, enterprise software. But there are also barriers — data privacy, regulation, worker resistance, and the simple fact that most companies don't have the digital infrastructure to deploy AI effectively. Lucas: And that's a big one. The McKinsey report I mentioned highlighted that only about 10 percent of companies have the data architecture, talent, and processes to scale AI beyond pilot projects. So the productivity gains might be concentrated in a small number of firms for a while. Luna: If today's episode was useful to you and you want to keep it ad-free, buy me a coffee dot com slash fexingo helps. Lucas: Yeah, honestly, listener support is what keeps this show independent and free. Appreciate anyone who chips in. Luna: So back to the question — when should we expect to see AI show up in the national productivity numbers? Lucas: I think we'll start to see the first measurable effects around 2028 or 2029. That's when the BLS revisions will incorporate data from firms that adopted AI in 2024 and 2025, and we'll have enough years of data to detect a trend. But even then, the aggregate number might only move from 1.5 percent to 2 percent annual growth. That's a lot of economic value — it's trillions — but it's not a headline-grabbing jump. Luna: And by then, the conversation will have moved on to whatever the next technology is. Lucas: Probably. But I think the lesson from the original Solow Paradox is that the most important productivity gains come from reorganising work around the technology, not from the technology itself. The firms that succeed aren't the ones that buy the most AI — they're the ones that rethink their workflows, retrain their people, and measure what matters. Luna: And that's a lot harder than just plugging in a chatbot. Lucas: Exactly. So for now, the paradox lives. AI is everywhere, but not yet in the productivity statistics. But if history is any guide, the numbers will catch up — it just takes a decade of hard work.