Learn to fly

Read the full article (PDF)open_in_new

The AI productivity J curve and why the data are still idling on the runway

« I’m looking to the sky to save me,
Looking for a sign of life,
Looking for something to help me burn out bright. ”
– Foo Fighters

Executive summary

The legendary Yogi Berra was right in saying that “It’s tough to make predictions, especially about the future.” Rarely has this quotation been more appropriate than when we try to predict the future of a world with artificial intelligence (AI).

But history, as always, can give us a flight path. As has been the case time and again, new technologies arrive with promise, capital pours in, and everyone starts looking to the sky for signs of takeoff. Then the productivity numbers shrug. That gap between excitement and statistics isn’t new. It’s often the price of learning to fly.

Continuing our previous piece about AI and the future of work (found here), this paper argues three points.

First, productivity is hard to measure and goes beyond output per hour worked. Digging through the history of economic growth, more specifically the broad concept of total factor productivity (TFP), a measure of efficiency and technology that includes how well labour, capital, and organization work together, we find that technology alone doesn’t lead to productivity growth; rather, the growth comes from societal reorganization around the technology and the new tasks that it creates.

Second, history suggests that technological revolutions often create a productivity J curve. Installation and disruption come first, followed by deployment and adjustment, then transformation and upside. The delay in the data, when the magic actually happens, is called co-invention. In other words, before seeing radical gains, firms need to redesign workflows, retrain workers, and rethink what they produce, not just bolt the new tool onto old processes.

Third, AI is most likely in the dip of its J curve today. Executives report real productivity gains, but measured outcomes lag. Part of the reason is organizational friction and intangible investment. Another part is measurement bias, especially when hyperscalers build AI infrastructure internally, and when model development is recorded late in national accounts. For investors, the key is to watch for markers that the economy is truly learning to fly, and to be realistic about how quickly takeoff can and will happen.

What is productivity?

Productivity is the simplest concept in economics but the hardest to measure cleanly. At the most basic level, it is how much output we produce for each unit of input. The most common headline metric is GDP per hour worked. It is intuitive and it affects living standards, profits, and the long-run path of interest rates.

But GDP per hour has well-known limitations. It can rise because workers obtain better tools, because capital intensity increases, because the workforce becomes more skilled, or because firms reorganize more effectively. Moreover, it struggles with intangibles, quality improvements, and free digital services. That’s why economists also rely on total factor productivity, which captures the portion of growth not explained by labour and capital inputs. TFP is where we usually find the signature of technology and broad efficiency.

TFP has rarely moved in a straight line through history, nor has it been uniform across countries. Periods of rapid TFP growth tend to coincide with clusters of innovation and diffusion, followed by slowdowns when the easy gains are exhausted or when institutions fail to spread the benefits widely.

Beyond technology: Productivity and society

Before digging into the expected behaviour and impacts of the AI revolution, let’s take a few moments to scope out a vital piece of the puzzle : Technology’s impact on productivity goes well beyond scientific breakthroughs. It has as much, if not more, to do with the societal and organizational changes that the breakthroughs enable.

Electricity wasn’t just a better way to power machines. It changed cities, households, factory design, and labour markets. The automobile didn’t merely replace the horse. It created suburbs, highways, shopping malls, and entirely new industries. Computers didn’t just speed up arithmetic. They rebuilt supply chains, generated new business models, and eventually changed how consumers search, shop, and communicate.

That is why the productivity debate hinges so much on diffusion and institutions. When the surrounding system adapts, productivity surges. If it doesn’t, the technology can be ubiquitous yet fail to show up in the statistics.

Electricity’s long climb to real gains: 1880 to 1940

Electricity is the classic reminder that transformation takes time. It’s tempting to surmise that, with the advent of electricity, productivity jumped immediately. In reality, it took decades for electrification to lift economy -wide productivity because complementary investments and redesigns were necessary.

The household channel mattered immensely. Electric lighting extended productive hours and changed urban life. Appliances reduced the time cost of domestic work, supporting labour force participation (especially for women) and changing the allocation of time across the economy.

Manufacturing was even more instructive. Early factories often used electricity to power the same centralized layouts they had used under steam, essentially swapping motors while keeping the factory floor unchanged.

The major productivity gains came later (as much as 40 years later in some cases), when firms reorganized around unit drive motors, redesigned layouts, and reengineered workflows. In other words, progress is real, but transformation requires starting fresh rather than layering the new technology onto the old world.

As we search the sky for a sign of life in the productivity data, electricity teaches us patience. The sign often comes after the unglamorous work has been done, not during the hype.

The automobile’s redrawing of the map: 1910 to 1955

The automobile wasn’t simply a transport upgrade. It created a major industry, then sparked secondary waves in steel, oil, chemicals, construction, retail, and leisure. Productivity gains in the car industry helped pull along productivity in upstream and downstream sectors. New infrastructure followed. Roads, highways, logistics networks, and suburban housing reshaped consumption patterns and commuting.

This lesson is key for AI. The biggest productivity gains may not come from AI inside existing workflows. They may come from the new industries, the new businesses, and the new consumption possibilities that AI enables.

It’s said that America made the automobile, and then the automobile made America. For AI to come even close to the auto industry’s impact on modern society, it will have to reshape society itself. While we can certainly don our optimist hats and argue that this will be the case, we must admit that the bar has been set high.

Robert Solow’s early but correct take on computers and the internet: 1970 to 2010

The computer and internet era gave us the most famous productivity paradox. As the pathbreaking economist Robert Solow famously wrote in 1987: “You can see the computer age everywhere but in the productivity statistics.” The 1990 s eventually produced a productivity acceleration in North America, but the lag was long enough to confuse observers in real time.

This recent episode offers a policy lesson. In the mid- 1990 s, firms were ramping up IT investment, but official productivity measures did not reflect it clearly. Later, data revisions revealed the acceleration had begun earlier than the initial estimates suggested. The lesson is not that the data are useless. The lesson is that, in a transformation, the right indicators are often disaggregated, and the aggregate statistics can lag the reality on the ground.

This parallel is central to AI today. Many firms report that AI tools save time and improve workflow. Yet productivity metrics remain noisy and contested.

We may be living through Solow paradox 2.0, not because AI is fake, but because the economy is still learning to fly.

Of course, it might be tempting to say that this time will be different, because AI is much more revolutionary than computers and the internet . Still, looking back, we find that productivity was still somewhat slow and generally elusive by historical standards during the computer age.

More precisely, despite the computer age and the wonders of the internet, TFP annual growth in the United States has averaged 1.2% since 1980 , versus 2.1% from the 1940 s to the 1970 s . Again, it’s not just the technology that counts, but its societal impacts.

We can add a modern warning for AI specifically. Even if AI raises output, it can further widen the gap between capital income and labour income. Moreover, some new AI-driven activities may have questionable, if not outright detrimental, social value. Think of the use of AI to sow disinformation or to hack computer systems. In other words, new “bad tasks” could act as productivity counterweights to new “good tasks”.

What is a productivity J curve?

A productivity J curve describes a dynamic adjustment process in which a new technology, policy, or structural change initially reduces measured productivity, then ultimately raises it far above its original level.

The mechanism has three phases.

First comes installation and disruption. At the outset, heavy investment in technology, reorganization, and training is necessary and entails large amounts of time and money. As processes are disrupted and the learning curve dominates, measured productivity falls or stalls. We could argue that we’re still in this phase in mid-2026.

Second comes deployment and adjustment. In this phase, systems start working, skills improve, and complementary changes take hold. As productivity recovers gradually, the gains from the new technology become clearer.

Last comes transformation and upside. Further down the line, full efficiency gains arrive along with new business models, scale effects, and spillovers. Productivity accelerates above trend, and the new technology’s full benefits become widespread.

This is the runway leading up to flight. Many observers look only at the destination, forgetting the takeoff roll.

The AI productivity J curve and why the boom lags the breakthrough

Solow paradox 2.0: perception versus data

A recent piece of evidence from a corporate executive survey displayed in full force the gap between the expectations and the reality of AI’s current impacts.

After surveying nearly 750 corporate executives to study AI’s effects on productivity and the workforce, the Federal Reserve Bank of Atlanta published its findings in March 2026. The conclusions were predictable : Executives report a sizable discrepancy between perceived productivity gains and revenue -based productivity gains. Many executives report meaningful improvements, but the implied gains in measured output per worker are smaller in the near term. The authors interpret this gap as delayed output realization and quality improvements not yet captured in revenue statistics.

That is exactly what a J curve looks like at the firm level. Time is saved first, and money is spent to adopt new technologies. The workforce is forced to adapt, and some firms fire and then rehire some workers. Revenue and measured output follow later, after processes and products change enough to monetize the time savings.

Further down the dip: co-invention costs

Co-invention is the hidden capex cycle where the magic happens.

General-purpose technologies require firms to invest in complementary intangible capital, retraining workers, redesigning workflows, and renegotiating contracts before seeing full returns.

Figuratively, using AI to automate steps in a process is like replacing a steam motor with an electric motor while leaving the factory layout unchanged. It saves time and money but falls short of transformation. The big gains come when firms start fresh and redesign the process itself. This part of the process is what we could label as the largest known unknown: How long will it take corporations to tame the AI beast and adjust so as to unlock its potential? And what will the adjustments ultimately be?

While it’s impossible to know for sure, it’s amply clear that this phase of the process toward productivity gains is vital yet unpredictable.

Timing the turn, the institutional investor’s playbook

It’s tempting to imagine technological adoption as an exponential curve : smooth, relentless, and inevitable. In practice, adoption looks more like an S, with spurts and stalls, false starts, and plenty of turbulence along the way.

The reason is simple : Deploying a general -purpose technology is a sequence of costly choices. Integration takes time, regulations evolve, and the real economy imposes speed limits through scarcity and prices. An obvious roadblock for AI is that computing power, or tokens, is not free – it is in fact highly energy-intensive. When such constraints tighten, the curve bends. When they loosen, momentum returns. That is why the path to scale can look like a series of highs and lows rather than a clean takeoff.

Going from technological innovation to broad -based productivity is a multifaceted problem . We highlight four markers to help assess where the technology is on the curve:

The first is organizational rewiring. Early AI adoption is often scattered, but transformation occurs when AI integrates into core operations. This shift involves redesigning workflows, redistributing tasks between humans and machines, and creating new job roles. The shift becomes clear when firms ask not just where to use AI, but how to build their business around it.

The second marker is the emergence of measurable economic spillovers. At first, AI improves task throughput, which feels fantastic internally but is hard to monetize externally. The upward turn becomes visible when time saved turns into revenue earned, when product cycles shorten, when service quality becomes a competitive advantage, and when firms can scale without scaling headcount at the same pace.

This marker appears in the data as a broadening beyond the infrastructure layer, with adoption benefits spreading from the obvious winners to the unglamorous sectors that make up most of the economy.

The third marker is the buildout of the physical layer. As adoption moves from firms to whole industries, constraints become more binding. You see rising investment in data centers, networking, and specialized hardware. You see bottlenecks in grid connections, power procurement, and construction capacity. You see firms signing long-term energy contracts and redesigning their infrastructure plans around reliability and cost. When the constraint shifts from “Can it do the task?” to “Can we power it, secure it, and scale it?”, you’re watching the economy inch toward takeoff.

The fourth marker is institutional and regulatory clarity. Adoption accelerates when the rules become clearer. Liability frameworks, data governance standards, procurement rules, and industry-specific guidance reduce uncertainty and lower the cost of integration.

Tighter regulations aren’t necessarily a bad thing ; they can slow down adoption in the short run while enabling broader diffusion in the long run by increasing trust. Investors should therefore pay attention not only to whether regulation is restrictive, but also to whether it is legible.

Conclusion

Productivity isn’t waiting for one breakthrough, one killer app, or one magic quarter in the data. It depends on organization, measurement, infrastructure, energy, skills, and institutions moving in the same direction, often unevenly. The path to a full takeoff stretches far into the future. But along the way the signals will most likely become evident, first in how firms are rebuilt, then in how capital is deployed, and finally in how output is produced and priced.

The primary risk for investors is not that AI will fail. The primary risk lies in misreading where we are on the runway, pricing in the destination while ignoring the co -invention processes in the middle. That’s how markets get ahead of the economy.

There’s nothing automatic about technology bringing prosperity. The size and timing of productivity gains will also depend greatly on regulation, diffusion, competition, and sharing of benefits. AI can help us learn to fly, but we need to be patient before expecting sustained altitude.