The Electric Motor Problem.

In 1821, Michael Faraday demonstrated the first electric motor at the Royal Institution in London: a wire rotating around a magnet in a pool of mercury. It was crude, impractical, and one of the most consequential demonstrations in the history of science. Over the following decades, the technology matured. Nikola Tesla introduced the alternating current motor in 1887, and by the 1890s, commercially viable electric motors were ready for industrial use. They made possible capabilities no prior power source could offer, from precision machining to continuous automated production.

It might come as a surprise, then, that there was a thirty-year lag between widespread adoption of the electric motor and the day factories saw any real benefit from it.

For most of the 19th century, factories ran on waterwheels and steam engines. These power sources were large, centralised, and immovable. The entire factory was designed around them. Overhead rotating shafts, called line shafts, ran the length of the building, transferring mechanical energy to individual machines through an elaborate web of leather belts and pulleys. The factory floor was organised around the flow of power, not the flow of materials or labour. Machines had to sit close to the shaft. The building had to be compact. Reorganisation was a nightmare. The belts shed dust, the shafts lost energy to friction at every interface, and the whole system was inflexible and dangerous. Workers routinely lost fingers to exposed belt drives.

When electric motors arrived, factory owners did what seemed obvious: they ripped out the steam engine and dropped in a large, centralised electric motor. The vestigial belt-and-shaft system was left untouched. A newer, superior power source was adopted. Efficiency, output, and the bottom line barely moved. The factories that electrified early saw so little return that it actually discouraged others from adopting the technology at all. As the Stanford economist Paul David documented in his landmark 1990 paper “The Dynamo and the Computer,” the problem was that managers “simply overlaid one technical system upon a preexisting stratum.” They upgraded the power source but left the architecture of the factory completely unchanged. No wonder nobody was rushing to follow suit.

It wasn’t until the 1920s that a new generation of factory designers recognised the centralised belt distribution system as the true bottleneck. The electric motor was small, cheap, and independently controllable. They used those properties to place individual “unit drive” motors in every machine and workstation. Freed from the constraints of mechanical power transmission, factory layouts could finally be redesigned around the flow of materials and people rather than the flow of energy. The results were dramatic: electrification accounted for roughly half of all manufacturing productivity growth during the 1920s.

Progress doesn’t come from better technology alone. It is only realised through the intelligent application of that technology to a system’s true bottlenecks.

For the past few years, every business has been asking, “how can we use AI?” That is the wrong question. When Malcolm McLean revolutionised global trade in the 1950s, he didn’t start with a steel box and ask what to put in it. He watched longshoremen spend days loading cargo onto ships piece by piece and asked, “why is this so slow?” The standardised shipping container wasn’t a technology looking for a problem. It was a bottleneck demanding a solution. It cut loading times from days to hours and reduced shipping costs by over 90%.

We need to be asking, “what are our bottlenecks?” and “which of those can we use the unique strengths of AI to solve?”

The Productivity Paradox

Nobel Laureate Robert Solow formulated the Solow Productivity Paradox in 1987, famously remarking that “You can see the computer age everywhere but in the productivity statistics.” In the 20 years preceding this observation, US productivity growth had plummeted from 3% to 1%, despite computing capacity increasing a hundredfold. Actual ROI on tech investment was 1/4 of expected ROI. Solow and many others opined that the more money invested in IT, the more worker productivity would decrease.

If you read about the Productivity Paradox today, consensus dictates that it was resolved. By the late 90s, productivity doubled, investments paid off, computers delivered everything they promised. The ‘paradox’ was chalked up to measurement error or lag. This standard narrative is wrong.

These long-awaited productivity gains came from just 6 industries: retail, wholesale, securities, semiconductors, computer manufacturing, and telecoms. The other 53 sectors of the US economy increased IT spend and saw almost no productivity growth. Even amongst the 6 growing sectors, gains were unequal.

In retail, Walmart alone drove a huge share by redesigning supply chain logistics with real-time data, forecasting and optimisation. Competitors sought to “use computers in retail”. Walmart redefined retail as a computational problem.

Bezos didn’t start by asking how to sell books online. He asked what a store looks like when it’s designed from scratch around the internet: no physical inventory constraints, infinite shelf space, logistics optimised for shipping rather than foot traffic. Amazon wasn’t a bookshop that went online; it was a logistics and data company that happened to start with books.

Jim Simons didn’t ask his traders to “start using IT”. In fact, he didn’t hire traders at all. He hired mathematicians and physicists, built statistical models to find patterns in market data, and let the models trade. Renaissance’s Medallion Fund returned 66% annually for three decades.

Simons, Bezos, Walton. Three men who surveyed an existing industry, ignored the existing workflow entirely, and rebuilt it from scratch around the new technology. Competitors bought the same tools but overwhelmingly misapplied them. Once it was clear who had won the race, they scrambled to copy, and some succeeded. Every retail chain now runs on Walmart’s playbook, every e-commerce operation borrows from Amazon. But Renaissance remains essentially inimitable, which tells you something about how rare genuine system redesign really is.

Productivity peaked at 3% in the early 2000s, before collapsing back to 1% in the early 2010s. The boom was temporary and concentrated. Computers may have consumed our lives, but that doesn’t mean they elevated us. The uncomfortable possibility is that most of the economic growth in my lifetime, which we attribute to technology and innovation, was actually driven by workforce expansion, cheap capital, and the contributions of a remarkably small number of people.

The IT revolution didn’t occur because computers finally worked. A handful of people figured out how to redesign entire systems around what computers could uniquely do, while everyone else replaced ‘outdated’ and ‘inferior’ methods with computational ones, just to keep having the same meetings.

The conversion rate from buying a powerful technology to capturing value is terrible. It has always been terrible. With electric motors, with computers, with AI. The value accrues to the minuscule number of people with the taste and vision to redefine the world around these new capabilities, not to the millions who buy the tool. So when I hear that 2026 is the year that AI transforms business, I hear 1987 all over again. The technology is ready. The organisations are not.

The AI Disappointment

Against this backdrop, AI numbers should surprise no one. But they do. Of the $40+ billion poured into enterprise AI investment, 95% of pilots show zero P&L impact. Only 5% of custom enterprise tools ever reach production. 42% of companies scrapped their AI initiatives in 2025, up from 17% in 2024. The failure rate of AI projects is double that of conventional IT projects.

There is currently a schizophrenic divide in the market. Scroll through Silicon Valley Twitter, and you will be told 2026 will usher in the end of human labour. Talk to a Fortune 500 COO, and they will tell you AI has done absolutely nothing for their bottom line. Companies are spending money on the wrong use cases. MIT found the biggest ROI is in back-office automation, but over half of AI budgets go to sales and marketing tools.

Again and again, we are applying the technology where it’s visible, not where it’s valuable.

The root cause is exactly what history predicts. When MIT analysed these failures, they found the core barrier was learning, not infrastructure or talent. Most GenAI systems ‘do not retain feedback, adapt to context, or improve over time.’ The most common enterprise complaints are hallucinations and unreliability. These are all symptoms of bad deployment, not bad models.

Every legacy CRM has bolted a chatbot onto the same broken database. Billions have been spent on ‘AI coding assistants’ that just autocomplete text inside the same brittle CI/CD pipelines. Every law firm has subscribed to an AI document review tool that sits on top of the exact same agonising manual workflow. Consulting teams are using AI to generate the exact same mediocre slide decks at ten times the volume.

We swapped steam for electric but kept the belt drives. We bought computers but kept the same workflows. Now, we are trying to “force generative AI into existing processes with minimal adaptation”. Across three cycles, we continue to make the same mistake: focusing on the general, not the purpose.

What Actually Works

If 95% of AI initiatives are failing, it means 5% are succeeding. Why?

Goldman Sachs’ CIO recently revealed that the firm had spent 6 months with embedded engineers from Anthropic. Rather than buy licenses for AI-enabled software to tack onto their existing workflows, the bank opted to design bespoke solutions for two specific, high-volume bottlenecks: trade reconciliation and client onboarding.

These processes “have resisted automation for decades because they require processing large volumes of data against strict regulatory frameworks.”

Which is exactly where AI excels.

Goldman Sachs is now seeing 30% faster client onboarding, 20% developer productivity gains, and “thousands of hours” saved weekly on trade ledger reconciliation. To an enterprise market currently drowning in failed pilots, these numbers look like magic. To an economic historian, they look inevitable.

Just like the new generation of factory designers. Just like Bezos, Simons and Walton. Anthropic and Goldman Sachs found the bottleneck first, and only then did they apply the technology’s unique strengths to resolve it.

This is unit drive. Individual motors on each machine. Not a big, central, highly capable AI model bolted onto the existing system.

The playbook is simple. Find the bottleneck before reaching for a tool. Embed engineers who know both the models and the domain. Build for the specific workflow. Target the back-office, where the value sits, not the front-office, where the demos look good.

The data reflects this reality: MIT found that specialised, deeply integrated vendor partnerships succeed 67% of the time. Internal corporate IT builds, which usually default to buying generic SaaS wrappers, succeed just 22% of the time. You cannot buy a unit-drive factory off the shelf. Success requires deep, targeted integration between the physics of the technology and the reality of the domain.

Interfaces

What’s the hardest part of running a large organisation?

Is it strategy? Talent? Cashflow? No. It’s managing the interfaces between people. Fred Brooks formalised this in 1975: the number of communication channels in a team grows as n(n-1)/2. Ten people, forty-five channels. Fifty people, twelve hundred. Every interface is a point where information degrades: context is lost, intent is misread, nuance is flattened. The entire discipline of management exists, fundamentally, to minimise this loss.

Bezos’s response when his executives pitched better cross-team communication: “Communication is a sign of dysfunction. It means people aren’t working together in a close, organic way.” His solution to this information loss was to minimise the number and throughput of the interfaces by which it occurs. Every effective organisation in human history has been structured to combat this problem. The great Prussian military theorist Carl von Clausewitz saw this problem two centuries ago: “The military machine... is composed entirely of individuals, each of which keeps up its own friction in all directions.” The Joint Commission has found that eighty percent of serious medical errors trace not to misdiagnosis or surgical error, but to communication failures during handoffs. The information existed, but it didn’t survive the interface. Just like the belt-and-shaft system, organisations are predominantly bottlenecked by the degradation of information as it crosses structural and conceptual boundaries, rather than by the generation of those ideas.

A few years ago, a human’s programming ability was roughly their ability to translate an idea from their head into a format a computer can execute effectively. AI programming agents are exceptional at translating ideas to code that works, and they will only get better. Tens of billions of dollars and many of the brightest minds in the world are working to ensure that. But what idea is being translated? By introducing these pair programmers, or any AI assistants that help with any tasks, we have created a new interface between human intent and machine interpretation.

It’s common to hear from people that AI is incredible at completing the first 90% of a task, and infuriatingly poor at finishing the last 10%. If you’re making something of low importance, or working outside your area of expertise, the first 90% might be good enough. But if you’re building complex, high-performance systems, the AI will eventually become more of a hindrance than an asset. The root of this seemingly nonsensical collapse, right at the finish line, is the interface. If there is a persistent gap in a model’s understanding of your ideas, it will barely be noticeable in that first draft, but you might go backwards when working on the final touches. I’m sure this is an annoyingly relatable experience, which makes you feel like you’ve gone from working with an expert to a pre-schooler.

The most effective software engineers today are commanding swarms of dozens of agents. Other industries will move in the same direction. The benchmarks make it clear this is how we produce the most ‘intelligent’ outputs. No matter how smart models become, there is still an interface between human and machine. In fact, the future I see is one where the best models comprise many specialised sub-agents, in which case the information lost to interfaces will increase, and these human-machine interfaces will become the primary bottleneck of human productivity.

Some believe that models will soon become so advanced that they will be capable of generating better ideas than humans, and so this interface problem will solve itself: remove the ‘human’ from ‘human-machine interface’. If you believe this is true, the only rational course of action is to be on your hands and knees at a frontier lab, begging them to take your money, and then enjoy the last few years where human thought and existence still matter. If that’s not the case, then solving these interfaces is a trillion-dollar question that the market is overlooking.

Redesigning the Factory

If the chat box is the leather belt of the AI era, a lossy, centralised interface that degrades human intent, then what does “unit drive” look like?

It requires a fundamental shift in how we deploy non-deterministic models. For the last three years, the tech industry has been obsessed with making models better conversationalists. But humans don’t build skyscrapers, design microchips, or run military operations by simply chatting with each other. We use blueprints, architectural schematics, interface control documents, and rigid building codes. We use structured, high-bandwidth interfaces that force alignment, expose unstated assumptions, and make intent executable.

To eliminate interface friction, we must stop treating AI as an oracle and start treating it as a compiler. Right now, we are trying to write the source code for the global economy using the equivalent of conversational poetry. It is no wonder that the first 90% of a task feels like magic, while the final 10% collapses under the weight of misaligned assumptions and degraded context.

If we are going to build the “unit drive” for the cognitive era, we have to redesign the factory floor around the frictionless flow of human intent. This is the defining challenge of our generation. It’s the one we’re working on. Solving this requires a massive divergence from the current trajectory of Silicon Valley. It demands a two-front war.

First, it is an engineering mandate. We must move away from brittle chat wrappers and prompt-engineering, and instead build deterministic, model-based architectures capable of orchestrating non-deterministic agents without losing fidelity at internal and external interfaces. We have to design the actual schematics for this new compiler.

Second, it is a philosophical threshold. If we successfully build systems that execute complex intent, we are forced to rigorously define what “intent” actually is. We must answer what the fundamental, irreducible role of the human mind becomes when execution is entirely commoditised.

We have the motors. Now, it is time to build the factory.