Elevator Boys of GenAI
In the 1920s, every office building had an elevator boy. Although automated elevators already existed, people found the idea of riding in a driverless box dangling by steel cables terrifying, and delegated the role to a uniformed human they trusted, accepting the necessity to pay for the privilege. Over the years, people got used to elevator buttons, but the force of habit and the preference for human touch kept the profession flourishing for years.
The operators were so confident in their indispensability that they went one step too far: in 1945, a New York strike brought the city to a grinding halt, costing it hundreds of millions of dollars. That was the final straw, leading to a massive push to upgrade to automated systems. Within a short while, the job ceased to exist, entering the history books as the only major job category to be completely wiped out from the U.S. Census purely due to automation. The only one so far, that is.
Elevator boys were the cleanest definition of a middleman: someone who exists not because they create value, but because of information asymmetry or transaction friction. The history of modern commerce is largely a history of those “toll booth” trades, and of the technologies that remove them one by one.
Newspapers were once the only viable printed information channel, using their middleman position to bundle content with ads and classifieds, fattening their revenues. People accepted it because nothing else existed – but then the Internet broke their business model. Craigslist alone did more damage to newspaper economics than any editorial failure ever could; Twitter and Facebook finished the job. Today the newspaper is a diminished thing, sustained largely by institutional inertia and nostalgia.
Real estate agents had an informational moat – access to listings, knowledge of comparable sales, relationships with buyers – which was valuable in a world without Zillow. Once that information became freely available, their commission became very hard to justify. The agent survived by clinging to the execution layer, but that too is shrinking.
Here is where the story gets interesting. The companies that dismantled the old middlemen wasted no time building new ones. Uber eliminated the taxi dispatcher and the phone-in booking system, then inserted itself between driver and passenger for a fat slice of every fare. DoorDash did the same between restaurant and customer. Expedia aggregated what travel agents used to know and charged airlines and hotels for access to their own customers. These were genuine technological improvements, but the business model was identical to what they replaced: find a friction point, own it, and extract rent from both sides. The market rewarded them handsomely for this, for a while. Then the next wave arrived.
Generative AI is driving a change of extraordinary scale and speed. We cannot assess the impact of the tsunami from inside it, but we can see the fish floating belly-up, and extrapolate. The agentic economy is eliminating many roles that just a couple of years ago seemed staple of our service-based economy. AI agents will (if they haven’t yet) replace secretaries, clerks of all kinds, brokers, advisors, recruiters, customer service representatives, paralegals – and the buck won’t stop there.
What is happening now to companies like Capgemini, Accenture, and McKinsey is structurally identical to what happened to newspaper classified departments and taxi dispatchers. AI agents do not merely reduce friction – they eliminate the information asymmetry that made the intermediary necessary in the first place. A system embedded inside an enterprise does not need a consultant to explain what it is doing. It does it, iterates, and reports back.
OpenAI and Anthropic understood this early, which is why both recently announced joint ventures – in a parade lockstep – to deploy engineers directly inside corporate clients. OpenAI has built an elite, highly technical consulting wing – a multi-billion dollar venture backed by TPG, Brookfield, Bain Capital and others. Anthropic teamed up with alternative asset titans like Blackstone, Hellman & Friedman, and Goldman Sachs to form a dedicated AI services company. AI labs are moving fast into services and deployment because model commoditization is a risk, and because adoption bottlenecks hurt revenue growth.
The Big Four are seemingly fine for now, touting alliances with the AI leaders, helping them scale AI implementation across their enterprise clients. However, professional consultants are clearly the next elevator boys, hanging by the thread of the “human in the loop” habit. The only chunk of the consulting business that is accelerating involves embedding the AI revolution into enterprises – and very soon, Anthropic and OpenAI will not require the help of PwC or Deloitte for that. They are the owners of the technology: why would they pay a toll for a booth on their own road?
The irony is pointed: the companies building the technology that makes middlemen obsolete are inserting themselves as the new middlemen between the AI model and the enterprise. But even this layer is temporary. Once AI agents can deploy themselves, even that layer compresses. The OpenAI and Anthropic JV story is the last gasp of the middleman era.
The real and durable beneficiaries of the AI economy are not the model builders. Raw intelligence, reasoning, and pattern-matching are no longer rare, expensive breakthroughs – they are becoming cheap, standardized, and universally accessible. Core AI technology is turning into a commodity, just like electricity once did – and the value moves both up and down the stack.
“Up the stack” is a constantly moving target. Right now, it sits in hyper-specific vertical applications – defense, aerospace, finance, and medicine – where proprietary data, regulatory compliance, and domain expertise create durable moats. It also lives in the integration layer: the software plumbing that turns raw AI reasoning into auditable, legally compliant enterprise actions.
In the near term, value will shift further to agentic orchestration – the “Agent Overlords” that coordinate swarms of specialized AI agents, manage workflows, handle exceptions, and maintain oversight across complex business processes. These control planes will become the new scarce and valuable layer, much as operating systems and databases once did. What comes after that is harder to predict, but the pattern is clear: as each layer commoditizes, the economic prize moves to the next bottleneck.
“Down the stack” is the physical layer underpinning everything, and that’s where the true moat is. Every agentic transaction, every automated workflow, every AI-mediated business relationship runs on cloud compute, which runs on power, which runs on tangible assets unlikely to be replaceable for at least the next decade. After a century of disruption, humanity has come full circle: the “boring” material world – acres, bricks, pipes, wires, water, and power – has once again become the real source of scarcity and enduring value.
OpenAI and Anthropic are the last of the middlemen: brilliant, richly capitalized, yet ultimately dependent on infrastructure they do not own. What sits beneath them is not a new intermediary – it is bedrock.
