This started with a video I came across on X.
Matt Kim was making an argument about AI and jobs that I hadn’t quite heard framed that way before. His point, roughly, was that we may be looking for job losses in the wrong place.
The usual story is that an existing company adopts AI, automates some of the work its employees currently do, and eventually needs fewer employees.
Kim’s argument was that the bigger disruption could come from somewhere else.
Established companies have years of technology, processes and organisational structure built around the old way of doing things. A new company doesn’t. It can be designed around AI from day one, operate very differently, and potentially take business from the incumbent.
The jobs then disappear because the old company loses, rather than because somebody inside it installed AI.
I wasn’t convinced by the stronger version of the argument, that incumbents are effectively too burdened by legacy systems to adapt. Large companies have advantages of their own, and the evidence actually shows they are adopting AI faster than smaller firms.
But the mechanism stuck with me.
Partly because, a few weeks earlier, I had installed a 4-billion-parameter language model on my iPhone. Nothing exotic. Just my regular iPhone 15. It sits on the device and can answer me without an internet connection.
Obviously, this is nowhere near the frontier models running in enormous data centres. But companies don’t need frontier intelligence for every email, document review, customer query, reconciliation, coding task or piece of internal analysis.
Put the two observations together and a different question emerges.
What happens when useful machine intelligence becomes infrastructure?
If increasingly capable AI is cheap enough to run on an ordinary phone, simply having access to AI eventually becomes less interesting. The scarce capability moves elsewhere.
And perhaps that is where Kim’s argument gets particularly interesting.
We keep asking which jobs AI will take. A more useful question may be:
What kind of company becomes possible when intelligence gets much cheaper?
Because AI doesn’t need to take your job directly. It can change the economics of the company that employs you.
We are looking for disruption inside the company
The standard AI jobs story is easy enough to understand.
A company employs 100 people. AI automates work previously performed by 20. Eventually the company needs 80.
That is happening. Just not yet at anything like the scale the public debate sometimes implies.
The US Census Bureau recently surveyed how American businesses are actually using AI. During the survey period, 18% of firms reported using it in at least one business function. Because larger companies were more likely to use AI, those firms accounted for roughly 32% of employment.
Most adoption was still fairly shallow. Among companies using AI, 57% had deployed it in three or fewer business functions. Two-thirds said they were using AI solely to assist people with tasks. Only 2% reported employment reductions associated with AI.
So far, this doesn’t look like mass replacement.
But something more interesting appears when you look at where the employment effects are showing up.
Stanford researchers using payroll data covering millions of US workers found no widespread economy-wide displacement. Among 22-to-25-year-olds in highly AI-exposed occupations, however, employment was about 19% below where it would have been if it had kept pace with less-exposed occupations.
And much of the adjustment appears to be happening through reduced hiring rather than increased firing.
Take a company with 100 employees that would previously have needed another 15 people to support its growth. With AI, perhaps it hires ten.
Nobody receives a redundancy letter saying an algorithm took their job. There are simply five jobs that never get created.
But even that is still happening inside an existing company.
There is another route.
What if the company never hires the five people?
Imagine two businesses trying to solve the same problem.
One already exists. It has employees, departments, software, reporting lines, budgets and contracts. Its natural question is:
How can AI make this organisation more productive?
The second company hasn’t been built yet. Its founder gets to ask:
If AI can do this work, why would I build that part of the organisation at all?
Those questions lead to very different companies.
We are starting to see evidence of that.
Hyunjin Kim of INSEAD and Rembrand Koning of Harvard studied Y Combinator companies and a broader sample of US venture-backed startups. In their Y Combinator sample, AI-native firms had approximately 25% fewer employees than comparable non-AI startups. The broader venture-backed sample showed a smaller, but similar, difference of around 12%.
Their workforces were also structured differently: a greater proportion of engineers, fewer entry-level employees and managers, and flatter hierarchies. Yet valuations were broadly comparable.
The qualification matters. These are young venture-backed companies. Comparable valuations tell us very little about eventual profitability, durability or market power.
Still, the organisational difference is hard to ignore. The researchers also found that the effect was strongest where AI was built into what the company actually produced, rather than simply added as a tool for existing workers.
There is some early evidence on new-company formation too. Research published this year found about 20% more startup formation in industries with greater exposure to generative AI following its diffusion.
Again, early evidence. But the direction is worth watching.
AI may be changing more than how productive an individual worker can be. It may be changing how much organisation you need in the first place.
That gives us a different route through which jobs can disappear.
An incumbent automates a role. An entrant never creates it.
And if the entrant eventually takes business from the incumbent, the employment effect has happened through competition between companies rather than automation within one.
We do not yet have enough evidence to say that last step is happening at meaningful scale.
But the mechanism is now plausible enough to watch.
Existing companies have to cross a distance
Buying AI is increasingly easy.
Rebuilding a company around it isn’t.
A business that has existed for twenty years has accumulated far more than technology. It has workflows, reporting structures, employment arrangements, outsourcing contracts, compliance procedures, compensation systems, software integrations and internal power centres.
Changing these things costs money. It also creates losers.
If AI removes the need for part of a manager’s team, it may remove part of that manager’s authority. If three departments can become one, two department heads have a fairly obvious reason to prefer a slower transition.
And if an existing process is embedded in customer contracts, regulatory approvals or hundreds of software integrations, changing it can be difficult even when everyone agrees the new process is better.
None of this started with AI.
Research on earlier general-purpose technologies has repeatedly found that companies need to invest in new processes, skills and ways of organising work before the full productivity gains show up.
AI makes that old problem particularly interesting because the thing that needs reorganising may be the company itself.
Every established business therefore carries what I think of as a reset cost: the cost of getting from the organisation it inherited to the organisation it would build today.
Technology is only part of it. There are workflows to change, contracts to renegotiate, people to retrain or replace, regulatory requirements to satisfy and internal constituencies to manage.
And you have to do all of this while continuing to serve customers.
A new entrant avoids much of that. Of course, it has a different problem: it has to build a business from zero.
Which leaves us with a surprisingly useful way to think about disruption:
Does it cost more to transform the old company than to construct the new one?
This matters because it changes what boards and investors should be measuring.
A company can spend heavily on AI and still preserve almost all of the organisational assumptions that existed before it. Another can spend much less and redesign an entire workflow around the technology.
The first may report more AI adoption. The second may have changed its economics more.
So when assessing an incumbent, the useful question isn’t simply how much AI it uses. It is how much of the organisation would still exist if the company were being built today.
Some legacy becomes more valuable. Some becomes baggage.
An incumbent’s history doesn’t suddenly become worthless because AI arrives.
Customers and distribution still matter. Proprietary data may matter even more. A banking licence doesn’t disappear because somebody has access to a large language model.
Neither do trusted brands, regulatory permissions, capital, network effects or long-standing commercial relationships. A new entrant has to reproduce or find a way around those advantages.
Other forms of legacy behave differently.
A twenty-year customer relationship may be extremely valuable. The twenty-year software stack used to service that customer may not be.
A nationwide distribution network may remain a formidable moat. Seven management layers designed to move information around the company may not.
A regulatory licence remains scarce. A 100-person document-processing operation may not.
This is what I think of as Legacy Inversion.
Something accumulated as an advantage under one technological architecture becomes a migration cost when the architecture changes.
The useful distinction is between durable legacy and architecture-specific legacy.
One gives the incumbent something the challenger has to reproduce. The other gives the incumbent something it has to unwind.
The entrant has freedom to build. The incumbent has assets that are difficult to copy.
Who wins depends on which advantage matters more.
“Does this company use AI?” tells me very little
Imagine three companies.
Company A gives AI assistants to 10,000 employees.
Company B redesigns a workflow so 40 people can produce what previously required 100.
Company C is founded after the technology arrives and never develops the 100-person structure.
All three can say they use AI. Economically, they are doing three completely different things.
This is why I’m increasingly sceptical of AI adoption as a useful measure of corporate transformation.
Large companies actually have some obvious advantages here. They have money, data, technical staff and established digital systems. The Census data show that larger firms are substantially more likely to adopt AI.
The interesting question comes after adoption.
Can they convert access to the technology into an organisation designed around what the technology makes possible?
For a board, CEO or investor, I think three questions tell you much more than whether a company has an “AI strategy”:
Rebuild: If this company were founded today, which parts of the existing organisation would we still build?
Reset: What would it cost, financially and politically, to get from the current company to that version?
Replicate: Which advantages do we possess that a new AI-native competitor would genuinely struggle to reproduce?
The first tells you how much organisational baggage exists. The second tells you how difficult it is to remove. The third tells you whether carrying that baggage is worth it.
Can the incumbent reorganise faster than the entrant can reproduce what makes the incumbent valuable?
That, to me, is the more useful competitive test.
Africa may have less to unwind
There is an African angle here that I think is worth exploring, with a fairly large warning label attached.
African companies are usually assumed to be disadvantaged by having less sophisticated enterprise technology. Often they are.
But picture an established Western company with thousands of employees, decades of enterprise software, hundreds of integrations, outsourcing agreements and several layers of management.
Now put it beside an African growth company running a surprising amount of its operation through cloud software, spreadsheets, WhatsApp and a much smaller workforce.
The African company may objectively have weaker technology.
But suppose the efficient AI-native company looks radically different from both of them.
Then the question changes:
Who has further to travel?
Africa has seen a version of this before.
Across much of the continent, limited fixed-line telephone infrastructure was an obvious development disadvantage. Then mobile arrived. Countries that had never built extensive fixed networks had much less installed infrastructure to replace.
AI raises the possibility of a similar inversion at company level. A firm that never accumulated thirty years of enterprise architecture may have less to dismantle when the optimal architecture changes.
I would be very careful about taking that analogy further.
Africa did not magically become technologically advantaged because it had fewer landlines.
AI still needs electricity, connectivity, compute, usable data, skills and capital. It needs customers who can actually pay for whatever is being produced. Regulation can matter enormously.
An African company can have less legacy to unwind and still lose because its data is poor, electricity unreliable, capital expensive or market too shallow.
The proposition is more specific: low legacy burden may give some African companies an option to build AI-native organisations without first dismantling an older one.
That doesn’t mean the option can always be exercised.
For African founders, this has a practical consequence. Copying the organisational structure of a developed-market incumbent may actually throw away part of the advantage. If AI changes the minimum efficient structure of the business, the opportunity is to build toward that structure directly.
For investors, the interesting companies may therefore be those combining low legacy burden with the things AI cannot simply supply: proprietary data, distribution, regulatory access, trusted customer relationships and sufficient demand.
The strongest candidates should be businesses where information processing is a large part of the cost base, physical-capital requirements are relatively low, customers can be reached digitally and regulatory barriers are manageable.
Parts of professional services, software, customer operations, financial services and media fit that description. The argument becomes much weaker where physical infrastructure remains the main constraint.
So I wouldn’t bet on “Africa leapfrogs with AI”.
I would look, sector by sector, for African companies that have less to unwind and enough of everything else required to take advantage of it.
That is a much more interesting group.
The scarce capability is moving
Which brings me back to the model running on my phone.
Saying intelligence is becoming abundant would be going too far. Frontier models, proprietary data, compute and specialist technical capability remain concentrated.
But useful machine intelligence is becoming much easier to obtain.
As that continues, access becomes a weaker differentiator. Scarcity moves somewhere else.
For one company, proprietary information. For another, distribution. For another, trust. For a bank, perhaps regulatory permission and accountability.
And for many established companies, the scarce capability may turn out to be something much less glamorous:
the ability to rebuild the company around increasingly cheap machine intelligence.
That changes how I think about who is most exposed to AI.
It may not be the company employing the largest number of people doing work that AI can automate. It may be the company carrying the greatest distance between the organisation it has and the organisation the technology now makes possible.
We don’t yet know whether AI-native entrants will become a major mechanism through which AI destroys jobs. The market-share evidence isn’t there yet.
But we can now see enough of the pieces to take the possibility seriously.
AI-native firms are already being built with smaller teams and flatter structures. Startup formation is increasing faster in industries more exposed to generative AI. Among young workers in AI-exposed occupations, some of the first measurable labour-market effects are showing up through reduced hiring rather than mass firing.
And increasingly capable machine intelligence can now run on a device sitting in my pocket.
So perhaps the most useful AI question for an established company is a thought experiment.
Imagine you were handed your company’s customers, brand, proprietary data, licences and capital tomorrow morning.
But none of its employees, software, contracts, workflows or reporting lines.
You can rebuild the business from zero using today’s technology.
How much of the old company would you rebuild?
Whatever your answer leaves behind is where the AI transition gets interesting.
Some companies will discover that decades of accumulated infrastructure remain an enormous advantage. Others may discover that part of what looked like a moat was simply the architecture required when machine intelligence was expensive.
The AI winners may therefore be determined less by who gets the technology first, and more by who can rebuild around it without giving up the advantages that still matter.