Why
the disruption narrative is more complicated, and more hopeful, than it appears
For years, the AI narrative has been relentlessly linear and tinged with
apocalypse: models get smarter, cheaper, and more capable, and humans get edged
aside, role by role, sector by sector. Then came a headline that disrupted the
script.
“Microsoft is limiting internal use of expensive AI coding tools as
enterprise AI costs surge.”
It sounded like a contradiction from the company that bet its future on
AI, poured $80 billion into data centres, and plastered Copilot across every
product it makes. It was not a contradiction. It was a revelation, though, as
we shall see, a more layered one than the headline suggests.
The Economics Behind the Curtain
Inside Microsoft’s engineering divisions, Claude Code, the AI coding
assistant from Anthropic, was not cancelled because it failed. It was cancelled
because it succeeded too well. Rolled out to roughly 5,000 engineers in the
division behind Windows, Microsoft 365, Outlook, Teams and Surface, it reached
usage rates of 84 to 95 percent within months. Engineers used it relentlessly,
and token-based billing, where every prompt, every agentic loop, every
code-generation cycle costs real money, ran to an estimated $500 to $2,000 per
engineer per month. The internal memo set a cancellation deadline of June 30,
2026.
The pattern is wider than Microsoft. Uber’s CTO has confirmed the company
burned through its entire planned 2026 AI coding budget in four months, after
actively incentivising engineers to maximise usage. Meta built an internal
leaderboard called “Claudeonomics” to track which employees were consuming the
most AI tokens. Amazon encouraged “tokenmaxxing”, gamifying maximum AI
consumption as a proxy for productivity.
Two honest caveats belong in this story, and most commentary has skipped
both.
First, Microsoft’s decision was not purely about cost. Engineers
reportedly preferred Anthropic’s tool to Microsoft’s own Copilot CLI, and the
cancellation conveniently redirects them into Microsoft’s own stack. Cost was
real; so was competitive strategy. Second, the escape route is no escape:
GitHub Copilot itself is moving to usage-based billing from June 2026. The
token meter is not a Claude problem. It is the emerging price structure of
frontier AI itself.
The collective result stands nonetheless: for many enterprise tasks
today, undisciplined AI usage is more expensive than the humans it was meant to
augment. The promise was frictionless efficiency. The reality is that when
thousands of employees use frontier models without governance, the economics
invert.
The Objection This Argument Must Survive
Before drawing conclusions, the cost story has to face its strongest
counter-argument: the price of intelligence is falling, fast. The cost of
frontier-quality inference has been dropping several-fold every year. What
looks like a “cost ceiling” in 2026 could look like a rounding error by 2028.
So is the Microsoft episode just a temporary blip?
Not quite, and the reason is an old one. Economists call it the Jevons
effect: when something useful gets cheaper, we do not spend less on it; we use
vastly more of it. Microsoft’s engineers did not hit a budget wall because
tokens are expensive. They hit it because usage exploded faster than prices
fell, agentic loops, always-on assistants, code generated and regenerated at
industrial scale. Every cost decline to date has been swallowed by appetite.
The lesson is not that AI is permanently expensive. The lesson is that AI
consumption, like cloud computing before it, will always expand to consume the
budget available, and therefore governance of usage, not the price of tokens,
is the durable management problem.
Does This Slow the Replacement of Humans?
Only partially, and only temporarily. The cost ceiling buys time. It does
not change direction.
The “AI will replace humans” narrative was always too blunt. AI is not a
flat substitute for human labour. It has a cost curve. At low usage it is
remarkable. At scale, without discipline, it is ruinous. Companies will not
replace human beings wholesale. They will replace them selectively, deploying
AI where ROI is unambiguous, retaining humans where judgment, accountability,
ambiguity, or trust cannot be priced away.
The most exposed roles are not at the bottom of the skills ladder or the
top. They are in the middle: paralegals, junior coders, financial analysts,
content writers, customer service agents, roles that are routine,
pattern-based, and high-volume. At Uber, around 70 percent of committed code
now originates with AI. That number should concentrate the mind of every
mid-career professional whose work is pattern recognition at volume.
The Radiology Test: Why “Exposed” Is the Wrong Word
But “exposed” is a one-dimensional lens, and one profession shows why.
Consider radiology, the example most often cited, for a decade now, as the
first white-collar casualty of AI.
On capability, the pessimists are right: AI can read many scans as well
as or better than humans, and the per-unit economics are unanswerable. On
accountability, the pessimists are early: in most jurisdictions, a diagnosis
requires a licensed human signature, and regulators, courts and insurers will
keep it that way for some time, not because the human is always more accurate,
but because someone must be answerable when the machine is wrong.
And on access, the pessimists have the sign of the effect backwards, at
least in a country like India. The binding constraint on radiology in
small-town and rural India has never been an oversupply of radiologists. It is
that there are almost none. AI-assisted reading changes that arithmetic. A
local physician in a taluk hospital, supported by AI triage and a remote human
radiologist for sign-off, can now order and act on imaging that was previously
out of reach. The realistic effect in such markets is not fewer radiology jobs
but more radiology, more referrals, more scans, more diagnostic activity, and
new paramedical and technician roles around it, in places where the alternative
was not a human radiologist but nothing at all.
The same technology, in the same year, can displace work in saturated
markets and create it in underserved ones. Radiology is not an exception; it is
the template. Apply the same three questions, can AI do it, who must answer for
it, and where was the service never available at all, to law, to accounting, to
software, to education, and the picture that emerges is not a single wave of
obsolescence but a redistribution: of tasks within professions, and of services
across geographies.
The Real Disruptor: Robotics, Not Software
While enterprises wrestle with token bills, robotics companies are
solving a different equation. A humanoid robot is largely a capital
expenditure: its “salary” is electricity, maintenance and software. Tesla
Optimus, Figure and Boston Dynamics are targeting price points intended to
undercut the minimum wage in developed economies within this decade, beginning
with exactly the jobs that employ hundreds of millions globally, fast food,
warehouse picking, hotel housekeeping, retail stocking.
Two qualifications keep this honest. First, the cost structure is
different from software AI, not free of recurring costs: many robotics firms
are pricing robots-as-a-service, with subscriptions and teleoperation support,
a cousin of the token meter, not its opposite. Second, the crossover point
depends on the wage it must undercut. A robot that beats a $15-an-hour wage in
Ohio is nowhere near beating a ₹15,000-a-month wage in Kanpur. Which leads to a
striking inversion: the Global South will likely adopt software AI fastest,
because models are cheap and skilled labour is scarce, and adopt robotics
slowest, because physical labour is abundant and cheap. The displacement map of
the next decade will not be uniform. It will be a patchwork drawn by local
wages.
The China Factor: The Cost Floor May Collapse
DeepSeek’s breakthrough in early 2025, and the rapid succession of
low-cost Chinese models since, changed the global cost equation in ways that
have not been fully absorbed. Frontier-level reasoning delivered at a fraction
of Western pricing, backed by structural advantages: subsidised compute and
energy, lower infrastructure costs, and less shareholder pressure to monetise
quickly.
If such models gain wide adoption across India, Southeast Asia, Latin
America and Africa, markets where price matters more than geopolitics, the cost
barrier falls years ahead of current projections. And note that this cuts both
ways: cheap models accelerate displacement of routine work, but they equally
accelerate the access story told above. The same collapsing cost floor that
threatens the call-centre agent makes the AI-assisted rural clinic viable. The
West will move more cautiously, constrained by data sovereignty and regulatory
anxiety. The Global South may move faster, precisely because it cannot afford
to be slow.
This creates a two-speed world of AI adoption, and, by extension, a
two-speed world of both displacement and inclusion. That asymmetry deserves far
more attention than it receives in the global policy conversation, which
remains written almost entirely from the vantage point of high-wage economies.
The Jobs That Don’t Exist Yet
Every major technological shift destroys familiar work and creates
categories of work invisible until they become indispensable. The steam engine
did not just displace handloom weavers; it created railway engineers and
factory inspectors. The internet did not merely kill travel agencies; it
created cloud architects and UX designers. The honest difficulty is that new
jobs are not legible until they exist.
Still, the outlines are forming at the margins: people who design how
humans and AI divide work and supervise each other; specialists who generate
and govern synthetic data as real data becomes regulated and scarce;
professionals who arbitrate between AI outputs and human decisions in finance,
healthcare and law; a new blue-collar profession maintaining and supervising
robot fleets; AI safety and governance analysts, a field that will grow to the
scale of cybersecurity; and millions of AI-enabled one-person enterprises that
would have been operationally impossible a decade ago. Above all, as AI absorbs
the burden of logic and pattern, the premium on trust, empathy, cultural
context and meaning rises, precisely because AI cannot credibly supply them.
AI as an Equaliser: The Welfare Dimension
If AI dramatically reduces the cost of delivering essential services, the
welfare gains could, under the right policy conditions, outweigh the disruption
to employment. AI-optimised irrigation and supply-chain prediction could raise
yields and cut food waste across the Global South. AI triage, diagnostic
support and remote monitoring could bring quality care to populations who currently
have none, the radiology story above, repeated across a dozen specialities.
Adaptive AI tutors could personalise learning for hundreds of millions of
children for whom quality schooling remains a geographic accident of birth.
If AI reduces the cost of food, health and education by half or more, the
question is no longer simply “who loses their job?” It becomes “what kind of
society do we build with the surplus?” That is a question of political will,
not technology.
Where Humans Still Win on ROI
Despite the compression of timelines, some domains will remain higher-ROI
for humans through this decade: trust-based, relationship-driven work, where
clients pay a premium for human accountability; novel problem-solving in
genuinely ambiguous environments, where no training data exists for the
situation at hand; skilled trades in unstructured physical settings, the
plumber navigating an unfamiliar home, the electrician improvising under
deadline; and regulated roles where law or professional standards mandate human
sign-off regardless of AI capability.
The pattern across these safe harbours is consistent. What makes humans
irreplaceable is not intelligence alone. It is accountability, physical
adaptability, and the fact that in some relationships the human presence is
the product, not merely the mechanism of its delivery.
The Question That Actually Matters
The debate has moved on from whether AI will displace human workers. The
debate now is: which humans, doing which tasks, in which geographies, on what
timeline, under which cost structures, and, critically, will the new categories
of work emerge quickly enough, and be accessible enough, to absorb those
displaced?
The cost ceiling revealed by Microsoft, Uber and others is real. It slows
the slope, forces selectivity, and creates space for societies to adapt rather
than absorb a vertical shock. But it does not alter the destination. The
direction of travel has not changed, only the gradient of the curve. And the
gradient, as the radiology test shows, points in different directions in
different places: downward for routine work in saturated markets, upward for
services in markets that never had them.
The most important investment any individual, institution, or government
can make right now is not in AI itself. It is in the human capacity to navigate
the transition: to identify which skills will compound in value, which roles
are building toward the new categories, and which paths are quietly narrowing.
“AI will not erase human value. It will redraw the map of where that value lives, and whether we prepare to inhabit that new terrain is the defining challenge of this decade.”
Sources:
The Verge (Tom Warren, Notepad, May 14, 2026) on Microsoft’s internal Claude
Code cancellation; The Information (April 2026) on Uber’s AI coding budget;
Fortune (May 22, 2026) on Meta’s “Claudeonomics” and Amazon’s token incentives.
