The threshold, not the frontier
Most arguments about AI and employment are arguments about capability: can it really do the job, is it as good as a person, look at the mistakes it makes. That framing gets the economics wrong, because a company deciding whether to replace a role is not asking whether the machine is better. It is asking whether the machine is good enough at a low enough price.
That threshold is much lower than the frontier. Plenty of work is already below it. And the two things that most obviously separated a capable model from a working employee — durable memory of what happened last week, and the ability to operate the same screens a person operates — are engineering problems being solved incrementally, not scientific mysteries awaiting a breakthrough.
Remote work already did the hard part
The thing that made large parts of the economy automatable was not a model release. It was the decision, made at scale from 2020 onwards, to reduce enormous quantities of work to a screen, a keyboard and a video call.
If a job can be done from a laptop in another city, then everything that job consists of has already been converted into text, images and clicks — the exact interface a model operates through. The office did the format conversion. Whether that job is done by a person at home or by software is now a question about cost, not about feasibility.
This is why the exposure is so much broader than the industries that talk about AI. Insurance underwriting, marketing, paralegal work, bookkeeping, claims processing, customer support, internal reporting — none of these are technology companies. All of them are already fully digital.
It reaches past the desk, too. A quality-assurance function that consists of a trained person looking at objects and judging them is, in a growing number of settings, a few cameras and a cheap model.
Why the human touch saves fewer jobs than people think
The usual reassurance is that people will pay for a human. Some will, for some things, on principle — and those jobs will survive.
But the argument proves less than it appears to. It only holds where the customer both notices and cares who did the work. Most work in a modern economy is not like that: the customer never meets the person who processed the claim, reconciled the account, drafted the contract or wrote the report. The preference for a human cannot protect a role the customer cannot see.
And where the preference does exist, it is priced. A human therapist, a human teacher, a human doctor may be genuinely better and genuinely preferred — and still be unaffordable for most people relative to something that costs almost nothing and is available immediately. The human option becomes a premium good. Premium goods employ few people.
The : why this happens in a step
Here is the property that makes AI different from every previous labour-saving technology, and it is the single most important idea on this page.
Humans do not copy. Software does.
When a company hires a brilliant employee, it gets one of them. Their knowledge is trapped in one head, in one place, working one shift. Scaling that capability means finding, hiring and training more people — which is slow, expensive, and imperfect, because the second hire is never quite as good as the first.
An AI that reaches the threshold has none of those limits. The same system can be assigned to every employee in the company at once, as an assistant. It sits alongside the accounts team, the support desk, the analysts and the legal department simultaneously. It watches how each job is actually done — not how the process document says it is done — and it does not forget.
Then one day the company stops using it as an assistant.
Nothing about that transition is gradual. There is no phase where the firm needs half as many analysts, and then a third as many. There is a period where the tool is helping, and then a decision, and then the department is a piece of software running a million copies of itself, none of which needs onboarding.
This is why forecasts built on historical adoption curves mislead. Past technologies diffused slowly because each installation required capital, physical delivery and people trained to operate it. This one is a configuration change.
Replacement is mandatory, not optional
Suppose a chief executive believes automating their workforce would be bad for their staff, their community and their country, and does not want to do it.
They will do it anyway, or they will be replaced by someone who will.
The mechanism is competition and it is not subtle. The first firm in an industry to cut its cost base by a large multiple can undercut everyone on price, or keep prices and take the margin, and either way it wins. Its competitors then face a choice between matching the and losing the market. Shareholders, seeing one company's margin expand, ask the others why theirs has not. Boards act on that question.
This is already visible. In February 2026, Block — the payments company formerly known as Square — announced it was cutting over 4,000 jobs, taking the company from over 10,000 employees to just under 6,000, a reduction of more than 40 per cent. The company's leadership pointed to productivity gains from embedding AI across the business as part of the rationale, though it also said there was no top-down percentage target and the cuts were not a straightforward swap of people for software.
That caveat matters and we will not overstate the case: this was not a clean experiment in machine substitution. What it demonstrates is the thing that actually drives the dynamic — that a large company can now remove 40 per cent of its people and expect to keep functioning, and say so publicly.
Which points at something uncomfortable that predates AI entirely. When Elon Musk cut Twitter's headcount from roughly 7,500 to around 1,500 — close to 80 per cent — the widespread expectation was that the service would fail. It did not. Whatever one thinks of what X became, the site stayed up. A great deal of work in large organisations turns out to be defensible only until somebody stops doing it.
AI does not have to eliminate genuinely necessary work to cause mass unemployment. It only has to give every management team a credible reason to find out how much of their payroll was necessary.
The new jobs are the part that does not arrive
The strongest objection to all of this is historical, and it has a very good record. Every previous technological revolution destroyed occupations and created more than it destroyed. Weavers, farriers, switchboard operators, typesetters — all gone, and employment is higher than ever. Why not this time?
Because of what those technologies actually were.
The automobile destroyed the horse economy and created millions of jobs — but look at what those jobs were: assembling cars, paving roads, pumping fuel, selling insurance, repairing engines, designing the suburbs that cars made possible. Every one of them was a job for a human, because the car was a machine that needed humans to build it, operate it, maintain it and sell it. The technology was a tool. Tools need hands.
The innovation engine does not stop with AI. New products, new industries and new categories of work will keep appearing — that part of the historical pattern holds.
What breaks is the assumption that a human gets to do them. A new industry that emerges into a world where AI can already do any screen-and-keyboard task will be staffed accordingly, from the beginning. There is no lag in which humans are needed while the technology catches up, because the technology is already there. The jobs get created. They just do not get created for us.
That is the difference, and it is not a matter of degree. Past technologies were complements to human labour — they made a person's hour worth more. This one is a substitute for it.
What follows
If the jobs go, incomes go, and incomes are what other people's jobs are made of. That is the subject of the next page: the spiral, and why UBI cannot stop it.