How it works
Two properties, and the second one is the whole argument.
The first is the threshold. A firm deciding whether to replace a role is not asking whether the machine is better than a person. It is asking whether the machine is good enough at a low enough price. That bar sits well below the research frontier, which is why demonstrations of models getting things wrong tell you less about employment than people assume. We set this out in Why the jobs go: remote work already did the hard part by converting enormous quantities of labour into text, images and clicks, which is precisely the interface a model operates through.
The second property is the one with no historical precedent. Humans do not copy. Software does.
Hire a superb analyst and you get one of them, in one place, on one shift, with their knowledge locked in one head. Doubling that capability means recruiting, onboarding and training, which is slow, expensive and lossy — the second hire is never quite the first. A system that clears the threshold has none of those constraints. The marginal cost of the ten-thousandth instance is a compute bill, and the ten-thousandth instance is exactly as good as the first, knows everything the first one knew, and starts on Tuesday.
The transition tends to run in two phases. First the system is deployed as an assistant, sitting alongside the accounts team, the support desk and the legal department simultaneously, watching how each job is actually done rather than how the process document says it is done. Then someone decides to stop using it as an assistant.
As an illustration with invented round numbers: a firm with 300 claims processors does not pass through a stage where it needs 150, and then 90. It has a period where a tool is helping, then a board decision, then a workflow running as many copies of itself as the queue requires. Past technologies diffused slowly because every installation needed capital, physical delivery and trained operators. This one is a configuration change.
That is what "scalable human equivalent" names: not a smarter model, but the moment a unit of labour stops being something you recruit and becomes something you provision.
Why it matters to this crash
Everything we argue about employment rests on this one property. A slope is survivable. Institutions built to absorb the pace of ordinary business cycles can absorb a slope. What they cannot absorb is a whole occupational category being provisioned out of existence between two quarterly results.
Replacement is also not optional for the individual firm. The first company in a sector to cut its cost base by a large multiple can undercut on price or keep the margin, and either way it wins. Everyone else then explains to their shareholders why their margin has not moved.
Our own coverage points at Block, which in February 2026 announced cuts of over 4,000 jobs, taking the company from over 10,000 employees to just under 6,000 — more than 40 per cent — citing productivity gains from embedding AI across the business. Block also said there was no top-down percentage target and that this was not a straight swap of people for software. We will not overstate it: it was not a clean experiment. What it shows is that a large firm can now remove 40 per cent of its people, expect to keep functioning, and say so out loud. The precedent predates AI entirely. Twitter went from roughly 7,500 staff to around 1,500, close to 80 per cent, and did not stop working.
What would make this dangerous
The thing to watch is the handover from phase one to phase two, because that is where the step happens and it is visible if you know what you are looking at.
A firm saying that a named function is now run by copies of one system rather than by a team, rather than saying that a team has been made more productive, is the moment the assistant stops being an assistant. So is a firm that cuts a department and does not backfill any part of it, ever, at any level. Growth that arrives with no corresponding hiring is the same signal in a happier costume.
The second thing is simultaneity. One firm doing what Block did is a management decision. Several firms doing it to the same function inside the same short window is the difference between a slope and a step, and it is the version that gives nobody time to retrain into anything.
And the inverse risk, which is just as real and which we hold open honestly: if the assistant phase simply continues, year after year, with the systems helping and the headcount holding, then the step never arrives and we are wrong about the shape of this. Nobody outside the firms actually doing the deploying knows which of those we are in right now, and the firms have every incentive not to say.