Anthropic's growth model: how fast workers can switch jobs is a key variable in who benefits from AI

The sliders are the product. Anthropic's model lets whoever runs it set how capable AI becomes, how widely businesses deploy it, how much of the work runs without a person in the loop, how much more productive the remaining humans get, and how easily someone displaced lands in another occupation ^[5]. Set them low and the result looks like earlier technology waves, inside historical norms ^[7]. On Anthropic's own account, that last variable is among the most important in the whole model, because work disappearing faster than people can move is what turns displacement into unemployment ^[17].

The middle setting is where the arithmetic gets awkward for anyone defending a pay band. An 8.3% uplift landing at $36.3 trillion implies a no-AI 2030 of about $33.5 trillion, so the model puts roughly $2.8 trillion of extra annual output on the table ^[11][1][2]. Knowledge-worker wages in that same run do not move ^[12]. The extreme run's 32.4% is about 3.9 times the middle run's uplift ^[3], and that is the run where those wages fall more than 10% ^[3]. Output and pay are separate lines here, which is not how they behave in most internal forecasts.

The mechanism is the unit of analysis. Anthropic does not treat a job as one thing AI either can or cannot do; it splits occupations into tasks, some of which AI augments, some it automates, and some that still need a person ^[13]. A model can draft the discharge instructions and organise the care schedule, and it cannot bathe a patient ^[14]. If design and permitting get cheap enough that more infrastructure projects clear the bar, the crews who build them see more demand and possibly better pay ^[15]. Digital work and physical work move in opposite directions inside the same growth figure.

What teams tell themselves is that adoption here is slow, so their people have time. Anthropic's substantial scenario grants that and then takes most of it back. AI is capable of half of all knowledge-work tasks by 2030 and can do most of them autonomously ^[8], while adoption stays patchy enough that people still perform the majority of knowledge work across the economy ^[9]. Growth runs at roughly twice its normal rate ^[10] and the wage line still does not move ^[12]. Slow adoption protected the task count, not the price of doing the tasks.

The audience for this is not the macro desk. It is whoever signs requisitions in the spring, and the useful exercise is two numbers against every role on the chart. First, the share of that role's tasks a current model can already run without a human checking the output. Second, the count of adjacent roles inside your own organisation that person could hold within two quarters, on training you would actually fund. High task coverage with high mobility is a redeployment plan you can write down. High coverage with low mobility is a bill, retraining or severance, and it lands on your budget rather than the economy's. Low coverage with low mobility is where the bedside and the job site sit. Low coverage with high mobility is your bench for everything else.

Anthropic's own illustration of the migration is coders and call-center workers moving into nursing or electrical work ^[16]. That route runs through a licence, a training programme and an employer willing to hire a career changer, none of which appear on a 2030 chart. The model cannot say which scenario arrives, and Anthropic says it is not predicting one ^[4]. It does say which variable to instrument, and mobility is the one measurable without a macro model: how many people changed roles inside your company last year, and how long each move took.