The economic models of AI displacement tend to focus on labour market effects: which jobs are automated, on what timeline, with what wage effects, and whether retraining programmes can close the gap. These are real questions and the models are increasingly sophisticated. What they don’t typically model is the feedback loop.
The feedback loop: successful AI deployment in a firm reduces its labour costs and increases its output per worker. Aggregate this across an economy and you get rising corporate productivity, rising profits, and — under standard assumptions — rising wages for the workers who remain, rising incomes for shareholders, and rising aggregate demand that funds the purchasing of AI-enabled products. The story ends well if the productivity gains distribute broadly enough.
The loop that breaks this story: if AI deployment reduces labour income faster than it creates new labour demand, aggregate consumer purchasing power falls. The firms that deployed AI to increase their efficiency are selling into a market with reduced demand. The productivity gains that made individual firms more efficient create a collective action problem: each firm rationally deploys AI to reduce costs, but the aggregate effect is to reduce the consumer spending that supports demand for their products.
This is not a new observation. It’s a version of the argument Henry Ford made when he raised wages — not because he was altruistic, but because he wanted his workers to be able to buy the cars they built. The purchasing power of workers is a component of the demand for what workers produce. If AI deployment reduces purchasing power faster than it creates new demand, the efficiency gains are partly self-defeating.
What makes the current moment different from previous waves of automation is the speed of potential displacement and the breadth of the affected categories. Previous automation primarily substituted for routine physical labour, expanding the relative demand for non-routine cognitive labour. The affected categories were specific and the transition period, while disruptive for the people involved, was slow enough that structural adjustments could partially compensate.
Current AI systems substitute for non-routine cognitive labour directly. The categories most immediately affected are the ones with the most consumer purchasing power — professional services, knowledge work, administrative functions. The timeline for disruption in these categories is faster than the historical pattern for physical automation, and the retraining alternatives are less clear, because the AI capability is advancing in the cognitive domains where the displaced workers would be retrained into.
None of this is deterministic. There are plausible paths through which the productivity gains distribute well: higher wages for the workers who remain, lower prices for consumers whose real purchasing power rises, new categories of demand created by AI that weren’t previously possible, policy interventions that accelerate redistribution. The optimistic scenario is real.
The point is not that the pessimistic scenario is inevitable. It’s that the feedback loop isn’t being modelled seriously enough in most discussions of AI economic impact. The models that show net positive outcomes over a multi-decade transition period typically assume that the transition period doesn’t break the demand conditions that fund the transition. That’s a load-bearing assumption and it deserves more scrutiny than it usually gets.
The policy implication is not obvious but points in a direction: the speed of displacement matters as much as the scale. If AI deployment reduces labour income gradually enough that new demand categories and new labour roles emerge in parallel, the feedback loop may be manageable. If the displacement is fast enough that the aggregate demand fall precedes the new demand creation, the positive-sum outcome becomes harder to achieve. Speed is the variable that most directly affects whether the transition period sustains itself.
This is not a reason to slow AI deployment. It’s a reason to take seriously the distributive and demand-side implications of AI deployment speed, and to treat “the transition will work out” as a hypothesis that requires specific supporting conditions rather than an assumption that can be treated as given.
The loop is real. Modelling it honestly is the start of responding to it well.
P.S. The most useful question for evaluating AI economic impact studies: does the model include a mechanism for how aggregate demand is affected by the labour income effects of deployment? Most don’t. The ones that do tend to reach more cautious conclusions about the pace of transition required for positive outcomes.