Companies don't swap in AI as fast as it improves — they move at whatever pace fits their own setup.

This explores whether firms swap workers for AI at the speed the technology's abilities would predict, or whether something else sets the pace. The short version: the corpus says there is no single adoption speed. How fast a firm substitutes depends less on what AI can do in general and more on how much AI capability that particular firm has already built. One study of firms hiring from online labor marketplaces found that the most AI-exposed firms replace those workers faster and more cheaply than less-exposed firms Do firms substitute labor for AI at different rates?. That pattern looks like returns to scale: once a firm has the internal setup to use AI, each further substitution gets cheaper. So it isn't the same technology spreading evenly across the economy. Two firms with access to the same models can move at very different speeds.

Where substitution does happen, it follows capability fairly closely, though not the capability you might expect. Tracking where workers have actually handed tasks to AI inside structured workflows shows delegation concentrated in information-heavy jobs. It follows what the technology can technically do more than how popular chatbots are Where have workers actually delegated tasks to AI?. That pattern also breaks with older automation forecasts, which expected routine tasks to go first. But 'capability' is slippery, because the usual measuring stick misleads. Agents win benchmark contests yet fail at long, multi-step professional workflows, so benchmark scores overstate how ready they are for real work Why do agent benchmarks not predict real economic value?. Cost and speed over a whole task can matter more than peak ability, and smaller models trained on real execution can be the practical choice for actual work Does model efficiency matter more than peak capability for real work?. If you compare adoption to headline capability, firms look slow. If you compare it to capability you can actually deploy, the gap shrinks.

The less obvious point is that how exposure is spread across a job matters as much as how much there is. Task-level data from 2010–2023 shows that when AI touches only a few tasks in a role, workers shift to the remaining tasks and net job losses stay modest. When exposure is high across many tasks, labor demand falls more Does concentrated AI exposure enable workers to adapt and reallocate?. So the speed of substitution is partly about whether a job is reshaped or removed. Incentives also push the direction of substitution: firms earn more from automating expertise than from building AI that creates new tasks for workers, so investment leans toward replacement Why do firms build automating AI instead of pro-worker AI?.

One gap to flag: the corpus doesn't have a direct measurement of the lag between when AI becomes able to do a task and when firms actually substitute for it. The firm-level, delegation and benchmark studies get at it from different sides, but none measures the time gap itself. For the longer-run stakes of uneven, step-by-step substitution, see Anthropic's scenarios, where labor's share of income falls as AI accelerates Does AI growth inevitably shift wealth away from workers?. Also see the argument that gradual replacement removes a hidden safeguard. Institutions have stayed aligned with human interests partly because they depend on people who care about outcomes Does incremental AI replacement erode human influence over society?.