If an AI takes over just a few tasks in a job, does that feel different from it taking over most of them?

This explores whether swapping workers for AI gets easier or harder as automation reaches more tasks. The short answer from the corpus: no paper here measures a single substitution elasticity changing over time. What the corpus does show is that the answer depends heavily on *how* automation spreads, meaning which tasks it reaches, in what order, and how those tasks connect to each other. Several notes, taken together, suggest that substitution is not a fixed rate but something that shifts with the shape of exposure.

The most direct evidence is about concentration versus breadth. Firm-level data from 2010 to 2023 shows that when AI exposure is concentrated in a few tasks, workers move into the tasks that weren't touched, and net job losses stay modest. When average exposure across tasks rises, labor demand falls Does concentrated AI exposure enable workers to adapt and reallocate?. Put another way, early automation spread thinly gives people somewhere to go. As it covers more of a job, those escape routes shrink, and substitution likely speeds up. A related point from Acemoglu: AI's wage effects look more evenly spread than past automation because tasks get reshuffled between capital and labor rather than simply eliminated, yet capital income still pulls ahead overall Will AI increase inequality less than past automation?.

An opposing force points the other way. When tasks multiply together rather than add up (the 'O-ring' idea, where one weak link ruins the whole product), automating some steps makes the remaining human steps *more* valuable, not less Does automating some tasks actually increase the value of remaining work?. Autor's bank-teller example makes this concrete: teller employment rose alongside ATMs because the human part of the chain became worth more Does automating some tasks make other human tasks more valuable?. In these setups, substitution can actually *slow* as automation spreads, because the leftover tasks become the bottleneck. One version of that bottleneck is human verification. As AI execution gets cheap much faster than checking its output does, the human capacity to audit results becomes the scarce input What becomes scarce when AI execution grows cheap?. A shipbroking case study shows this happening inside one firm: manual classification work wasn't eliminated but turned into auditing and retraining the algorithm Does automation eliminate manual work or repurpose it?.

Two notes show that the rate isn't uniform across firms or tasks either. Firms with more AI exposure replace online freelancers faster and more cheaply, which suggests that building AI capability in-house gets easier the more you do it, rather than the technology simply diffusing at one even pace Do firms substitute labor for AI at different rates?. At the task level, the fixed cost of fine-tuning means high-volume tasks get automated first, while equally hard low-volume tasks stay with humans. Task scale explains 63% of the variation in whether automation pays off Does AI's automation advantage depend on how often a task runs?. Together, these imply that automation spreads from big, repetitive tasks toward the long tail, and each further step costs more per task automated.

What you might not expect: the spread of substitution may also be steered by choices, not just by technology. Firms underinvest in AI that works with people because automating expertise pays them better Why do firms build automating AI instead of pro-worker AI?. Brynjolfsson warns this can lock in a trap where workers lose the bargaining power to change course Does automating labor concentrate power more than augmenting it?. At the largest scale, the gradual-disempowerment argument says that as human labor becomes unnecessary, society loses one of the informal ways it keeps institutions in line with what people want Does incremental AI replacement erode human influence over society?. So 'how fast does substitution grow' is partly a question about who decides what gets built.