The best historical analogy for AI may be electrification: the key disruptions came not when the technology first appeared, but when firms reorganised production around it (David 1990). Electric power had made new factory layouts possible for years before Henry Ford redesigned Highland Park around the moving assembly line in 1913–14. When that reorganisation came, productivity surged, but so did labour market strain: annual worker turnover reached 370% (Raff and Summers 1986). The lesson is that adoption can lag invention for years, and that the largest labour shock may come not from a new technological leap, but from the rapid diffusion of capabilities that already exist.

That historical episode matters again because AI is moving from augmentation to reorganisation. For most of the past decade, adoption was incremental: firms layered AI tools onto existing workflows, speeding up particular tasks without redesigning production around them. That is changing. In early 2026, Project Prometheus, a venture led by Jeff Bezos, announced plans to raise tens of billions to acquire and restructure industrial firms around AI-enabled production systems, while Jack Dorsey has been re-engineering Block around AI-first workflows.
The common thread is organisational diffusion at scale: the shift from using AI at the margin to redesigning production around systems already in hand.

This is where the argument departs from, but also complements, the Acemoglu-Restrepo (2018) framework. Their task-based approach is fundamentally about the automation frontier: which tasks are automated, which new tasks are created, and how that reallocation shapes wages and employment in the long run.

Alternatively, we can take the frontier as given and ask a transitional question: how fast is the economy pushed toward that frontier, and what happens when retraining capacity cannot keep up?

Why speed can do permanent damage

I developed this intuition formally in a new paper (Levy Yeyati, 2026) that introduces a simple dynamic model of AI adoption and labour reallocation. Workers displaced by automation enter a retraining pipeline with finite capacity, while firms adopt AI technologies at a speed denoted by κ. A higher speed of adoption means that the economy moves more quickly toward the same long-run automation frontier. The key question is what happens when the inflow of displaced workers generated by faster adoption exceeds the retraining system’s ability to absorb them. The model shows that adoption speed alone can determine whether an AI transition leaves the labour market largely intact or produces permanent labour-force exit. This mechanism is illustrated in Figure 1.

The distinction matters because transition costs are not a simple scaled-down version of steady-state effects. When AI displaces routine workers, they enter a retraining pipeline with limited throughput. Faster adoption compresses the displacement window: the same total mass of workers arrives in much less time, overwhelming the system. Peak displacement inflow scales with adoption speed even when total displacement is unchanged (see Figure 1).

The mechanism operates through the rising opportunity cost of waiting in the retraining queue. As congestion in the retraining pipeline increases, staying attached to the labour market becomes more costly: workers face longer waits, lower expected wages, and greater uncertainty about re-employment. When waiting time exceeds the expected present value of future wages in the non-routine sector, some workers make a rational calculation to leave the labour force. That exit is absorbing: there is no coming back. The social cost of rapid adoption is therefore not captured only by where employment ends up, but by how steeply participation falls, how long it remains depressed, and how much labour income is compressed along the way.

Two economies can converge to the same long-run automation frontier and still experience very different transitions. In one, firms adopt AI at a pace the labour market can absorb. In the other, adoption outruns reallocation capacity. The long-run destination may be identical; the permanent social damage is not.

As Figure 1a shows, faster adoption produces a sharper and more persistent decline in labour-force participation.

Figure 1 Transition dynamics under slow (solid blue) versus fast (dashed red) AI adoption

Note: The parameter κ denotes the speed of adoption: higher κ means the same long-run automation occurs more quickly. Panel (a): labour force participation — permanently lower under fast adoption. Panel (b): non-routine employment — paths cross as early gains reverse later. Panel (c): non-routine wages — mirror image of panel (b). Panel (d): displacement flow — the same total displacement compressed into a much shorter interval under fast adoption.

Congestion, misleading signals, and market overshooting

The first result is intuitive: faster adoption concentrates the same total displacement into a shorter interval, raising peak unemployment and queueing pressure without changing the long-run automation frontier (Figure 1d). The same number of workers eventually need to retrain; they simply all arrive at the door at once.

That congestion is where permanent damage begins. Workers who encounter long queues and worsening prospects do not just wait longer; some give up altogether. The discouraged stock rises directly with adoption speed, and the labour share stays lower throughout the transition. Figure 2 illustrates how permanent labour-force exit increases with adoption speed in the calibrated model.

This is consistent with evidence from earlier adjustment shocks. Autor et al. (2013) show persistent employment and wage losses in US local labour markets exposed to Chinese import competition, while Jaimovich and Siu (2012) document routine-job losses that do not return after downturns.

Figure 2 Permanent exit as a function of adoption speed

Note: Under the baseline calibration, permanent exit, D(∞), ranges from 3% at κ = 0.05 to 15% at κ = 1.20 — a fivefold increase. The dashed line marks the socially optimal κ* = 0.36; the dotted line marks an illustrative market speed κ_mkt = 0.90. The gap between them is excess permanent exit from market over-adoption.

The less obvious result is that early indicators can be misleading. Under fast AI adoption, non-routine employment rises sooner and productivity gains arrive earlier. A policymaker focused on short-term growth could read that as successful adaptation, even as the stock of discouraged workers quietly builds. Then the crossing arrives: wages fall more sharply during the transition as retrained workers enter the non-routine sector faster than demand expands, while permanent labour-force exit continues to rise. In the long run, wages for the workers who remain may end up higher precisely because fewer displaced workers successfully retrain and return.

These dynamics – congestion, permanent exit, and misleading early signals – generate the central normative result: social welfare is concave in adoption speed and is maximised below the market rate. Figure 3 shows this relationship between welfare and adoption speed. Individual firms internalise the productivity gains from adopting AI, but not the congestion they impose on other workers’ retraining queues, the irreversible exits that congestion triggers, or the wage losses borne elsewhere in the economy. Adoption speed is therefore not just a technological choice. It is a coordination problem.

Figure 3 Social welfare is concave in adoption speed

Note: Welfare, W(κ), is maximized at κ* = 0.36 because output gains saturate as the logistic diffusion completes, while transition costs scale with peak displacement. The shaded region is the welfare loss from adopting at the illustrative market speed κ_mkt = 0.90. The planner’s problem is not whether to automate but whether institutions can absorb automation at the chosen speed.

The race that matters

Goldin and Katz (2008) described inequality as a race between education and technology. The relevant race in the AI transition is slightly different: adoption speed versus retraining capacity, with permanent labour-force exit as the finish line. The policy intuition follows directly: this is an argument for aligning the pace of diffusion with the economy’s absorptive capacity.

Two levers matter. The first is retraining capacity itself: active labour market policies, credential reform, mobility support, and institutional mechanisms that increase the efficiency of worker reallocation. The second is timing. Capacity built before the displacement peak is far more valuable than capacity built after it, because discouragement rises when congestion is highest. More broadly, economies with access to the same frontier technologies can experience very different social costs depending on how much retraining their institutions can sustain.

The model also nests the Acemoglu-Restrepo benchmark as a limiting case. Where task creation is rapid and new non-routine roles emerge quickly enough to absorb displaced workers, wage pressure is milder and the optimal adoption speed is higher. In that sense, the automation frontier and the transition problem are complements, not substitutes: the more dynamic the frontier of new tasks, the less costly rapid diffusion becomes.

This also points toward an empirical agenda. Comin and Hobijn (2010) document large cross-country variation in the diffusion speed of general-purpose technologies. If AI follows a similar pattern, transition costs will depend not only on the technology itself but on the institutions that shape the uptake. The central empirical challenge is therefore to explain differences in labour-market damage across economies facing similar frontier technologies but very different adoption speeds and retraining capacity. Even in a world where frontier progress slows and large language models plateau, the labour shock may intensify as firms finally reorganize around systems already in hand.

References

Acemoglu, D and P Restrepo (2018), “The Race between Man and Machine”, American Economic Review 108(6).

Acemoglu, D and P Restrepo (2022), “Tasks, Automation, and the Rise in US Wage Inequality”, Econometrica 90(5).

Autor, D, D Dorn and G Hanson (2013), “The China Syndrome”, American Economic Review 103(6).

Brynjolfsson, E, D Rock and C Syverson (2017), “Artificial Intelligence and the Modern Productivity Paradox”, NBER Working Paper 24001.

Comin, D and B Hobijn (2010), “An Exploration of Technology Diffusion”, American Economic Review 100(5).

Goldin, C and L Katz (2008), The Race between Education and Technology, Harvard University Press.

Jaimovich, N and H Siu (2012), “The Trend is the Cycle: Job Polarization and Jobless Recoveries”, NBER Working Paper 18334.

Levy Yeyati, E (2026), “Too Fast to Adjust: Adoption Speed and the Permanent Cost of AI Transitions”, working paper, UTDT School of Government.

Raff, D and L Summers (1986), “Did Henry Ford Pay Efficiency Wages?”, NBER Working Paper 2101.