Complementarity with remaining non-automated tasks substantially limits the output gains from partial AI automation.
Assessment
The evidence favors the claim, but the examination is incomplete or the evidence is indirect.
The claim restates the Baumol or "weak link" logic that runs through the growth literature on AI. When tasks are gross complements, output depends on all of them together. Automating some tasks then makes those tasks cheap and abundant, but output stays tied to the tasks still done by humans, whose supply grows slowly. In a standard CES task model with an elasticity of substitution below one, automating a fixed share of tasks raises output by a bounded multiple, even if the automated tasks become free. Aghion, Jones and Jones (2017) put it this way: growth is limited by what is "essential and yet hard to improve". Epoch's GATE model and Erdil and Besiroglu's review accept the same mechanism for partial automation. The debate in that literature is mostly about full automation, where the bottleneck disappears.
The claim depends on whether the elasticity of substitution between tasks or sectors is below one. Cross-sector estimates generally put it well below one. A second line of evidence is the historical rise in the GDP share of slow-productivity sectors. Task-level estimates are thinner and less direct. The strongest objection is that these elasticities are measured on today's way of organising production. If AI lets firms reorganise production and create tasks that avoid bottlenecks, the constraint would bind less over time. Electrification is the usual historical example, but there the reorganisation took decades.
On balance the evidence favours the claim, especially over horizons of years to a decade. Two questions remain open: how large "substantially" is in practice, and whether reorganisation can loosen complementarity over longer periods. Better task-level estimates of the elasticity would narrow the uncertainty, as would firm-level evidence on whether AI adoption is limited by the steps that are not automated.
Full reasoning: the evidence and decisions behind this verdict
Prior model before searching: the claim is the standard implication of CES aggregation with sigma<1. If a share β of tasks is automated and automated inputs go to infinity, output is bounded at a multiple set by (1-β) raised to a power that depends on sigma. With sigma near 0.3 to 0.5, automating half of tasks gives roughly a 2x to 4x level gain at most, not unbounded growth. That is "substantial limitation" relative to the frictionless benchmark. The open questions are whether sigma<1 holds at the task level, and whether it holds for future AI-reorganised production.
Evidence checked: the Aghion, Jones and Jones NBER paper (w23928) models exactly this case and finds Baumol effects that keep growth balanced. Erdil and Besiroglu (arXiv 2309.11690) treat the bottleneck objection as plausibly decisive for partial automation when sigma is about 1/3, and argue only that full automation overcomes it. Epoch's GATE model (arXiv 2503.04941) builds in task complementarity and says the extent of the non-automated bottleneck is a major determinant of macro impact. No source found argues that partial automation escapes complementarity limits under the current task structure. The dissent is about reorganisation and new tasks, and about whether sigma could exceed one for AI-produced inputs.
Weighing: the Baumol argument gives likely support, conditional on sigma<1. That premise is well supported across sectors and less directly supported across tasks. The reorganisation argument weakens the claim over long horizons but rests on a speculative premise. The status is supported rather than verified because "substantially" is vague and the task-level elasticity is not pinned down. What would change this: robust task-level estimates of sigma≥1, or evidence that AI-driven redesign quickly dissolves bottlenecks.
Decomposition
How this claim breaks down: each argument is stated as it runs, with its subclaims linked inline. ↗︎ opens a subclaim; the map shows how they fit together.
Because tasks are gross complements in aggregate production, making automated tasks arbitrarily cheap leaves output limited by the inputs to the tasks that remain manual, whose supply grows slowly; in a CES task model with elasticity below one, the gain from automating a fixed share of tasks is bounded by roughly the inverse of the non-automated share's weight. The historical rise in the GDP share of slow-productivity sectors is the observed signature of this mechanism.
- Does it follow?
- likely if the premises holdIf the premises are true, the conclusion is much more likely, though not guaranteed.constitution →
- Do the premises hold?
- 2 not separately assessed
Granting that tasks are gross complements, the bound on gains from automating a subset follows almost mechanically within the CES framework. The argument still falls short of certainty, because "substantially" depends on the size of the elasticity and the automated share, and the framework's fixed task structure is itself an assumption. It rests chiefly on The elasticity of substitution between tasks or sectors in aggregate production is less than one., with Historically, sectors with slower productivity growth have gained GDP share relative to faster-productivity sectors. as corroborating historical evidence.
Complementarity estimated on today's task structure is not fixed: if AI lets firms reorganize production and create tasks that avoid bottleneck tasks, the effective elasticity between automated output and remaining human tasks rises over time, so the bottleneck binds less than static models imply and partial automation can yield large gains.
- Does it follow?
- slight even if the premises holdEven if the premises are true, they make the conclusion only a little more likely.constitution →
- Do the premises hold?
- 1 not separately assessed
Granting that AI enables reorganisation and task creation that avoids bottlenecks, the bottleneck would weaken. But a capacity to avoid some bottlenecks does not show that complementarity stops limiting gains substantially, particularly over short and medium horizons. The argument lives or dies on how fast and how completely AI automation can reorganize production and create tasks that let firms avoid bottleneck tasks. holds, and that remains speculative.
Provenance
Where this claim has been said, linked to its canonical form.
economic growth may be constrained not by what we do well but rather by what is essential and yet hard to improve
Modeling automation of a growing set of tasks under a CES with elasticity below one, the authors show Baumol effects keep aggregate growth balanced rather than explosive.
elasticity of substitution across tasks σ ≈1/3, then it is quite plausible that this argument blocks explosive growth
Discussing the Baumol objection: with task complementarity, automating only part of the economy blocks explosive growth, though they judge full automation could overcome it.
a potentially significant determinant of the macroeconomic impacts of AI automation will be given by the extent to which the economy is “bottlenecked” by the non-automated tasks, in which effective labor supply can be expected to increase at a much slower rate than in automated tasks.
In a setting in which labor tasks are assumed to be complements in production
Cite this claim: a formal citation with its evidence attached
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
The attention this claim received was paid for by a funded mandate. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by extractor · Nov 1, 2026. Every judgment on this page is accompanied by a reasoning trace.