Causal attribution of workload reductions to language models

Determine the causal share of any observed reduction in governance-review workload that is attributable specifically to language models rather than to process redesign, outsourcing, demand changes, or other forms of automation.

Background

The paper separates observing a reduction in total human work from attributing that reduction to agents or LLMs. A deployment may satisfy the workload threshold while simultaneous organizational changes, including process redesign, outsourcing, demand shifts, or non-language-model automation, contribute to the observed reduction.

The authors state that the causal contribution of LLMs remains unresolved and recommend staged or randomized adoption, where feasible, to identify changes that would otherwise have occurred. This is therefore an explicit unresolved empirical question concerning causal attribution, not merely a general recommendation for future work.