When should AI automate work versus augment human decisions?

Fügener, Walzner and Gupta propose a task allocation framework that treats automation (AI substituting for humans) and augmentation (AI advising humans) as distinct roles whose value depends on which of two complementarity types is present. "The use of automation increases with higher levels of between-task complementarity," they write, while "the use of augmentation increases with higher levels of within-task complementarity" — between-task complementarity being performance differences between humans and AI across different tasks, within-task complementarity being the gain available from interacting on the same task. They validate the framework with an image-classification experiment: humans alone reached 68% accuracy, automation alone reached 77%, augmentation alone reached 80%, and the full framework — combining automation, augmentation, and reallocating the human resources automation frees up — reached 88%, "greatly outperforming the other options."

Automation's benefit splits into substitution (replacing humans where AI performs better) and reallocation (redeploying freed human resources to harder tasks); augmentation's benefit is narrower, limited to interaction improving human performance on the same task. The framework is scoped to judgment tasks meeting two conditions: the "Polanyi condition," where the optimal procedure cannot be formalized so humans and AI solve tasks differently, and the "ground truth condition," where tasks have one correct answer reachable through aggregation such as majority voting. Their model extends Boland et al.'s (1989) wisdom-of-crowds formulation to advice-taking, showing AI advice helps or hurts a crowd depending on whether the correct answer stays the most likely choice after the advice is factored in.

The paper explicitly borrows How should humans and AI systems divide decision-making tasks?'s framing of the human-AI pair as a "multiagent, goal-oriented system," then extends it by making complementarity type, rather than task type alone, the variable deciding whether automation or augmentation is the better role — a finer cut than Puranam's Type A/B/C typology. It also pushes back on the pattern behind When do human-AI teams actually outperform either alone?: "the general notion of previous human-AI research seems to be that augmentation is preferable to automation," the authors write, but their complementarity analysis makes the better role depend on which complementarity type dominates — and the 88% figure comes from splitting work by comparative advantage rather than the redundant whole-task designs that review blames for decision-task losses. Against Can humans and AI improve decisions by predicting together?, which finds ensembling can help without specialization, this paper's benefits are explicitly conditioned on complementarity existing — different regimes, not a conflict. Do humans and AI divide labor differently in knowledge work? offers a parallel typology organized around who picks the task and method, a different axis from this paper's complementarity cut.

The validation is a single experimental study of image classification — a generic, idealized judgment task chosen for tractability — so the 88% figure and the surrounding accuracy gaps support the framework's logic, not a general estimate for any workplace task allocation. The authors note the framework applies only where the Polanyi and ground-truth conditions hold, and that the complementarity driving its benefits rests on a knowledge gap between humans and AI that could narrow with AI improvements, task and process redesign, or loss of tacit knowledge — meaning any automation/augmentation split recommended today is not fixed and would need re-evaluating as those factors shift.