Selective automation keeps humans in the AI learning loop
Industrial companies do not have to choose between fully automated decision-making and permanent human intervention. Research by Kai Hoberg of Kühne Logistics University, Christina Imdahl of Eindhoven University of Technology and William Schmidt of Emory University proposes a more selective model: automate decisions where people add little value, but retain them where their judgement can improve the system.
The approach links operational efficiency with the continuing development of artificial intelligence. Instead of treating employees simply as a fallback for difficult cases, it identifies situations in which human corrections are most likely to produce better outcomes and useful learning data.
Four stages for targeted automation
The researchers’ framework begins by examining when human changes to algorithmic recommendations improve decisions. A predictive model then learns to estimate the likely value of expert intervention for new cases. Decisions with limited expected benefit from human involvement can be automated, while cases with greater potential value remain subject to human review.
The final step is a feedback cycle: human corrections are returned to the algorithm so that its recommendations can improve over time. This design is intended to prevent automation from becoming static as operating conditions change.
Evidence from commercial-vehicle operations
The study examined more than 475,000 order-related decisions recorded across four years at a major commercial vehicle manufacturer whose annual revenue exceeded €4 billion. Rather than asking whether people or AI perform better overall, the researchers looked for the conditions under which human decisions outperformed algorithmic recommendations.
In the empirical application, approximately 84% of decisions could be automated. The finding does not support removing people from the process altogether. It suggests that automation rates can be high when human attention is directed towards cases where it has measurable impact.
For industrial leaders, the practical starting point is to measure the added value of intervention before setting automation targets. That evidence can guide the allocation of expertise, reduce repetitive work and maintain a route for algorithms to adapt as demand, products and operating environments evolve. Kühne Logistics University is among the institutions involved in the research.
