Conformalized prescriptive machine learning for uncertainty-aware automated decision making: the case of goodwill requests

Stefan Haas, Eyke HüllermeierInt. J. Data Sci. Anal 2025 · DOI

Introduction

In many industrial sectors, the abundance of historical transactional data makes the automation of manual business processes an attractive goal. However, in high-stakes financial domains—such as evaluating automotive "goodwill" warranty requests—the cost of an incorrect decision is significant. A car manufacturer must balance maintaining brand loyalty through customer compensation against the financial risk of unnecessary payouts. This paper introduces a framework that utilizes Conformal Prediction to safely automate these decisions by quantifying risk and selectively routing uncertain cases to human experts.

The Challenge of High-Stakes Automation

Standard machine learning models often lack a reliable mechanism to signal uncertainty, or to say "I don't know." In the context of goodwill requests, both epistemic (model-related) and aleatoric (inherent randomness) uncertainties can lead to errors with heavy financial and customer-satisfaction penalties. To address this, the authors propose a selective classification system designed to maximize the Degree of Automation (DoA) while strictly controlling the risk of incorrect decisions by knowing exactly when to abstain.

Proposed Framework: Hierarchical Conformal Prediction

The core of the approach is a hierarchical architecture wrapped in Inductive Conformal Prediction (ICP). This design mimics human decision-making through two distinct layers:

  1. Ranking Layer: An XGBoost model performs a qualitative classification, categorizing requests into "NO", "PARTIAL", or "FULL" contribution.
  2. Regression Layer: For requests classified as "PARTIAL", a Feed-Forward Neural Network predicts the specific percentage of contribution (ranging from 10% to 90%).

By applying conformal predictors to both layers, the system generates prediction sets for the ranking layer and prediction intervals for the regression layer. The size of these sets or the width of the intervals serves as a proxy for uncertainty. If the uncertainty exceeds a predefined threshold, or if the predictor outputs an "empty set" (indicating an out-of-distribution anomaly), the request is automatically routed for human review.

Data Analysis and Model Architecture

The framework was evaluated using a proprietary dataset of 15,397 historical goodwill requests from a BMW National Sales Company. The data includes 24 features detailing vehicle specifications and mechanical issues, with labels representing human expert decisions on cost contributions for both labor and parts.

Distribution of goodwill contributions across raw, collapsed, and partial categories.

As shown in the figure above, the distribution of contributions is highly skewed, with significant peaks at 0% (NO) and 100% (FULL). The hierarchical model effectively handles this by first addressing the broad categories before attempting fine-grained regression on the partial cases. Notably, the data reveals a human tendency to default to a 50% contribution, which the regression layer must navigate.

Performance and Business Impact

The system offers a tunable, Pareto-optimal trade-off between automation and accuracy. Key results include:

  • Baseline Performance: Without any abstention (100% automation), the model achieves an accuracy of ~91-93%.
  • Selective Automation: By tuning the conformal risk thresholds to abstain on the 30% most uncertain queries, the accuracy for the automatically processed requests jumps to 98%.
  • Error Reduction: In the regression layer, Mean Absolute Error (MAE) decreases monotonically as the degree of automation is reduced, confirming that conformal intervals accurately capture model uncertainty.

One critical insight is the utility of "empty sets" in conformal prediction. These act as a built-in safety net, identifying edge cases that do not conform to training patterns and ensuring they are handled by humans rather than being misclassified.

Conclusion and Future Directions

This research demonstrates that conformalized prescriptive ML can bridge the gap between experimental models and production-ready automated decision-making. By presenting a Pareto front of "Accuracy vs. Automation," business leaders can choose the exact trade-off that fits their operational capacity.

Future work will focus on treating human labels as weak supervision to account for inherent biases and integrating Explainable AI (XAI) to provide transparency into why specific cases are deemed uncertain. Additionally, the human-AI interaction dynamics, such as automation bias, remain a vital area for further study.

Summary updated

BibTeX
@article{haas2025conformalized,
  author = {Stefan Haas and Eyke Hüllermeier},
  title = {{Conformalized prescriptive machine learning for uncertainty-aware automated decision making: the case of goodwill requests}},
  journal = {Int. J. Data Sci. Anal.},
  year = {2025},
  doi = {10.1007/s41060-024-00573-2},
}
Figure