Agentic AI for Autonomous Value Generation in Data Models: Reducing Manual Intervention in Predictive Pipelines
Predictive machine learning systems for structured data frequently depend on manual intervention when new observations arrive, input distributions change, prediction thresholds are modified, or model outputs must be regenerated. Although automated machine learning has reduced effort in model development, operational decisions surrounding deployed predictive pipelines often remain dependent on predefined procedures or human supervision. This paper proposes an agentic artificial intelligence layer for autonomous value generation in data-driven predictive pipelines. The proposed framework operates above conventional tabular learning models, including CatBoost and linear regression, and uses an intelligent agent to monitor incoming data and pipeline state, identify conditions requiring re-evaluation, select appropriate pipeline actions, execute those actions through controlled tools, and validate updated outputs before acceptance. The framework is grounded in two practical case studies: a CatBoost-based German Credit Risk Assessment application and a Linear Regression-based Crop Yield Prediction application, both deployed through Streamlit. The architecture is designed to remain model-agnostic and can also interface with XGBoost, LightGBM, stacking pipelines, and tabular foundation models. Rather than replacing the predictive model, the proposed system automates selected operational decisions around it. A simulated evaluation protocol is defined using intervention count, response time, update success, false-trigger rate, and human-approval frequency. The paper further argues that autonomy must be bounded by validation, auditability, and risk-sensitive human oversight, particularly for consequential applications such as credit assessment.
Authors
- Ronit Balouria
Publication Details
- Journal
- Dandao Xuebao/Journal of Ballistics
- Published
- 2026-11-16
- Primary Topic
- Data Stream Mining Techniques
- Type
- article
- Field-Weighted Citation Impact
- 0.00