Organizations are rapidly moving from predictive Artificial Intelligence (AI) toward Generative Artificial Intelligence (GenAI) and increasingly agentic systems that can recommend, create, coordinate, and sometimes execute managerial actions. Yet higher model capability does not by itself produce better organizational decisions: underreliance wastes useful machine competence, whereas overreliance can amplify hallucination, bias, automation complacency, accountability diffusion, and skill erosion. This conceptual study develops a decision-calibration framework for human–AI collaboration that treats the central design problem as aligning the degree of AI influence with verified AI competence, task risk, decision reversibility, evidentiary verifiability, and human expertise. The study integrates research on organizational decision design, trust in automation, algorithm aversion and appreciation, explainable AI, responsible AI governance, and recent evidence on GenAI-enabled knowledge work. The resulting framework distinguishes four operating modes, human-led, AI-augmented, AI-led with human veto, and dual-control, and links them to a lifecycle governance architecture comprising task qualification, evidence-grounded generation, confidence and uncertainty signaling, independent verification, escalation, audit logging, and post-decision learning. Eight propositions explain how calibration quality should affect decision accuracy, speed, accountability, organizational learning, and resilience. The paper further introduces practical metrics for calibration error, override quality, verification coverage, and decision reversibility. The framework shifts managerial attention from the simplistic question of whether humans or AI should decide toward the more consequential question of when, how, and under what governance conditions each should influence a decision. This perspective offers a theoretically grounded and operationally actionable basis for designing reliable human–AI decision systems in contemporary organizations.