The intellectual roots of human–AI decision design precede contemporary machine learning. Automation research distinguishes stages such as information acquisition, information analysis, decision and action selection, and implementation, allowing different levels of machine involvement at each stage [9]. This decomposition remains valuable for GenAI because a system can be permitted to retrieve information and draft alternatives without being permitted to select or execute an action. The literature on trust in automation further argues that reliance should correspond to the system's capabilities and the task environment, not to generalized positive attitudes [10]. Longitudinal trust is shaped by prior experience, system performance, user disposition, and situational factors [11]. Empirical work also demonstrates that trust affects whether people rely on automated advice [12], [13].

In organizational research, the central shift has been from 'technology use' to joint decision production. Shrestha et al. [3] identify delegation, sequential hybrid arrangements, and aggregated human–AI decision structures.

Puranam [5] shows that division of labor and learning configurations are organization-design choices, while Raisch and Krakowski [2] caution that automation and augmentation are mutually entangled rather than mutually exclusive. The implication is that the same organization may need different combinations of AI and human authority across tasks and over time. This is particularly important when AI changes the knowledge distribution itself: learning algorithms may alter who is considered expert, which activities remain visible, and how decisions are monitored [6], [8].

Behavioral research demonstrates that reliance on algorithms is neither uniformly low nor uniformly high. Dietvorst et al. [14] show algorithm aversion after people observe an algorithm make errors. In contrast, Logg et al. [15] find algorithm appreciation in several judgment contexts, with people sometimes weighting algorithmic advice more than human advice. Giving users even limited control over algorithmic forecasts can reduce aversion [16], and task characteristics matter: algorithms are trusted less for tasks perceived as subjective than for tasks perceived as objective [17]. A systematic review by Burton et al. [18] organizes causes and remedies around expectations and expertise, autonomy, incentives, cognitive compatibility, and divergent rationalities.

These findings imply that adoption is a poor proxy for decision quality. A manager can overrule a superior algorithm because of aversion, or accept a weak output because of automation bias or the persuasive fluency of a generated explanation. The managerial problem is therefore bidirectional: Preventing both disuse and misuse. Calibration requires the organization to identify when the human is likely to be wrong about the AI, not only when the AI is likely to be wrong about the task.

Explainability is frequently proposed as a remedy for inappropriate reliance, but evidence is more nuanced. Alufaisan et al. [19] found that AI predictions could improve human decision accuracy in their experimental setting, whereas adding explanations did not yield conclusive additional gains. Bansal et al. [20] similarly show that explanations can increase acceptance of AI recommendations without necessarily improving complementary team performance. Buçinca et al. [21] demonstrate that cognitive forcing functions can reduce overreliance, but at a user-experience cost. These studies suggest that explanation should not be treated as a universal trust-building mechanism; it is a decision-control intervention whose value depends on whether it helps users discriminate correct from incorrect AI advice.

Human–AI interaction guidelines emphasize expectation setting, feedback, correction, and appropriate control across the interaction lifecycle [22]. The human-centered AI perspective likewise argues that high levels of automation can coexist with high levels of meaningful human control [23]. Together, these literatures support a governance design in which users are not merely shown an explanation; they are given the evidence, uncertainty cues, checkpoints, and authority needed to verify consequential outputs.

Responsible AI research has achieved substantial convergence on broad principles while remaining less settled on operationalization. A global review of AI ethics guidelines identifies recurring principles such as transparency, justice and fairness, non-maleficence, responsibility, and privacy [24]. The AI4People framework similarly articulates beneficence, non-maleficence, autonomy, justice, and explicability [25]. Yet principles alone cannot guarantee ethical AI because implementation requires institutional mechanisms, professional norms, incentives, and accountability processes [26].

Reviews of responsible-AI tools show a continuing challenge in translating high-level principles into concrete methods and practices [27]. End-to-end internal algorithmic auditing is one response because it embeds documentation and accountability throughout the system lifecycle [28]. The National Institute of Standards and Technology (NIST) AI risk management framework provides a complementary governance vocabulary organized around govern, map, measure, and manage activities [29].

Recent information-systems scholarship reinforces this operational focus. Responsible AI governance can be understood through structural, relational, and procedural practices across the AI lifecycle [30]. The remaining gap is to connect governance practices directly to decision rights and reliance behavior. A policy may require fairness testing and audit logs while leaving unanswered a basic operational question: When a specific AI recommendation is uncertain, who is required to verify it, who may overrule whom, and what evidence is necessary before execution? Decision calibration addresses this missing link.

The diffusion of GenAI raises the stakes of calibration because evidence shows both substantial productivity gains and task-dependent limitations. Noy and Zhang [31] report that access to a generative AI tool improved speed and quality in professional writing tasks.

Brynjolfsson et al. [32] find productivity gains from generative AI assistance in customer support, with particularly strong benefits for less experienced workers. A large field experiment on product innovation shows that individuals with AI can match the performance of human teams without AI and that AI can help bridge functional knowledge silos [33].

However, survey evidence among knowledge workers indicates that higher confidence in GenAI is associated with lower reported critical-thinking effort, while task stewardship increasingly shifts toward verification and integration [34].

These results imply a central paradox. GenAI can democratize expertise and compress coordination costs, yet the same convenience can reduce the cognitive effort applied to verification. In team settings, AI is also evolving from tool to teammate [35]. As organizations move further toward agentic AI, the locus of risk changes from generated content to generated action. Decision calibration must therefore govern not only whether a recommendation is believed, but whether an AI system is permitted to act.

Decision calibration is the alignment between the influence granted to AI in a specific decision and the system's verified competence for that decision, conditional on the decision's risk, reversibility, verifiability, and the availability of competent human oversight. This definition differs from trust calibration in two ways. First, the unit of analysis is not only the user's psychological trust; it is the organizationally designed influence that the AI actually exerts. Second, the target is not a subjective belief alone but a governance outcome: whether decision rights, verification effort, and execution authority are proportionate to demonstrated

Let I denote AI influence on a bounded scale from 0 (no decision influence) to 1 (full autonomous authority). Let C denote verified task-specific AI competence on a comparable scale. A simple descriptive calibration error can be written as CE = |I − C|. This quantity is not a complete risk measure because the same mismatch matters more in a high-impact irreversible decision than in a low-impact reversible one.

Accordingly, organizations should use a Risk-weighted Calibration Error (RCE), RCE = W × |I − C|, where W is a context weight derived from impact severity, irreversibility, uncertainty, and the weakness of independent verification. The purpose of these expressions is managerial rather than predictive: they make explicit that excessive AI authority and excessive human exclusion are governance choices that can be Competence must be task-specific and evidence-based. A single benchmark score for a foundation model is insufficient because managerial tasks combine factual retrieval, interpretation, synthesis, causal reasoning, forecasting, stakeholder judgment, and execution. The relevant question is whether the configured system, model, prompt, retrieval layer, tools, data, and safeguards perform adequately on the class of decision episodes it is being asked to influence. Competence should therefore be estimated from validated historical cases, shadow-mode testing, adversarial tests, and post-deployment monitoring rather than vendor claims or user impressions.

Influence also has multiple components. An AI system can influence a decision by framing the problem, selecting evidence, generating alternatives, ranking alternatives, recommending an action, drafting a justification, or executing the selected action. Organizations should avoid collapsing these into one binary variable called 'AI use.' A system may legitimately have high influence over information search but low influence over execution, especially when the decision is consequential or difficult to reverse.

The framework identifies four operating modes. In the human-led mode, AI may retrieve, summarize, or simulate, but the human frames the decision and selects the action. This mode is appropriate when task competence is uncertain, contextual, or moral judgment is dominant, or consequences are difficult to reverse. In the AI-augmented mode, AI provides recommendations or alternatives and the human makes the final choice after specified verification. This is likely to be the default mode for many managerial decisions because it combines machine scale with human accountability.

In the AI-led with human-veto mode, the system selects or executes an action unless a human intervenes. This mode is appropriate only when competence is well validated, monitoring is continuous, and actions are reversible or tightly bounded. Examples include routine replenishment within limits, classification with post-hoc review, or workflow routing. Finally, dual-control requires both AI and an independent human (or two independent controls) to authorize an action. It is appropriate for high-impact or high-uncertainty decisions where neither fully autonomous AI nor unstructured human discretion is acceptable. Dual-control is not simply 'more oversight'; it is a designed separation of proposal, verification, and authorization roles.

Stage 1 (task qualification). Define the decision class, its consequence level, reversibility, evidence requirements, excluded uses, and minimum human expertise. No model should be granted authority before the task itself is qualified.

Stage 2 (evidence-grounded generation). Constrain AI to approved information sources when factual accuracy matters. Record provenance for retrieved evidence and distinguish retrieved facts from model-generated interpretation.

Stage 3 (uncertainty and limitation signaling). Require the system interface to disclose uncertainty, missing evidence, conflicts between sources, and conditions outside the validated task envelope. Confidence should not be inferred from linguistic fluency.

Stage 4 (independent verification). Specify what must be checked, by whom, and against what source. Verification coverage should rise with consequence and irreversibility. Critical outputs should be checked independently rather than by asking the same model to self-confirm.

Stage 5 (decision-right allocation and escalation). Route the episode to one of the four operating modes. Escalate when confidence falls, evidence conflicts, the task is novel, policy limits are reached, or the proposed action crosses an impact threshold.

Stage 6 (execution safeguards). Use sandboxing, transaction limits, staged commitments, human approval gates, and reversible actions to prevent a reasoning error from becoming an uncontrolled real-world action. Stage 7 (audit logging and post-decision learning). Log prompts, evidence, recommendations, overrides, approvals, tool calls, and outcomes. Use these records to update task competence estimates and detect drift, recurring human override failures, or emerging automation bias.

The architecture is depicted in Fig. 2 as a closed-loop system. The feedback loop is essential: governance should update when either the model changes or the organization learns that users are interacting with it differently than expected. A model upgrade can invalidate earlier competence evidence; similarly, experienced users may become more efficient verifiers over time, or conversely may become complacent.

Proposition 1. Decision performance will be higher when AI influence is calibrated to verified task-specific competence than when influence is determined by generalized trust, user preference, or a fixed organization-wide automation policy.

Proposition 2. The negative effect of AI competence errors on decision outcomes will increase as decision consequence and irreversibility increase.

Proposition 3. Independent verifiability will increase the safe level of AI influence because it lowers the probability that an unsupported output propagates into action.

Proposition 4. Explanation features will improve human–AI decision performance only to the extent that they increase users' ability to discriminate correct from incorrect AI recommendations rather than merely increasing acceptance.

Proposition 5. Human oversight will reduce AI-related error only when reviewers possess relevant expertise, adequate time, independence, and genuine authority to reject or escalate the AI output.

Proposition 6. Reversible execution and bounded tool permissions will permit higher levels of agentic AI autonomy without proportionally increasing organizational risk.

Proposition 7. Organizations that record and analyze both human overrides and non-overrides will achieve better calibration over time than organizations that monitor only model accuracy.

Proposition 8. Excessive reliance on GenAI will reduce critical verification effort and weaken organizational learning unless workflows explicitly preserve problem formulation, evidence checking, and post-decision reflection as human responsibilities.

A framework becomes managerial only when it can be monitored. Table 2 proposes a compact measurement system. None of the metrics should be interpreted in isolation. For example, a low human override rate can indicate either excellent AI performance or severe automation complacency. The metrics must therefore be read together with verified task performance and decision outcomes.

between granted AI governance mismatch. units or decision influence and verified classes.

competence.

Verification Share of required AI Low coverage is Audit adherence to coverage claims/actions dangerous in high-impact verification policy.

independently checked. contexts.

Unsupported- Share of material AI Signals grounding or Track by model, claim rate claims lacking hallucination risk. prompt, and task authoritative evidence. class.

Override Share of human overrides Separates useful oversight Train reviewers and precision that improve the final from reflexive aversion. refine escalation rules.

outcome.

Missed- Share of harmful AI Signals overreliance or Redesign interface and override rate recommendations that weak review. cognitive forcing.

humans failed to stop.

Reversible- Share of agentic actions Higher values expand safe Prioritize reversible action share that can be automatically autonomy. workflow engineering.

rolled back or safely staged.

Outcome Time from observed Long latency slows Improve monitoring learning outcome to organizational adaptation. and feedback latency governance/model ownership.

update.

The framework reframes several established debates. First, it moves beyond the automation-versus-augmentation dichotomy by treating both as variable configurations of influence. The relevant theoretical object is not whether AI replaces or supports humans, but how problem framing, evidence selection, recommendation, authorization, and execution rights are distributed. This is consistent with the organizational perspective that AI changes structures of decision making [3–5] while extending it through a measurable calibration construct.

Second, the framework distinguishes trust from governance. Trust research remains essential for predicting reliance [10–13], but an organization cannot safely rely on psychological calibration alone. Users differ in expertise, incentives, risk tolerance, and exposure to automation. Organizational controls must therefore be designed so that an overconfident user cannot grant unlimited authority to a weak system and an algorithm-averse user cannot silently discard superior evidence without accountability. Calibration becomes a property of the socio-technical system, not only of individual cognition.

Third, the framework clarifies the role of explainability. The goal of explanation should be error discrimination and contestability, not generalized persuasion. This interpretation is consistent with evidence that explanations may increase acceptance without improving complementary performance [19–21]. In management contexts, useful explanations should expose decision-relevant evidence, assumptions, uncertainty, and counterfactual sensitivity in a form that permits a qualified reviewer to challenge the recommendation.

Fourth, the framework links responsible AI to everyday decision operations. Ethical principles and lifecycle governance are necessary [24–29], [36] but organizations ultimately experience AI risk through specific decisions and actions. By tying governance controls to decision rights, the framework offers a bridge between policy-level responsible AI and operational management.

For executives, the immediate implication is to stop treating 'human in the loop' as a sufficient control. A human presence is meaningful only when the role is defined: What must be reviewed? What evidence must be checked? What decisions can be vetoed? How much time is allocated? What happens when the reviewer disagrees with the AI? Organizations should replace generic human-in-the-loop policies with explicit decision-right matrices.

For AI product owners and analytics teams, performance evaluation should move from model-centric dashboards to decision-centric dashboards. Average benchmark accuracy should be supplemented with unsupported-claim rates, error severity, distribution-shift indicators, human override precision, and downstream outcome measures. This is especially important for GenAI systems because linguistic quality can remain high even when factual reliability deteriorates.

For internal audit and compliance, AI logs should record the evidence path and the authority path. Evidence path means what information the system used and which claims were verified. Authority path means who proposed, approved, modified, escalated, and executed the action. Together these records allow an organization to reconstruct not just what the model said, but why the socio-technical system allowed the decision to occur.

For workforce design, GenAI should be deployed in ways that preserve learning. Productivity gains are valuable [31–33], but work should not be redesigned so that junior employees only approve outputs they no longer know how to independently produce or evaluate. The evidence that GenAI use can shift critical-thinking effort toward verification and integration [34] suggests a need for deliberate skill architecture: organizations should preserve opportunities for problem formulation, first-principles reasoning, and independent judgment while using AI to accelerate search, drafting, simulation, and routine synthesis. Fig. 3 provides a practical routing matrix. As consequence and irreversibility increase, governance should move upward from autonomous or AI-led modes toward dual-control and human-led modes unless verifiability and competence are exceptionally strong. Conversely, low-impact and easily reversible tasks can support more autonomous agentic execution, provided monitoring remains active.

The framework generates a research agenda across individual, team, organizational, and institutional levels. At the individual level, experiments should compare generalized trust interventions with calibration-oriented interventions such as mandatory evidence checks, uncertainty displays, delayed reveal of AI advice, or justification-before-advice designs. The dependent variables should include complementary accuracy and error discrimination, not only trust or acceptance.

At the team level, research should examine whether AI reduces or redistributes functional diversity. Recent field evidence indicates that AI can bridge expertise silos [33], but it remains unclear whether prolonged use homogenizes reasoning or suppresses minority expertise. Studies could compare teams in which AI is a shared resource with teams in which members use independent agents, testing effects on information pooling, dissent, and coordination.

At the organizational level, longitudinal studies should test whether calibration metrics predict operational incidents, productivity, innovation, and learning. Natural experiments created by staged AI rollouts could compare business units using fixed automation rules with units using dynamic routing based on consequence, reversibility, and verified competence. Researchers should also examine governance drift: whether approval rates rise over time as users become accustomed to AI, even when model reliability does not change.

At the institutional level, future research should study how regulation, professional liability, standards, and procurement rules shape calibration choices. High-level governance principles are increasingly common, but organizations may operationalize them differently depending on sector, country, and risk exposure. Comparative research can identify when structural, relational, and procedural governance practices [30] actually change the influence granted to AI in real decisions.

Finally, agentic AI requires research methods that capture action chains rather than isolated prompts. A consequential failure may emerge from a sequence of individually plausible steps: an agent retrieves incomplete data, forms an incorrect plan, invokes a tool, receives a misleading response, and commits an action. Evaluation should therefore measure chain-level reliability, intervention opportunities, rollback effectiveness, and the degree to which human supervisors can reconstruct and contest the agent's reasoning and actions.