Business analytics established the organizational logic on which ADI builds. Analytics creates value when firms develop not only technical infrastructure but also data quality, managerial skills, cross-functional routines, and mechanisms for translating insights into action [3–5]. This literature is important because it rejects technological determinism: Analytical capability is a configuration of resources and routines rather than a software acquisition. Agentic systems intensify this logic because the distance between model output and organizational action becomes shorter. When an AI system can execute as well as recommend, weaknesses in data governance, process ownership, or incentive alignment can propagate directly into operational consequences.

Traditional analytics can be organized around descriptive, predictive, and prescriptive functions. Descriptive analytics explains what has happened; predictive analytics estimates what is likely to happen; prescriptive analytics recommends what should be done. Agentic systems add a fourth operational dimension: controlled enactment. The system can decide how to pursue a delegated objective, invoke tools, coordinate subtasks, and revise execution based on observed feedback. This progression is summarized in Table 1. The distinction matters because the accountability burden changes as systems move from informational influence to behavioral intervention.

What may the system Delegate and Plan, use tools, ADI decide and execute? supervise act, learn

As Table 1 indicates, the defining change is not that AI becomes “smarter” in the abstract, but that organizational systems become capable of translating representations of managerial goals into sequences of consequential actions. This is precisely where established work on algorithmic management becomes relevant. Learning algorithms can reshape expertise, occupational boundaries, coordination, and control [11], while workplace algorithms create contested terrain around direction, evaluation, discipline, and worker autonomy [12]. ADI must therefore be designed as an organizational control system, not simply an advanced interface to a model.

Research on human–AI complementarity provides the second conceptual foundation. Human decision makers are generally stronger in contextual interpretation, moral reasoning, handling ambiguous goals, and negotiating equivocal social situations, whereas AI systems can process large information spaces, apply consistent rules, and rapidly evaluate alternatives [7]. Hybrid intelligence research formalizes the idea that human and machine capabilities can be deliberately combined so that the joint system achieves outcomes neither component could reliably produce alone [13]. Related work on machines as teammates emphasizes that AI artifacts alter not only task allocation but also communication, coordination, and institutional arrangements [14].

However, complementarity does not imply that more autonomy is always better. Classic automation research distinguishes different types and levels of automation and shows that the optimal allocation of functions depends on reliability, consequences, and human-performance effects [27]. Trust research likewise demonstrates that effective automation requires calibrated reliance rather than maximal trust [26]. In AI contexts, trust is shaped by the system's representation, capabilities, performance, process transparency, and user expectations [28]. Under-trust wastes useful automation; over-trust creates automation bias and weakens human monitoring. The implication for agentic systems is that delegation should be dynamically bounded by evidence about competence and risk.

Generative AI strengthens this argument. Controlled experiments have shown substantial gains in speed and quality for some professional writing tasks [15], while workplace deployment in customer support has produced average productivity gains and disproportionately benefited less experienced workers [16]. Yet the “jagged technological frontier” observed in knowledge work means that seemingly similar tasks can differ sharply in whether current AI improves or degrades performance [17]. The governance architecture of ADI must therefore support selective delegation: autonomy should be high where competence is established and consequences are containable, but constrained where capability is uncertain or failure is costly.

The third foundation concerns generative AI and autonomous agents. Large-scale foundation models create reusable capabilities across many downstream tasks, but they also concentrate risks because model limitations can propagate across applications [19]. The multidisciplinary literature on generative conversational AI highlights opportunities in productivity, creativity, decision support, and knowledge access alongside concerns regarding hallucination, bias, privacy, intellectual property, accountability, and misuse [18]. These characteristics make LLMs powerful cognitive interfaces but insufficient, by themselves, for reliable organizational agency.

Agentic architectures extend LLMs by adding memory, planning, tool use, environmental interaction, and feedback. Surveys of LLM-based autonomous agents converge on components such as perception, planning, memory, and action, while noting unresolved problems in evaluation, safety, long-horizon reliability, and social interaction [20], [21]. ReAct demonstrates how reasoning and acting can be interleaved so that a language model interacts with external environments rather than relying solely on internal generation [22]. Toolformer shows that language models can learn when and how to call external application programming interfaces, thereby combining general language capability with specialized tools [23]. Retrieval-Augmented Generation (RAG) connects parametric models to external knowledge stores, improving access to current or provenance-bearing information [24]. These techniques collectively enable a shift from conversational assistance toward action-oriented systems.

For management analytics, the technical significance of these advances is secondary to their organizational implication: An AI system can now traverse multiple decision stages that were previously separated by human handoffs. It can retrieve a policy, inspect a dashboard, calculate an indicator, draft a recommendation, open a workflow, request approval, and write the resulting state back to a business system. This compression of the decision cycle creates value through speed and consistency but can also create tightly coupled failures. ADI therefore requires architectural separation between evidence, reasoning, authority, and execution.

Human–AI interaction research emphasizes that systems should communicate what they can do, explain relevant behavior, support efficient correction, and adapt while preserving user control [25]. Explainability becomes especially important when AI outputs affect consequential decisions, although post-hoc explanations cannot substitute for inherently understandable models in every high-stakes setting [29]. Explainable Artificial Intelligence (XAI) research provides taxonomies of methods and identifies trade-offs between interpretability, performance, fidelity, and user needs [30]. For agentic systems, explanation must extend beyond why a prediction was generated to include what evidence was used, which tools were called, what actions were taken, and how the system's authority was determined.

Ethical and governance literatures reinforce this lifecycle perspective. AI ethics frameworks commonly converge on principles such as beneficence, non-maleficence, autonomy, justice, and explicability [31], while global guideline analyses identify recurring commitments to transparency, fairness, non-maleficence, responsibility, and privacy [32]. Earlier algorithm-ethics work also emphasizes opacity, bias, discrimination, autonomy, privacy, and responsibility as distinct but connected concerns [38]. The National Institute of Standards and Technology (NIST) Artificial Intelligence Risk Management Framework (AIRMF) operationalizes trustworthy AI around governance, mapping, measurement, and management [33], and its Generative Artificial Intelligence Profile extends the framework to generative-specific risks [34]. ISO/IEC 42001:2023 formalizes requirements for an Artificial Intelligence Management System (AIMS) [35]. The Organisation for Economic Co-operation and Development (OECD) AI principles, updated in 2024, emphasize trustworthy, human-centric AI [36], while the European Union AI Act establishes a risk-based regulatory regime for AI systems [37].

These sources point to a core conceptual conclusion: Governance cannot be reduced to an approval gate placed after technical development. Once AI systems can initiate actions, governance must be expressed as executable constraints within the architecture itself. Permissions, evidence requirements, human escalation, logging, and monitoring become part of the decision system's functional design. This conclusion motivates the ADI framework developed next.

ADI is defined here as an organizational capability in which AI-enabled systems participate in a governance-bounded loop of sensing, evidence construction, reasoning, tool-mediated action, and learning in pursuit of delegated managerial objectives. Three elements distinguish ADI from ordinary generative assistance. First, the system is goal-directed: It receives an objective or decision mandate rather than only a prompt for content. Second, it is action-capable: It can invoke tools or workflows that alter an organizational state. Third, its autonomy is conditional: Authority is constrained by explicit policies, risk thresholds, evidence requirements, and escalation rules. ADI is therefore not synonymous with “autonomous AI.” It is a socio-technical configuration in which autonomy is granted selectively and remains auditable.

This definition has important boundary conditions. A chatbot that drafts a memo but cannot access authoritative organizational knowledge or initiate an action is generative assistance, not ADI. A robotic process automation script that executes deterministic rules without adaptive reasoning is automation, not ADI. A predictive model that triggers a pre-approved action may be algorithmic automation, but it becomes ADI only when it participates in a broader loop involving contextual evidence, adaptive planning, or conditional decision rights. The framework is consequently oriented toward systems that combine heterogeneous reasoning and action capabilities under organizational control.

Fig. 1 presents the proposed architecture. The layers are deliberately separated because technical integration without governance separation encourages hidden coupling. Each layer has a distinct managerial purpose and produces artifacts that can be inspected independently.

an incorrectly framed problem.

Core Function Layer Capture events, goals, 1. Organizational states, and constraints sensing Retrieve and structure 2. Evidence and decision evidence knowledge Generate alternatives, 3. Reasoning and evaluate constraints, decision design plan Invoke tools and alter 4. Permissioned organizational state execution

Evaluate outcomes 5. Learning and and adapt safely assurance

Layer 1, organizational sensing, converts the environment into a decision context. Inputs may include structured enterprise data, events from operational systems, unstructured documents, user requests, customer communications, sensor streams, or managerial objectives. The central design requirement is not maximal data ingestion but contextual sufficiency. Agentic systems should know which objective they are pursuing, what constraints apply, which organizational unit owns the decision, and when the sensed state becomes stale. This layer prevents a common failure mode in which an apparently intelligent system reasons correctly about

Layer 2 constructs an evidence base. Retrieval should distinguish authoritative internal sources, external information, model-generated hypotheses, and user-provided statements. RAG and related retrieval mechanisms are valuable because they can connect language-model reasoning to explicit knowledge stores [24], but retrieval alone does not create trustworthy evidence. Each evidence item should carry provenance, time, access status, and where possible, an estimate of reliability. The goal is to make factual grounding inspectable and to prevent generated content from being silently promoted into organizational fact.

Layer 3 performs reasoning and decision design. The system decomposes objectives, generates alternatives, checks constraints, and estimates expected consequences. Here, LLM reasoning may be combined with statistical models, optimization, simulation, business rules, or human judgment. This heterogeneous architecture is preferable to treating the LLM as a universal decision engine. In high-stakes decisions, interpretable models or explicit rules may be required even when a generative model is used to orchestrate the workflow [29], [30]. The output of Layer 3 should be a decision object: proposed action, supporting evidence, assumptions, confidence or uncertainty indicators, and the required level of authority.

Layer 4 is permissioned execution. It is the defining operational layer of ADI because recommendations become actions. Tool calls may create purchase orders, modify schedules, contact customers, update records, deploy software, allocate resources, or initiate approvals. Least-privilege design is essential: the agent should receive only the permissions required for the current task, with transaction limits and reversible operations where possible. The system must never infer authority solely from technical capability. Organizational policy, not model confidence, determines whether a tool may be used.

Layer 5 combines learning and assurance. Agentic systems require feedback on both task success and governance quality. A decision can achieve its immediate objective while violating process requirements, creating downstream harm, or weakening accountability. Evaluation should therefore include outcome metrics, process metrics, exception rates, human overrides, escalation frequency, unauthorized-attempt counts, evidence quality, and drift indicators. Learning from feedback should itself be governed: incident data can inform improved prompts, rules, retrieval, models, and policies, but self-modification without controlled validation would undermine assurance.

The architecture operates as a closed loop rather than a linear pipeline. Fig. 2 depicts the cycle. A trigger creates a decision episode; evidence is assembled; alternatives and plans are produced; a governance function determines allowable authority; the system either executes, requests human approval, or escalates; outcomes are observed; and the assurance layer updates performance records and, where appropriate, prompts organizational learning. The loop can be repeated until the objective is achieved, abandoned, or transferred to a human owner.

The cycle reframes human oversight. Rather than forcing a person to approve every action, which can create meaningless “rubber-stamp” supervision, human involvement is concentrated where judgment adds value. This principle follows the literature on calibrated trust and levels of automation [26–28]. Humans should intervene when uncertainty, novelty, value conflict, irreversible consequences, or authority boundaries exceed pre-defined tolerances. Routine, well-evidenced, reversible tasks can be delegated more extensively. Oversight thus becomes risk-sensitive and exception-oriented.

The central governance proposition of ADI is that autonomy should be allocated by decision characteristics rather than by application category alone. Four factors jointly determine an appropriate level of machine decision rights: Consequence severity, environmental uncertainty, action reversibility, and evidentiary strength. Consequence severity captures the magnitude and distribution of potential harm or loss. Environmental uncertainty captures how incomplete, unstable, or ambiguous the decision context is. Reversibility captures whether an erroneous action can be promptly and fully undone. Evidentiary strength captures the quality, recency, consistency, and provenance of information supporting the action.

These factors produce three practical authority zones. In the autonomous zone, consequences are limited, evidence is strong, uncertainty is low to moderate, and actions are reversible. The agent may execute within predefined limits. In the supervised zone, at least one risk factor is elevated; the system may prepare and partially execute the workflow but requires human authorization at a designated checkpoint. In the prohibited or specialist zone, consequences are severe, evidence is weak, uncertainty is high, or the action is effectively irreversible; the system may support analysis but lacks authority to execute. Fig. 3 shows this concept as a governance envelope rather than a binary human-versus-machine choice.

This model operationalizes a recurring theme in AI governance: Risk controls must be proportional to context. NIST frames AI risk management as a continuous process involving governance, mapping, measurement, and management [33], while NIST guidance for generative AI emphasizes lifecycle-specific risks and mitigations [34]. ISO/IEC 42001:2023 similarly treats AI governance as an organizational management system rather than a one-time technical assessment [35]. The ADI model translates these general principles into decision-level allocation of authority.

The first mechanism is provenance. Every consequential decision episode should preserve a trace of the evidence used, model or tool versions involved, policy rules applied, prompts or structured instructions that materially affected the result, and human interventions. Provenance supports diagnosis, contestability, audit, and organizational learning. It also reduces a central danger of generative systems: Fluent outputs can obscure the distinction between retrieved fact, model inference, and fabricated content [18], [19], [40].

The second mechanism is permissioning. Agentic systems should follow least-privilege principles. Tool access should be scoped to role, task, data class, transaction value, and time. Permission boundaries must be enforced outside the language model whenever possible. A prompt stating “do not exceed $10,000” is weaker than a transactional control that technically prevents a larger commitment. Permissioning converts governance from language into infrastructure.

The third mechanism is escalation. Escalation should be triggered by explicit conditions such as conflicting evidence, low confidence, novel states, policy ambiguity, high-value transactions, sensitive data, unusual user requests, or failed validation checks. Effective escalation includes a clear human owner and a compact decision packet containing the proposed action, evidence, unresolved uncertainty, and reason for escalation. This is consistent with human–AI interaction guidance that systems should make their capabilities and limitations visible and facilitate user correction [25].

The fourth mechanism is continuous assurance. Pre-deployment testing is necessary but insufficient for systems that operate in changing organizational environments. Assurance should track technical performance, governance performance, and social consequences. Metrics may include outcome quality, override rates, escalation rates, incident severity, false-action frequency, retrieval provenance coverage, policy violations prevented, time-to-resolution, and differential performance across groups or contexts. The aim is calibrated autonomy: empirical evidence about performance can justify broader delegation, while incidents or drift should contract the authority envelope.

Agentic systems complicate responsibility because action emerges from interactions among models, data, tools, policies, developers, process owners, and users. Ethics research has long identified traceability and responsibility as central problems of algorithmic mediation [31], [32], [38]. An organization should therefore avoid assigning accountability to “the AI.” Accountability must remain attached to human and institutional roles. At minimum, four roles should be explicit: the business decision owner, who owns objectives and consequences; the system owner, who owns technical operation; the risk or compliance owner, who defines control requirements; and the data/knowledge owner, who governs evidence sources. Smaller organizations may combine these roles, but the responsibilities should still be distinguishable.

This role design also addresses the automation–augmentation paradox [9]. If management treats AI exclusively as automation, human expertise can atrophy and responsibility may be displaced. If management treats AI exclusively as augmentation, organizations may retain costly manual checkpoints that add little value. ADI instead treats human involvement as a portfolio of decision rights: Some decisions are delegated, some are jointly produced, some require approval, and some remain exclusively human. The portfolio should evolve with evidence about system capability and organizational risk appetite.

Table 3 illustrates that governance should be implemented as a combination of technical controls and human role assignments. The same underlying AI model may therefore operate with different autonomy in different processes. This context sensitivity is preferable to blanket claims that a model is “safe” or “unsafe.”

The framework generates six propositions intended for empirical testing. They are stated at a level that can support survey research, field experiments, process-mining studies, design-science evaluations, or longitudinal case research.

Proposition 1. The positive effect of agentic AI capability on decision-cycle performance is stronger when organizational knowledge is provenance-rich and operationally retrievable than when knowledge is fragmented or weakly governed.

Proposition 2. Conditional autonomy based on consequence severity, uncertainty, reversibility, and evidentiary strength produces higher joint decision quality and lower incident severity than either uniform human approval or uniform machine autonomy.

Proposition 3. Provenance visibility improves calibrated human reliance on agentic recommendations by enabling users to distinguish authoritative evidence, model inference, and uncertainty.

Proposition 4. Permission controls enforced outside the generative model reduce the relationship between model error and organizational harm more effectively than instruction-only constraints embedded in prompts. Proposition 5. Exception-oriented human oversight produces higher supervisory effectiveness than universal approval when routine decision volume is high and escalation criteria are well specified.

Proposition 6. Continuous assurance that can contract or expand an agent’s authority based on observed performance increases long-run organizational value relative to static pre-deployment approval.

Proposition 1 connects analytics-capability research [3–5] to agentic architectures by arguing that knowledge governance becomes a direct antecedent of action quality. Propositions 2 and 5 extend classic automation and trust theory [26–28] into dynamic decision-rights allocation. Proposition 3 connects explainability to reliance calibration rather than treating explanation as an intrinsic good [29], [30]. Proposition 4 differentiates policy statements from enforceable infrastructure, an increasingly important distinction when language models can plan around loosely phrased constraints. Proposition 6 treats autonomy as an adaptive organizational resource whose scope should depend on evidence accumulated through use.

Organizations should not implement ADI by moving directly from a chatbot pilot to broad autonomous execution. A staged roadmap reduces both technical and institutional risk. Stage 1 is assistive analytics: the system retrieves information, summarizes evidence, and drafts options, while humans retain all decision rights. Stage 2 is supervised agency: The system can construct plans and prepare tool calls, but each consequential action requires explicit approval. Stage 3 is bounded autonomy: The system executes low-risk, reversible actions within defined limits and escalates exceptions. Stage 4 is adaptive autonomy: The authority envelope is adjusted through continuous assurance, performance evidence, and formal governance review.

The transition between stages should be evidence-based. Useful readiness criteria include retrieval provenance coverage, test-set performance on representative cases, exception-handling accuracy, frequency and causes of human overrides, rollback success, incident severity, policy compliance, and user understanding of system limitations. Productivity alone is not a sufficient promotion criterion. A system that saves time while increasing tail risk or degrading accountability should not receive broader decision rights.

Implementation also requires process redesign. Organizations should identify decision episodes rather than merely automate tasks. A decision episode has an owner, trigger, objective, information requirements, permissible actions, consequences, and escalation path. Mapping these elements often reveals that existing processes rely on tacit judgment or undocumented exceptions. ADI deployment can therefore expose governance debt that predates AI. The appropriate response is not to encode ambiguity into prompts but to clarify policies, decision rights, and exception ownership.

Agentic systems require a broader performance model than predictive analytics. Four metric families are proposed. Decision quality includes accuracy, economic value, service level, or domain-specific outcomes. Process efficiency includes cycle time, cost, throughput, and coordination burden. Governance quality includes provenance coverage, policy compliance, escalation appropriateness, unauthorized-action prevention, and audit completeness. Human-system quality includes calibrated reliance, override quality, perceived control, workload, learning, and the preservation of relevant expertise. Evaluating all four families avoids the narrow optimization of speed at the expense of accountability or resilience.

A particularly important metric is recoverability. Because no complex system will be error-free, organizations should measure how quickly an erroneous action is detected, contained, reversed, and learned from. Reversibility is not only a decision-rights factor but also an engineering target. Workflows can be deliberately designed with staged commits, sandbox environments, transaction caps, and compensating actions so that AI errors remain locally containable. This design-for-recovery perspective is likely to become a defining capability of mature agentic organizations.

For management theory, ADI suggests that the unit of analysis should shift from the “AI tool” toward the human–AI decision system. Performance emerges from the configuration of data, models, knowledge, authority, users, workflows, and controls. This perspective aligns with research showing that AI management involves interdependent choices around autonomy, learning, and inscrutability [10]. It also offers a bridge between business analytics research, which emphasizes capabilities and value creation, and organizational AI research, which emphasizes control, expertise, and institutional consequences.

For practice, the framework provides a concrete design question: what is the smallest safe authority that allows the system to create meaningful value? Starting with bounded permissions and expanding them through evidence is more defensible than granting broad access and attempting to monitor after the fact. This principle is compatible with contemporary governance frameworks and standards [33–37], yet it is expressed in operational terms that can be implemented by process owners, AI teams, security teams, and managers.