A decision-level governance toolkit: delegating discrete rights to AI changes who actually controls choices and who must be held responsible, and firms should tighten human approval, escalation and monitoring as AI autonomy rises or reliability falls. The seven-rights/five-responsibilities framework makes hidden responsibility gaps observable and yields testable propositions for managing human–AI decision collaboration.

Main Finding

Sun (2026) develops a decision-level, theory-synthesis framework that explains how specific transfers of decision rights from humans to AI change effective control and evidence access within enterprise decisions, and how responsibility should follow (or fail to follow) those transfers. The paper derives seven separable decision rights and five responsibility domains, centers a rights–control–responsibility alignment mechanism, and shows how decision exposure, AI autonomy and AI reliability jointly determine governance configurations and failure modes. The framework yields eight empirically testable propositions and an auditable method for allocating rights, responsibilities, evidence and lifecycle controls.

Key Points

• Core conceptual contribution
  • Rights–control–responsibility alignment: governance succeeds only when the actor formally responsible also has the competence, authority and information to prevent, detect, contest or remedy failures created by delegated rights.
• Derived decision-rights (seven)
  • Information access
  • Recommendation
  • Selection (choose among alternatives)
  • Approval (permission to enact a chosen option)
  • Veto / override
  • Execution (implementing the decision)
  • Escalation (route to higher-level review or contest)
• Responsibility domains (five)
  • System design (problem framing, data, model validation)
  • Decision process (day‑to‑day use, case handling)
  • Outcome stewardship (cumulative effects, externalities)
  • Oversight (governance, monitoring, assurance)
  • Remediation (correction, redress, learning)
• Contingencies and mechanisms
  • AI reliability is treated as an antecedent and dynamic contingency: lower or uncertain reliability narrows permissible selection/execution rights and increases approval/escalation requirements.
  • Decision exposure (ethical salience/stakeholder reach) and AI autonomy determine four governance archetypes (low/high exposure × low/high autonomy).
  • Delegating a right changes who controls the decision and who has access to the evidence necessary to intervene; misalignments produce symbolic oversight, delayed intervention, or responsibility gaps.
• Empirical output and managerial utility
  • Eight testable propositions covering how reliability, exposure, vendor opacity, traceability, explanation design, and contestability affect rights allocation, performance and learning.
  • Two worked examples (low- and high-exposure decisions) illustrate how the framework yields more precise governance prescriptions than generic “human-in-the-loop” requirements.
  • The paper provides an auditable, lifecycle-aware design logic for enterprises allocating rights, evidence access and remediation responsibility.
• Methodological framing
  • Distinguishes normative design guidance (how to configure governance for auditable control) from empirically testable propositions (what firms will do or what outcomes follow).

Data & Methods

• Study type: conceptual theory-synthesis and model-building (not primary empirical research).
• Corpus: purposive sample of 44 sources assembled July 2026
  • 40 peer‑reviewed articles/proceedings, 2 standards/regulatory instruments, 1 risk-management framework, 1 conceptual-method source.
  • Four included streams: human–AI collaboration/delegation; organization/work/algorithmic management; calibrated reliance/explainability/meaningful human control; responsible AI governance/auditing/regulation.
• Identification strategy
  • Seed (anchor) works representing key streams, then backward/forward citation chaining and targeted keyword searches (e.g., “AI delegation approval veto execution”, “meaningful human control escalation”, “AI reliability decision rights”).
  • Inclusion criteria: defined authority relations; specified responsibility/accountability/remediation domains; identified contextual conditions changing permissible delegation; or described control mechanisms connecting evidence to intervention.
• Concept extraction and abstraction
  • First-order coding retained original vocabulary (recommend, select, approve, override, implement, monitor, audit, remediate, etc.).
  • Constant comparative method abstracted overlapping terms into distinct governance-relevant categories until two successive search iterations produced no new rights or responsibility domains.
  • Resulting taxonomy: seven rights, five responsibility domains.
• Saturation and validation
  • Saturation assessed at category level; negative-case testing mapped borderline items (e.g., protected non-reliance → veto/override; human reconsideration → escalation/remediation).
  • Propositions linked to theoretical foundations, causal mechanisms, observable outcomes and falsification conditions; an audit trail (Table 2 in the paper) documents sources per proposition.
• Outputs: prescriptive design principles, eight empirically testable propositions, and worked applications.

Implications for AI Economics

• Governance as productive-capital and transaction-cost factor
  • Rights allocation and responsibility alignment are governance capital that affect realized productivity from AI. Models of AI-driven productivity should include governance frictions (misalignment costs, monitoring overhead, audit/compliance expenses) as state variables that modulate technology complementarities with labor.
• Delegation choices change value capture and firm boundaries
  • Choice to grant selection/execution rights to vendors or in-house systems alters bargaining leverage, switching costs, and liability exposure. Economists should model how audit-access clauses, liability rules and evidence availability affect vertical integration and contracting for AI services.
• Reliability as an economic state variable
  • AI reliability (stability, calibration, domain validity) constrains permissible delegation and thereby labor displacement and task reallocation. Incorporate reliability dynamics (drift, validation costs) into models of adoption speed, labor re-skilling needs and capital investment returns.
• Heterogeneous effects by decision exposure
  • Decisions with high stakeholder exposure (e.g., lending, hiring, clinical care) will sustain more retained human rights, higher oversight intensity and slower automation — leading to sectoral heterogeneity in automation rates and productivity gains.
• Incentives, accountability and moral hazard
  • Misaligned rights/responsibility create symbolic oversight and moral hazard (actors exercise control without remediation incentives). Empirical work should measure how misalignment correlates with error rates, regulatory sanctions, or remediation costs.
• Vendor opacity, audit access and market structure
  • Opacity raises governance costs and reduces permissible operative rights; markets may favor vendors that grant richer audit access or standardized logs. This creates competitive advantage for transparent providers and may spur demand for certified assurance services — implications for industry structure and complementary service markets.
• Measurement and empirical research directions
  • Useables: internal decision logs, AI model logs, approval/escalation timestamps, incident and remediation reports, procurement contracts, audit records, role descriptions, and regulatory filings.
  • Testable outcomes suggested by the paper: scope of delegated rights, frequency of symbolic approvals, speed/accuracy of recalibration after drift, incidence of responsibility gaps, joint decision quality relative to single best performer.
  • Econometric designs: firm-level panel studies linking governance configurations to productivity, event studies around governance interventions/regulation, matched comparisons of decisions with differing exposure/autonomy.
• Policy and market implications
  • Regulation that mandates human oversight without specifying rights/evidence access risks symbolic control; policy should target alignment (who can stop, who has evidence, who must remediate).
  • Insurance pricing, compliance costs and capital allocation decisions should reflect governance maturity (traceability, audit completeness) as risk-reducing features.
• Managerial and operational economics
  • For practitioners, the framework gives a rubric to decide when to invest in monitoring/audit vs when to expand AI operative rights; it prescribes linking authority to evidence access and remediation budgets, which economists can monetize in cost–benefit analyses.

Suggested ways economists can operationalize the framework - Treat rights allocation as an observable categorical variable (which of the seven rights are held by AI vs human) and estimate its effect on outcomes (error rates, speed, cost). - Model AI reliability as an endogenous evolving process influenced by monitoring investments; estimate returns to monitoring (traceability, audits) via improvements in permissible delegation and productivity. - Study market responses to vendor transparency (audit access) by tracking procurement choices, switching behavior and contract terms.

If you want, I can: - Extract concrete variables and measurement strategies (survey items, event-log metrics) to operationalize the seven rights and five responsibility domains for empirical work. - Draft a simple empirical design (data sources, identification strategies) to test one of the eight propositions (e.g., Proposition 1 on reliability and scope of delegation).