Dr. Smrite Goudhaman smritegoudhaman.com All Research
Conceptual Framework · MERC 2026 · Paper ID MERC2026-0093 IT

Human Oversight Frameworks for Agentic AI in Enterprise Decision Systems

Designing Resilient Governance for Poly-Crisis Environments — a five-pillar framework that embeds human oversight into agentic AI as a structural feature, not an afterthought.

Authors
Dr Smrite Goudhaman · Tarushee Dahiya · Satvik Kumar
Institution
Golden Gate University, USA · Doctoral Programme
Conference
Management Education & Research Conference (MERC 2026) · IIM Kashipur · 29–31 May 2026
Paper ID
MERC2026-0093 IT
MERC 2026 · IIM Kashipur CERE 2026 · IIM Indore AAAI-26 · Singapore AITC 2026 · Nuremberg IASEAI · UNESCO Paris UN Global AI Governance · Geneva
P1
Oversight Layer
Audit · Human Gates · Dashboards
P2
Explainability Layer
Rationale · Audit Trails · Transparency
P3
Escalation Layer
Thresholds · Tiered Trees · Override
P4
Ethical Risk Layer
Bias Audits · Fairness · Privacy
P5
Adaptive Learning
Post-Crisis Reviews · Retraining · Policy

The Governance Gap No Existing Framework Was Built to Close

Agentic AI systems now execute multi-step, autonomous decisions across enterprise workflows — planning, executing, evaluating, and re-planning continuously, without human intervention at each step. This is a fundamentally different kind of AI from the single-output models that existing governance frameworks were designed for.

Microsoft AutoGen, agent swarms, and LLM-based orchestrators are already in active enterprise deployment. The governance architecture has not kept pace. And the consequences of that gap are not hypothetical.

When agentic AI eliminates the pause between signal recognition and action, organisational sensemaking collapses — precisely when human judgment matters most. HOF-AIDE was designed to restore that pause structurally.

"Sensemaking collapses when agentic AI eliminates the pause between signal recognition and action — exactly when human judgment matters most."

— Weick (1995), applied in HOF-AIDE theoretical framework · Goudhaman, Dahiya & Kumar (2026)

Continuous Autonomous Action Without Human Checkpoints

Unlike predictive AI, agentic systems plan, execute, evaluate and re-plan without pausing for human review. Existing frameworks treat oversight as post-hoc validation — not embedded design (Raji et al., 2020).

Frameworks Built for Single-Output Models

Current AI governance frameworks — including the EU AI Act and U.S. Executive Order — were designed for single-decision AI systems. They fail to address continuous autonomous action chains (Weidinger et al., 2021).

Accountability Gaps in Multi-Agent Ecosystems

When five agents are coordinating a customer operations decision in real-time, who is accountable for the outcome? HOF-AIDE answers this architecturally, not just procedurally.

Poly-Crisis as the New Operational Normal

Regulatory instability, operational disruption, reputational risk, cybersecurity threats and workforce anxiety now converge simultaneously — each amplifying the others (Tooze, 2022). Governance must be designed for this volatility, not for stable conditions.

Defining the Conceptual Boundaries of HOF-AIDE

How can enterprises design governance architectures that maintain effective human oversight of agentic AI decision systems operating in poly-crisis environments?

O1

Identify Governance Gaps

Diagnose structural inadequacies of existing frameworks using STS theory, HRO research, and responsible AI scholarship.

O2

Develop HOF-AIDE

Construct a five-pillar governance framework that embeds oversight into agentic AI as a design feature, not an afterthought.

O3

Derive Propositions

Generate four empirically testable propositions linking HOF-AIDE governance layers to organisational resilience outcomes.

O4

Demonstrate Applicability

Apply HOF-AIDE to a BPO Customer Operations Control Tower — illustrating real-world governance logic and crisis response capability.

Five Bodies of Scholarship. One Integrated Framework.

HOF-AIDE is built through conceptual framework development (Whetten, 1989) — a multi-theoretical synthesis approach that draws on five distinct scholarly traditions, each contributing a specific structural layer to the framework.

Socio-Technical Systems (STS)
Trist & Bamforth (1951) · Baxter & Sommerville (2011)

Organisations must jointly optimise technical and social subsystems. HOF-AIDE embeds both AI tools and human governance roles — rejecting the assumption that technical architecture alone can govern outcomes.

High Reliability Organisations (HRO)
Weick & Sutcliffe (2007)

HROs thrive in complexity through operational sensitivity and commitment to resilience. These principles are embedded directly in HOF-AIDE's Oversight and Escalation Layers — the two pillars that hold under crisis pressure.

Dynamic Capabilities
Teece, Pisano & Shuen (1997)

Organisational resilience requires adaptive sensing and reconfiguration. HOF-AIDE's Adaptive Learning Layer directly operationalises dynamic capability theory — governance that evolves as threats evolve.

Sensemaking Theory
Weick (1995)

Agentic AI disrupts organisational sensemaking by eliminating the natural pause where human interpretation occurs. HOF-AIDE's Explainability Layer provides cognitive cue structures to restore stakeholder understanding after AI-mediated disruptions.

Responsible AI Governance
Jobin, Ienca & Vayena (2019) · Raji et al. (2020) · Floridi et al. (2018)

Ethics principles must be institutionalised into auditable practices — not declared as values and left unimplemented. HOF-AIDE's Ethical Risk Layer operationalises this directly through bias audits, fairness checks, and privacy assessments.

The Integration Principle
Joint Optimisation · Trist & Bamforth (1951)

All five theories converge on one finding: governance cannot be selective. Partial implementation creates asymmetrical structural gaps that increase vulnerability under poly-crisis conditions. HOF-AIDE must be adopted as a complete architecture, not a menu of options.

Five-Pillar Human Oversight Framework for Agentic Intelligence Decision Ecosystems

Each pillar addresses a different dimension of the agentic AI governance problem. All five are connected by recursive feedback flows through the Human Governance Council — the enterprise AI accountability layer with crisis override authority.

HOF-AIDE Framework

Human Oversight Framework for Agentic Intelligence Decision Ecosystems · Goudhaman, Dahiya & Kumar (2026)

Poly-Crisis Environment — Regulatory Instability · Operational Disruption · Reputational Risk · Cybersecurity Threats · Workforce Anxiety
P1
Oversight Layer
Audit Logs
Human Gates
Dashboards
CAIO
P2
Explainability Layer
Decision Rationale
Audit Trails
Transparency Reports
CIO / Ethics Officer
P3
Escalation Layer
Threshold Triggers
Tiered Trees
Crisis Override
COO / Risk Head
P4
Ethical Risk Layer
Bias Audits
Fairness Checks
Privacy Assessments
CRO / Board AI
P5
Adaptive Learning
Post-Crisis Reviews
Retraining Governance
Policy Updates
AI Gov Council
↕ Recursive Bi-Directional Feedback Flows Between All Pillars and HGC ↕
Human Governance Council (HGC)
Enterprise AI Accountability Layer · Crisis Override Authority · All five pillars report to and receive direction from HGC

Bi-directional arrows (↕) indicate recursive feedback flows between pillars and the HGC. All five layers interact dynamically, not sequentially. Partial implementation creates structural vulnerability gaps.

Four Empirically Testable Propositions Derived from HOF-AIDE

These propositions are the framework's empirical contribution — each one is testable in matched-firm studies and provides a specific mechanism linking HOF-AIDE governance layers to measurable organisational outcomes.

P1
Explainability & Trust Resilience

Formal explainability governance positively moderates institutional trust resilience after agentic AI decision-events

Decision rationale records, immutable audit logs, and transparency reports restore stakeholder trust after AI-mediated disruptions — providing the cognitive cue structures that sensemaking requires.

Basis: Sensemaking Theory · Weick (1995)
P2
Escalation Architecture & Failure Propagation

Structured escalation architecture significantly reduces the risk of governance failure cascading through agentic AI ecosystems

Unstructured escalation processes allow failures to propagate through tightly coupled enterprise AI networks. HOF-AIDE inserts human judgment at the high-coupling nodes where cascade risk is greatest.

Basis: Normal Accident Theory · Perrow (1984)
P3
Adaptive Learning & Resilience Capacity

Organisations with adaptive learning governance demonstrate greater resilience capacity in poly-crisis conditions than those with static governance

Static governance frameworks become misaligned as AI capabilities evolve and regulatory requirements shift. HOF-AIDE's Adaptive Learning Layer builds reconfiguration into the governance architecture itself.

Basis: Dynamic Capabilities · Teece, Pisano & Shuen (1997)
P4
HOF-AIDE Integration & Governance Performance

Full five-layer HOF-AIDE implementation produces superior poly-crisis governance outcomes versus partial implementation

Asymmetrical structural gaps — where one or two layers are implemented and others are not — increase vulnerability disproportionately. Selective adoption of governance principles defeats the framework's purpose.

Basis: Joint Optimisation · Trist & Bamforth (1951)

HOF-AIDE Applied: AI-Powered Customer Operations Control Tower

A global BPO enterprise handling 45,000 daily customer interactions across financial services, telecom, and healthcare — running a five-agent AI ecosystem. HOF-AIDE applied across all five governance layers.

Global BPO · 45,000 Daily Customer Interactions

Financial Services · Telecom · Healthcare · Five-Agent AI Ecosystem

Triage Agent
Sentiment Escalation
SLA Breach Prediction
Workforce Routing
Governance Exception
↓ HOF-AIDE Governance Layers Applied to All Five Agents ↓
P1 · Oversight Layer
Real-time dashboards visible to Human Governance Council; HGC embedded in live operations; human confirmation required on all complex or ambiguous cases before AI action proceeds.
P2 · Explainability Layer
Structured audit trails for every agentic decision; decision rationale logs for sentiment escalation events; all records accessible to regulators and the HGC on demand.
P3 · Escalation Layer
SLA threshold triggers route automatically to human supervisors; Level 2 sentiment alerts handled exclusively by human agents; no AI-only resolution on high-stakes customer interactions.
P4 · Ethical Risk Layer
Workforce routing algorithms audited continuously for demographic bias; fairness checks on task assignment; privacy assessments run before any customer data inference is acted upon.
P5 · Adaptive Learning
Post-crisis model retraining completed within 3 months of any governance incident; governance policy updates triggered automatically by incident patterns; HGC approves all retraining before deployment.

Strategic Guidance for Enterprise Leaders and AI Governance Architects

CEO

Reframe Governance as Competitive Differentiator

HOF-AIDE embeds responsible governance as a trust asset — critical in markets where AI missteps rapidly erode customer loyalty. Governance is not a cost. It is a market position (Edelman Trust Barometer, 2024).

CIO

Build Governance Into the Architecture from Day One

Audit trails, explainability dashboards, and anomaly detection must be platform requirements, not retrofits. Governance that is built after deployment will always be insufficient.

CHRO

Counter Workforce Displacement Anxiety Structurally

HOF-AIDE's escalation and oversight roles require human judgment at every significant decision node. This confirms, architecturally, that AI augments rather than replaces human expertise.

COO / CX

Deploy Agentic AI in High-Stakes Environments Confidently

Defined human checkpoints protect both customer experience quality and organisational reputational standing. The escalation architecture is designed precisely for the moment when operational pressure is highest.

Regulators

HOF-AIDE Provides Auditable Proof of Meaningful Oversight

The Explainability and Ethical Risk Layers directly address EU AI Act (2024) and U.S. Executive Order 14110 (2023) compliance requirements — providing verifiable evidence of human oversight in high-risk AI deployments.

Poly-Crisis Preparedness

Design Governance for Volatility as the Norm

HOF-AIDE's Adaptive Learning Layer ensures governance evolves with regulatory shifts, cyber threats, and operational disruptions. Frameworks designed for stable conditions fail precisely when they are most needed (Tooze, 2022).

What the Framework Has Not Yet Done — and What Must Come Next

  • L1

    No Empirical ValidationHOF-AIDE is a conceptual framework. The links between governance architecture and organisational outcomes are theoretically grounded but require empirical testing in real enterprise environments.

  • L2

    Practitioner-Informed, Not Case StudyThe BPO use case is simulated from practitioner experience — not an independently documented case study. It illustrates logic, not evidence.

  • L3

    BPO/CX Context SpecificityThe framework was developed for high-volume, SLA-governed customer operations. Transferability to manufacturing, healthcare, and public sector is unconfirmed and requires separate validation.

  • L4

    Governance Infrastructure AssumptionsHOF-AIDE presupposes CAIO roles, Ethics Officers, and Board AI Committees — structural prerequisites that are currently beyond the capacity of smaller or governance-immature organisations.

FR1
Empirical Validation

Matched-firm studies comparing governance outcomes between HOF-AIDE adopters and non-adopters — across sectors, scale, and poly-crisis exposure levels.

FR2
Multi-Client BPO Scalability

How to build modular governance architectures that serve diverse client mandates, regulatory environments, and risk profiles simultaneously.

FR3
Regulatory Alignment

Whether HOF-AIDE inherently satisfies EU AI Act, U.S. EO 14110, and emerging Global South frameworks — and where supplementary compliance architecture is required.

FR4
Cultural Dimensions

Which organisational culture characteristics enable or inhibit effective human oversight governance adoption — particularly across different national governance traditions.

FR5
UNESCO AI Ethics Alignment

Cross-national comparison of HOF-AIDE alignment with UNESCO's 2021 AI Ethics Recommendation — testing transferability across jurisdictions and regulatory maturity levels.

"HOF-AIDE is not a constraint on AI capability — it is a blueprint for making AI trustworthy, governable, and resilient."

P1 · Oversight — Audit + Human Gates P2 · Explainability — Rationale + Trails P3 · Escalation — Triggers + Authority P4 · Ethical Risk — Audits + Fairness P5 · Adaptive Learning — Reviews + Updates

Human oversight is not a limitation of agentic AI.
It is the condition that makes enterprise intelligence resilient.

Goudhaman, Dahiya & Kumar (2026) · MERC 2026 · IIM Kashipur · Paper ID MERC2026-0093 IT

How to Cite This Work

Primary Citation

Goudhaman, S., Dahiya, T., & Kumar, S. (2026). Human Oversight Frameworks for Agentic AI in Enterprise Decision Systems: Designing Resilient Governance for Poly-Crisis Environments (HOF-AIDE). Proceedings of the Management Education and Research Conference (MERC 2026). IIM Kashipur, 29–31 May 2026. Paper ID MERC2026-0093 IT.

Arrieta et al. (2020). Explainable AI (XAI). Information Fusion, 58, 82–115.

Baxter & Sommerville (2011). Socio-technical systems. Interacting with Computers, 23(1), 4–17.

Dafoe (2018). AI governance: A research agenda. Future of Humanity Institute, Oxford.

Edelman (2024). Trust Barometer: AI and the future of trust.

European Parliament (2024). EU AI Act — Regulation (EU) 2024/1689.

Floridi et al. (2018). AI4People — An ethical framework. Minds and Machines, 28(4), 689–707.

Hemmer et al. (2025). Human-AI collaboration. European Journal of Information Systems.

Jobin, Ienca & Vayena (2019). Global landscape of AI ethics guidelines. Nature Machine Intelligence, 1(9), 389–399.

Li et al. (2023). CAMEL: Communicative agents for LLM society. NeurIPS 36.

Microsoft (2024). AutoGen: Enabling next-generation LLM applications. Microsoft Research.

Perrow (1984). Normal Accidents: Living with High-Risk Technologies. Basic Books.

Raji et al. (2020). Closing the AI accountability gap. ACM FAccT, 33–44.

Teece, Pisano & Shuen (1997). Dynamic capabilities. Strategic Management Journal, 18(7), 509–533.

Tooze (2022, Oct 28). Welcome to the world of the polycrisis. Financial Times.

Trist & Bamforth (1951). Social consequences of the Longwall method. Human Relations, 4(1), 3–38.

UNESCO (2021). Recommendation on the Ethics of Artificial Intelligence.

U.S. Executive Office (2023). Executive Order 14110: Safe & Trustworthy AI.

Wang et al. (2024). Survey on LLM-based autonomous agents. Frontiers of Computer Science, 18(6).

Weick (1995). Sensemaking in Organizations. SAGE Publications.

Weick & Sutcliffe (2007). Managing the Unexpected (2nd ed.). Jossey-Bass.

Weidinger et al. (2021). Ethical and social risks of harm from language models. arXiv:2112.04359.

Whetten (1989). What constitutes a theoretical contribution? AMR, 14(4), 490–495.

Xi et al. (2023). The rise of LLM-based agents: A survey. arXiv:2309.07864.

Zeitlin & Herrigel (2023). Poly-crisis and institutional resilience. Socio-Economic Review, 21(2), 735–757.

Request Full Paper See Research Page → ← Back to Website