A five‑domain enterprise framework and a 'DNA‑Coefficient' to measure gaps between formal and practiced decision rights that promises safer, auditable, and more cost‑efficient deployment of agentic AI across organizations.

Main Finding

The paper proposes the Authority Resolution Framework (ARF): authority is a relation that must be explicitly resolved across five loosely-coupled enterprise ontology domains (Social, Business vocabulary, Codified processes, Machine/executable, and External real-world). Treating authority as a cross-domain primitive (the Authority Relation, AR) and measuring divergence between documented and practiced authority (the DNA‑Coefficient) is necessary to safely and economically scale agentic AI. Proper cross-domain resolution yields (a) safer, auditable delegation to non-human agents and (b) measurable economic benefits for large-language-model (LLM)–driven systems through context compression (fewer tokens, lower inference cost, higher factual accuracy) and robust state-change propagation.

Key Points

• Problem framed as a category-level allocation and governance failure: agentic-AI spend is rising rapidly, governance and measurable returns lag, and authority inconsistencies across domains are a core structural cause.
• Five-domain ontology: Social (roles/influence), Business (domain vocabulary), Process (codified procedures), Machine (code/permissions), Real-world (external context).
• Authority Relation (AR): a six-element tuple binding actor, action, object, bounded context, justification chain, and DNA‑Coefficient (measures divergence between documented and practiced authority).
• DNA‑Coefficient: a novel, directional, per-instance score intended to quantify how much an organization’s lived decision-making diverges from its documented rules.
• Economic argument: resolved ARs let systems retrieve small, precise structured facts instead of large unstructured contexts—this “context compression” reduces token consumption, inference cost, and LLM factual degradation with long contexts.
• Safety extension: the AR primitive applies unchanged to embodied/physical AI (robotics, AVs, industrial automation) where unresolved authority has direct physical-safety consequences.
• Implementation-ready: provides JSON‑LD schema sketch, W3C PROV-based causal-graph treatment of Justification-Chain, worked knowledge-graph query patterns, and integration paths with Model Context Protocol and agentic-memory architectures.
• Governance mechanisms: proposes a 3+1 empirical methodology for estimating DNA (decision-log divergence analysis; organizational network analysis; structured elicitation; plus an audit instrument for agent-design authority), a tri-functional affirmation requirement (business, technical, governance), and a 6-level Authority Maturity Model.
• Limits: DNA‑Coefficient methodology is proposed, not yet empirically validated; some related literature remains paywalled/unobtained.

Data & Methods

• Methods are primarily conceptual/formal and implementation-sketch-based:
  • Formalization: defines Authority Relation (six elements) and shows cross-domain object-resolution implications for semantic interoperability and state-change propagation.
  • Implementation artifacts: JSON‑LD schema; knowledge-graph query examples; W3C PROV for provenance/justification chains.
  • Integration mapping: shows interoperability with Model Context Protocol and existing agentic-memory designs; no bespoke runtime required.
• Empirical measurement proposal (3+1 instruments) to estimate DNA‑Coefficient:
• Decision-log divergence analysis — compare logged actions and documented authorities.
• Organizational network analysis — map informal influence vs formal roles.
• Structured elicitation — targeted interviews/surveys to surface tacit practices.
• Audit of agent-design authority — inspect who controls objectives and permission boundaries of agents.
• Validation status: methodology and pilot are proposed; empirical pilot not yet run. Positioning is cross-checked against several existing organizational-ontology works; the framework is positioned to extend and stitch together prior approaches rather than replace them.

Implications for AI Economics

• Direct operating-cost reduction for LLM-based agents:
  • Context compression reduces tokens fed to models; since inference cost is approximately linear in token usage, reducing tokens yields proportional inference-cost savings and reduced latency.
  • Reduced need for repeated re-derivation of business meaning improves accuracy and cuts downstream error-correction costs.
• Improved ROI and lower project attrition:
  • By making authority machine-resolvable and auditable, enterprises can more reliably delegate to agents, improving throughput and lowering cancellation risk (Gartner predicts many agentic projects will be cancelled without better governance).
  • Tri-functional co-ownership (business/technical/governance) helps align incentives and unlock measurable productivity gains that currently evade many firms.
• Risk mitigation and regulatory value:
  • Provides auditable justification chains (W3C PROV) and traceability aligning with emerging regulation (e.g., EU AI Act), likely reducing expected compliance costs, fines, and insurance premiums.
  • For physical AI, reducing authority divergence translates directly to reduced accident/recall risk and associated liability costs.
• Scalable semantic-event propagation reduces integration overhead:
  • Cross-domain object-resolution reduces brittle point-to-point integrations, lowering maintenance and integration labor costs as agentic systems scale.
• Measurement and investment decisions:
  • The DNA‑Coefficient can become a quantifiable governance KPI: expected benefits include token-cost savings, fewer manual interventions, lower compliance remediation cost, and faster time-to-value for agentic deployments.
  • Economists and decision-makers can run pilots to estimate elasticities: e.g., measure token consumption and error rates before/after ARF-driven grounding to quantify per-agent cost reduction and extrapolate portfolio-level savings.
• Recommended short-term actions for economic decision-makers:
  • Pilot ARF on a high-value agentic use case (measure token usage, error corrections, time-to-resolution, compliance incidents) to empirically estimate cost/benefit.
  • Require tri-functional affirmation for agent delegation decisions to internalize governance externalities.
  • Integrate ARF artifacts (JSON‑LD AR records, provenance trails) into model context and retrieval pipelines to realize immediate token and accuracy benefits.

Summary takeaways: ARF reframes the scaling failure of agentic AI as a cross-domain authority-resolution problem with concrete, quantifiable economic payoffs (reduced LLM inference cost, improved accuracy, faster value capture, and lower regulatory and safety risk). The framework is implementation-ready and testable; empirical validation of the DNA‑Coefficient and pilots to quantify economic impact are the next practical steps.