PaperlayerView exact revision on arXiv

Source-linked AI summary

Corporate Language Model (CLM): Transforming Tacit and Fragmented Enterprise Knowledge into a Sovereign, Auditable, and Executable Corporate Intelligence Layer

Fabricio C. Avini, Guilherme Trez

arXiv:2609.04377v1cs.AI

TL;DR

Enterprise AI lacks access to the tacit, firm-specific knowledge needed to reflect how organizations decide and execute. The paper proposes CLM as an ontology-grounded, governed corporate foundation; in the reported deployment, the first three of six maturity stages were in production, the fourth in test, and the final two in design.

  • Problem

    Enterprise AI lacks firm-specific ontological priors and access to tacit operational knowledge, which remains fragmented and undocumented.

    Source: §1 · p. 2, §2 · p. 3

  • Method

    CLM encodes firm knowledge, reasoning patterns, governance constraints, and execution contracts in a composable, auditable foundation with ontology-grounded corporate memory and traceable governed actions.

    Source: §1 · p. 2, §4.1 · p. 8, §4.4 · p. 9

  • Results

    Stages 1 through 3 were in production, Stage 4 was in test, and Stages 5 and 6 remained in design in the reported rollout.

    Source: §7.2 · p. 17

  • Takeaways & Limitations

    The paper concludes that addressing firm-specific tacit reasoning, exceptions, governance, and execution contracts requires a structured foundation rather than selecting a better model.

    Source: §10 · p. 20

    Evidence comes from a single April 2026 snapshot, with no public-baseline comparison or formal ablation, and does not establish superiority, causality, or cross-organizational generalizability.

    Source: §7.5 · p. 18

Abstract

from arXiv · show

Enterprise AI deployments fail not from model inadequacy, but because organizations lack a structured substrate encoding how they decide, negotiate, and execute. Generic LLMs carry no firm-specific ontological priors; RAG remains brittle, with no path to executable action; static playbooks encode logic but cannot reason or adapt. This demands an architecture treating tacit-knowledge capture, ontological grounding, sovereign deployment, and auditable actuation as co-designed from the start. This paper introduces the Corporate Language Model (CLM), a framework transforming a firm's structured, unstructured, multimodal, and tacit knowledge into an ontology-grounded enterprise foundation upon which reasoning and governed execution are composed. CLM has five capability planes and four architectural pillars: a Neurosymbolic Mesh coupling generative models with a knowledge graph; a Skill Graph where reusable tactics, personas, objections, and goals are typed and composed; Living Digital Twins modeling functional areas as reasoning surrogates; and a Deep Security Layer enforcing sovereignty, traceability, and human oversight. A Spec-as-Code paradigm bridges grounded intent and executable artifact. CLM is one instantiation of this foundation-centric class. Four contributions follow: CLM is defined as a distinct object of study; the Skill Graph is introduced for compositional explainability by construction; the Wisdom Listener effect is proposed, whereby tacit-capable foundations compound in value with use, connecting to dynamic capabilities and organizational learning; and evidence from a JCI-accredited tertiary hospital in Brazil instantiates three of the six maturity stages under LGPD.

1 Introduction

The paper frames enterprise AI failure as an epistemic gap between generic models and undocumented firm-specific knowledge, and proposes CLM as a foundation rather than a model or pipeline. It presents CLM as a position-paper architecture with four contributions and distinguishes the system from the knowledge substrate it constitutes.

1 Introduction

  • The paper identifies the central gap as firm-specific operational knowledge that remains tacit and fragmented, beyond the priors of generic models and the reach of existing approaches.It argues that these constraints should shape architecture from the outset.Source: §1 · p. 2
  • CLM is a composable, auditable foundation encoding a firm’s knowledge, reasoning patterns, governance constraints, and execution contracts, rather than a replacement for a generic LLM.The paper describes CLM as one instantiation of a broader foundation-centric class.Source: §1 · p. 2
  • The paper defines CLM through five capability planes and four architectural pillars, introduces the Skill Graph and Wisdom Listener effect, and reports early operational evidence from regulated healthcare.The evidence instantiates three of six cumulative maturity stages under real sovereignty and auditability constraints; empirical validation remains future work.Source: §1 · p. 2
  • CLM denotes Corporate Language Model, while Enterprise Knowledge Foundation names the accumulated, ontologically grounded, queryable substrate of a firm’s explicit and tacit knowledge.The paper treats CLM as the architectural system and the enterprise knowledge foundation as the substrate it constitutes.Source: §1 · p. 2

2 Problem Statement

The paper treats five interlocking enterprise AI failure modes as architectural constraints: undocumented tacit knowledge, absent firm-specific priors, unauditable recommendations, sovereignty requirements, and the adaptability-accountability trade-off. Together, they motivate a foundation-centric design rather than features added later.

2 Problem Statement

  • Tacit organizational knowledge is structurally undocumented, while generic LLMs lack firm-specific ontological priors and cannot internalize organizational reasoning across sessions, domains, and edge cases.Prompt engineering helps only at the margins, and fine-tuning without ontological grounding inherits corpus biases and gaps [7].Source: §2 · p. 3
  • Recommendations without traceable logical paths can be unusable or reckless in regulated, financial, or customer-facing decisions, making traceability a governance requirement,.The paper characterizes inability to reconstruct an output’s basis as a governance failure that disqualifies a system from regulated contexts.Source: §2 · p. 3
  • Sovereignty constraints can prohibit sending operational data to external inference endpoints; LGPD makes this constraint legally binding in Brazil.The paper says cloud-only inference over firm data is structurally incompatible with such environments and notes that compact local models make sovereign deployment technically viable [5].Source: §2 · p. 3
  • Static playbooks sacrifice adaptability, while agentic systems sacrifice accountability, leaving enterprise AI without an acceptable option at scale [16].The paper treats the five failure modes as non-negotiable constraints, motivating a foundation-centric class of architectures rather than uniquely implying CLM.Source: §2 · p. 4

3 Related Work

The related-work review selects nine research streams for their relevance to the paper’s five design constraints, then positions CLM in relation to organizational knowledge management, retrieval, knowledge graphs, and neurosymbolic AI. Across these streams, CLM builds on prior ideas while locating them within an enterprise knowledge foundation.

3 Related Work

  • The review uses structured searches across four databases and includes streams addressing at least one of the five design constraints, with recent work emphasized in active areas.Coverage of active RAG and neurosymbolic areas is weighted toward publications from 2023 onward.Source: §3 · p. 4

3.1 Organizational Knowledge Management

  • The Foundation Plane operationalizes tacit-to-explicit knowledge conversion by capturing and reasoning over tacit knowledge at scale, connecting CLM to the SECI model and information-systems work on knowledge resistant to formalization.The paper characterizes its foundation and tacit-capture mechanisms as computational operationalizations of externalization and combination.Source: §3.1 · p. 4

3.2 Retrieval-Augmented Generation and Its Corrective Variants

  • CLM incorporates RAG retrieval primitives within a governed mesh but anchors retrieved content to an enterprise knowledge graph that constrains admissible entities, relations, and policies.This ontological constraint distinguishes CLM’s grounding from retrieval mechanisms alone.Source: §3.2 · p. 4

3.3 Knowledge Graphs and Ontological Grounding

  • CLM treats the corporate knowledge graph as the grounding structure for generation, building on graph and ontology work including GraphRAG [1] and ontology-constrained agent reasoning [9].Its relationship to Sandkuhl et al. [26] is complementary: their enterprise-modeling methods can be composed on top of CLM’s organizational knowledge foundation.Source: §3.3 · p. 5

3.4 Neurosymbolic AI

  • CLM couples generative language models with corporate knowledge graphs across an organizational knowledge substrate, extending neurosymbolic integration into an enterprise scope.The paper frames governance and sovereignty as properties of this coupling.Source: §3.4 · p. 5

3.5 Small and Locally Deployable Language Models

The review connects CLM to locally deployable models, governed agents, functional digital twins, and explainable AI. It positions local inference and governed tool access alongside reasoning surrogates and Skill Graph paths that make explanations part of the reasoning process.

3.5 Small and Locally Deployable Language Models

  • Compact, locally deployable models make sovereign corporate deployment technically viable, and CLM composes task-specific Small Language Models with larger frontier models when broader reasoning is needed [5].The paper treats local deployability as a design floor rather than an optional deployment choice.Source: §3.5 · p. 5

3.6 Agentic Systems and Model-Context Protocols

  • CLM preserves legacy systems as sources of truth and exposes them as governed tools, rather than replacing them, using open protocols such as the Model Context Protocol.This approach responds to auditability concerns in LLM-based agents [16] and is presented as relevant to organizational risk tolerance and adoption velocity.Source: §3.6 · p. 5

3.7 Digital Twins as Functional Reasoning Surrogates

  • CLM’s Living Digital Twins model organizational roles and reasoning patterns, shifting digital-twin use from physical-process simulation to organizational-decision simulation grounded in a corporate knowledge graph.The construct inherits the surrogate-reasoning principle from digital-twin work [21] but applies it to functional roles rather than physical entities.Source: §3.7 · pp. 5–6

3.8 Explainability and Governed AI

  • In CLM, each Skill Graph recommendation’s traversal path supplies its explanation, making explainability structural rather than a post-hoc module,.The paper says the path’s nodes and edges are coextensive with the reasoning process, eliminating the need for approximation methods.Source: §3.8 · p. 6

3.9 Domain-Specific and Fine-Tuned Language Models

CLM differs from fine-tuned models through continuous tacit-knowledge capture, compositional reasoning paths, and sovereignty as an architectural floor. Its positioning compares default affordances across adjacent approaches and integrates capabilities that are otherwise assembled separately.

3.9 Domain-Specific and Fine-Tuned Language Models

  • Unlike training-time adaptation, CLM captures tacit knowledge continuously, makes reasoning paths auditable, and treats sovereignty as an architectural floor.Fine-tuned or locally deployed models can still be composed as components within the CLM foundation.Source: §3.9 · p. 6

3.10 Positioning Summary

  • Table 1 compares default architectural affordances across six dimensions, rather than vendor-specific configurations.The dimensions cover tacit capture, ontological grounding, explainability, sovereignty, legacy integration, and compositional reasoning.Source: §3.10 · pp. 6–7, Table 1 · p. 7
  • CLM distinguishes itself from traditional digital twins by modeling organizational reasoning, and codesigns a full-stack foundation rather than assembling separate pipelines.Its capabilities include tacit capture, ontological grounding, hybrid retrieval, neurosymbolic coupling, twin-based reasoning, policy-aware execution, and sovereign deployment.Source: §3.10 · p. 7

4 The CLM Concept: Five Capability Planes

CLM comprises four sequentially dependent functional planes, wrapped by a transversal Deep Security Layer. Shared typed objects make the architecture composable across grounded memory, reasoning, actuation, and governance.

4 The CLM Concept: Five Capability Planes

  • Four functional planes depend sequentially on the output below them, while Deep Security wraps the stack and enforces sovereignty, auditability, and human oversight.This ordering underlies a cumulative rollout: stages cannot be inverted because planes require their predecessors.Source: §4 · p. 7, Figure 1 · p. 8

4.1 Foundation Plane

  • The Foundation Plane turns documents, interactions, records, and multimodal artifacts into continuously updated, ontology-grounded corporate memory.Its queryable output carries provenance, access permissions, and semantic typing, supporting retrieval by meaning, governance, and compositional reasoning.Source: §4.1 · p. 8

4.2 Tactical Plane

  • The Tactical Plane composes typed skills, personas, goals, and Living Digital Twins against current context, anchoring every reasoning step in inspectable objects.This makes its outputs explainable by construction rather than by approximation.Source: §4.2 · p. 8

4.3 Wisdom-Out Plane

  • The Wisdom-Out Plane renders recommendations for the intended persona, routes them with supporting evidence, and invokes governed tools against legacy systems.The underlying recommendation logic remains unchanged across persona-specific renderings.Source: §4.3 · p. 9

4.4 Governance Plane

  • The Governance Plane gives recommendations traceable Skill-to-Goal paths, updates skill weights from observed outcomes, and makes pause, rollback, and human fallback explicit.Audit is structured into each action rather than applied retrospectively.Source: §4.4 · p. 9

4.5 Deep Security Layer

  • Deep Security uses governance, human-approval, and sovereignty rings to gate execution, while a shared typed object model lets all five planes inspect and reason about common objects.The rings address auditability and input sanitation, high-impact approval gates, tenant isolation, safe legacy write-back, and rollback plans.Source: §4.5 · p. 9

5 Architectural Pillars

CLM gives its capability planes operational form through four coupled pillars, combining graph-grounded reasoning with governed execution across their interactions.

5 Architectural Pillars

  • The four pillars are coupled: the knowledge graph supports skill traversal and twin context, while the Deep Security Layer wraps every exposed interaction.Source: §5 · p. 9

5.1 The Neurosymbolic Mesh

  • The Neurosymbolic Mesh couples generative models to a corporate graph that constrains typed outputs and validates candidates before release, unlike retrieval-only RAG.Source: §5.1 · pp. 10–11

5.2 The Skill Graph

  • The Skill Graph types reusable tactics and context nodes, selects compatible skills by context, effectiveness, and risk, and makes recommendations auditable as graph paths.Uncovered personas, stages, or objections expose tactical blind spots.Source: §5.2 · pp. 11–12

5.3 Living Digital Twins

  • Living Digital Twins model functional-area reasoning with typed parameters updated from graph evidence, then hydrate context, simulate consequences, reconcile perspectives, and tailor output.Source: §5.3 · p. 12

5.4 The Deep Security Layer

  • The Deep Security Layer wraps execution with provenance, secret scanning, prompt-injection checks, human approval gates for high-impact actions, sovereignty, and rollback plans.Source: §5.4 · pp. 12–13

6 From Knowledge to Actuation: the Specas-Code Paradigm

Spec-as-Code addresses the gap between grounded recommendations and executable action by compiling business-authored intent into governed technical artifacts. Its derivation chain links acceptance criteria to scenarios, tests, and interface contracts.

6 From Knowledge to Actuation: the Specas-Code Paradigm

  • Spec-as-Code compiles grounded intent into four mutually consistent artifacts: business and technical specifications, behavior scenarios, and interface contracts.The healthcare example derives a discharge-summary technical specification with ontology, policy, and provenance checks; write-back is false.Source: §6 · pp. 13–14
  • Acceptance criteria deterministically generate BDD scenarios that bind tests, which reference interface contracts constraining the technical specification; this makes change-impact analysis tractable.Rollback plans are typed fields in the Business Specification.Source: §6 · p. 14
  • Domain experts author semantically typed requirements, then ontology validation, Skill Graph resolution, and governance checks constrain compilation before artifacts are emitted.The paper places proprietary scoring functions, registry structures, and compilation heuristics outside its scope.Source: §6 · p. 14

7 Early Empirical Evidence

The evidence demonstrates that CLM can be instantiated and deployed under regulatory constraints, but does not validate it through controlled comparison. It covers a single organization and supports feasibility, architectural coherence, and deployment viability.

7 Early Empirical Evidence

  • The evidence supports CLM’s instantiability and deployment viability under regulatory constraints, not its superiority, causal impact, or generalizability.The paper characterizes the evidence as structural, operational, and convergent, rather than a controlled empirical validation.Source: §7 · p. 14, §7.5 · p. 18

7.1 Cumulative Maturity Model

  • The six-stage maturity model treats each stage as valuable but non-skippable, from capture and grounding through governed actuation.Stages 1–3 are prerequisites for later capabilities, including reliable retrieval and twinization before governed actuation.Source: §7.1 · p. 16

7.2 Structural Evidence: Care Engine Platform

  • The Care Engine has 90 CLM-specific test methods, a 232-test repository suite, and a canonical Skill Graph with 41 typed nodes and 100 typed edges.The graph rebuilds deterministically from canonical registries; the Outcome Learning Loop is tested, but no production signals have been persisted.Source: §7.2 · p. 16
  • Decision activation uses declared weights and fixed risk coefficients, while production rollout reached Stages 1–3 before later stages were attempted.The observation is consistent with the cumulative-stage claim but does not validate it; pre-production constraints include keyword-matching false positives and non-persistent twin session state.Source: §7.2 · pp. 16–17

7.3 Operational Evidence: Healthcare Deployment

  • At a JCI-accredited Brazilian teaching hospital, the Clinical Aggregator structures EHR content within the hospital perimeter, and the quality team drove a JCI-aligned discharge-summary specification.Local-only de-identification was a precondition for hospital approval under LGPD; the evidence does not compare outcomes against alternatives.Source: §7.3 · p. 17

7.4 Convergent Research: Doctoral Validation

  • A doctoral project applies a graph-typed knowledge approach to value-based healthcare measurement at the same hospital, but its quantitative results are not reported here.Those results are planned for the doctoral dissertation.Source: §7.4 · pp. 17–18

7.5 Limitations of the Present Evidence

  • The evidence has no public-baseline comparison or formal ablation, covers a single April 2026 snapshot, and cannot establish superiority, causality, or cross-organizational generalizability.The Outcome Learning Loop has not accumulated production signals.Source: §7.5 · p. 18

8 Key Differentiators and Strategic Properties

CLM’s strategic properties center on capturing tacit knowledge, composing recommendations through typed skills, and grounding organizational reasoning in living twins. Sovereignty and auditability are built into the architecture, with explanations arising from graph paths.

Computable Tacit Knowledge

  • The Foundation Plane turns tacit decision patterns into a queryable, typed, ontologically grounded substrate, and usage is proposed to compound that captured knowledge through the Wisdom Listener effect.The paper connects this proposed feedback loop to dynamic capabilities and organizational learning, while distinguishing its typed, governed substrate mechanism.Source: §8 · p. 18

Compositional Reasoning via the Skill Graph

  • CLM composes recommendations from typed Skill Graph connections among tactics, personas, objections, and goals, making the reasoning path reconstructable.The paper frames this as constraint satisfaction over a typed graph rather than prompt engineering over a generative model.Source: §8 · p. 18

Living Twins as Organizational Reasoning Surrogates

  • Living Digital Twins are reasoning surrogates grounded in a corporate knowledge graph and accumulated interactions, rather than chatbots with static role descriptions.The paper proposes that accumulated tacit knowledge makes twins more accurate representations of how a functional area actually decides.Source: §8 · pp. 18–19

Sovereignty, Auditability, and Explainability by Construction

  • Sovereignty and auditability are architectural properties, while Skill Graph paths make explanations coextensive with reasoning and inspectable by domain experts.The paper describes a compounding sequence in which captured knowledge informs skill selection, twin accuracy, output quality, and subsequent human delegation.Source: §8 · p. 19

9 Discussion and Research Agenda

The discussion frames CLM as a conceptual foundation-centric argument whose early evidence demonstrates instantiability, not controlled comparative validation. It sets out empirical and theoretical work needed to test the architecture and its scope.

Discussion and Research Agenda

  • The paper argues that enterprise AI should be studied as a foundation, with sovereignty, auditability, and tacit capture treated as design constraints, while its early evidence is not a controlled comparative study.The proposed agenda includes empirical validation at scale, formal Skill Graph theory, and cross-vertical generalization.Source: §9 · p. 19

Empirical Validation, Formal Theory, and Generalization

  • Validation at scale requires comparisons across baselines and verticals; the Skill Graph’s compositional properties and conditions for generalization remain open research questions.The paper specifically identifies longitudinal measurement of the Wisdom Listener effect and cross-vertical replication as future work.Source: §9 · p. 19

Sovereignty, Organizational Conditions, and Protocol Dependency

  • The agenda also identifies unresolved trade-offs in local-model performance, organizational readiness, and dependence on open agent-to-system protocols.The paper notes that local-model performance gaps are narrowing but not eliminated, and that protocol fragmentation could reintroduce integration complexity.Source: §9 · p. 20

10 Conclusion

The conclusion reframes enterprise AI as a foundation-building problem: CLM structures organizational knowledge for corporate executable intelligence. It presents CLM as a conceptual case and identifies empirical and formal work still needed.

10 Conclusion

  • Enterprise AI is framed as a foundation problem: selecting a better model does not supply the organizational knowledge needed for decisions, negotiation, and execution.The proposed foundation is the structured layer around, below, or orthogonal to a generic LLM.Source: §10 · p. 20
  • CLM makes organizational knowledge infrastructure through five capability planes, four architectural pillars, and a transversal Deep Security Layer.The paper positions this foundation as a way to turn general linguistic competence into corporate executable intelligence.Source: §10 · p. 20
  • The proposed strategic implication is compounding value: use can deepen captured tacit knowledge, system accuracy, and audit trails across deployment cycles.The conclusion presents this as a conceptual case, with longitudinal measurement of the Wisdom Listener effect left for future work.Source: §10 · p. 20
  • Future work includes targeted-vertical evaluation, formal analysis of Skill Graph composition, sovereignty–performance studies with local small models, and longitudinal measurement of the Wisdom Listener effect.The conclusion says doctoral research will contribute the first quantitative before-and-after outcomes from a regulated healthcare deployment.Source: §10 · p. 20

AI-Assisted Writing Disclosure

The authors disclose generative AI assistance for manuscript writing while attributing the framework’s intellectual substance to the lead author. The paper also defines the Wisdom Listener as value accumulated by a tacit-capable foundation through use.

AI-Assisted Writing Disclosure

  • The authors used generative AI for language refinement, manuscript organization, and editorial assistance, under continuous author supervision and review.They state that AI tools did not contribute to the work’s intellectual substance.Source: AI-Assisted Writing Disclosure · p. 21
  • The authors attribute the framework’s concepts and integration to the lead author and state that references were independently verified.They accept responsibility for the paper’s claims and any remaining errors.Source: AI-Assisted Writing Disclosure · p. 21
  • The Wisdom Listener denotes operational value accumulated through use as interactions extend the tacit substrate available for later reasoning.The passage characterizes this as a self-reinforcing loop between system usage and competence.Source: AI-Assisted Writing Disclosure · p. 21
Loading 2609.04377v1…