The ELM AI Data Modelling Framework Book Launch with Remco Broekmans

Key Takeaways

• Business-Led Authority: Business stakeholders retain sole decision-making power over data definitions and rules, while AI acts strictly as a junior assistant or facilitator.
• Independence from AI: All six core framework services (Discover, Connect, Review, Facilitate, Publish, and Compare & Consolidate) are fully functional with or without AI tools.
• Universal Protocol: A standardised JSON schema serves as the central protocol linking all services, maintaining a single source of truth across iterations.
• Local, Air-Gapped Deployment: To prevent corporate data leaks and avoid heavy cloud API costs, the framework runs offline on lightweight local LLMs behind enterprise firewalls.
• Workshop Questions Over AI Guessing: Rather than hallucinating or assuming unverified business rules, the Discover service generates targeted workshop review questions for human clarification.
• Non-Automated Consensus: Business consensus cannot be automated; outputs from multiple LLM families are merged via Compare & Consolidate to spark human stakeholder debate.
• Controlled Model Modifications: The Facilitate service is the only component authorised to modify an established model through live Validation and Evolution modes.
• Tool Interoperability: Conceptual model reasoning is kept separate from visual publishing, allowing direct exports to tools like CaseTalk, Ellie, Mermaid diagrams, and DDL.

Webinar Details

Title: The ELM AI Data Modelling Framework Book Launch with Remco Broekmans
Date: 2026-09-21
Presenter: Remco Broekmans
Meetup Group: Book Launch with Technics Pub x MWS
Write-up Author: Howard Diesel

How was the ELM AI Data Modelling Framework developed?

How was the ELM AI Data Modelling Framework created?

Remco Broekmans, a Dutch data modeller and trainer at GNC Academy, developed the ELM AI Data Modelling Framework to bridge communication gaps between business stakeholders and IT teams. Building on his first book From Storage to Solution, Broekmans published The ELM AI Data Modelling Framework on September 2, 2024.

The guiding principle of the framework is that business stakeholders, not IT or AI, must retain ultimate decision-making authority over data logic. AI functions strictly as a communication facilitator across six structured services.

Guiding Principles

• Business-Led Authority: Business experts retain final decision-making power over models.
• AI as Facilitator: Large Language Models (LLMs) operate as junior assistants rather than autonomous creators.
• AI-Independent Logic: The core modelling methodology functions effectively even without AI tools.

Key Takeaways

• Business stakeholders hold sole authority over data model definitions and business rules.
• AI acts as a communication facilitator across structured data modelling workflows.
• The framework features a foreword written by data warehousing pioneer Bill Inmon.

FAQ

• Does the framework require AI to function? No, all six core services work fully with or without AI involvement.

Figure 1 The ELM AI Data Modelling Framework Book Launch

Figure 2 Introducing the Speaker

Figure 3 Author of Two Books

Figure 4 What the Presentation will Cover

How did AI Modelling Evolve to Local Micro-services?

How did AI data modelling shift from prompt engineering to local LLMs?

AI data modelling evolved from simple prompt engineering to structured, local micro-services due to security, cost, and context constraints. Early experiments with single prompts and custom ChatGPT GPTs extracted business concepts but suffered from context drift and unpredictable structures.

To solve corporate privacy concerns and heavy cloud resource consumption, Remco Broekmans transitioned from centralised cloud LLMs to small local models running on mini-PCs. Single monolithic prompts were broken down into specialised micro-services.

Evolutionary Phases

• Single-Prompt Engineering: Initial tests using monolithic prompts to generate entity-relationship diagrams.
• Custom GPTs: Ingesting knowledge files and templates to generate models and workshop questions.
• Local Micro-Services: Running lightweight open-source models on mini PCs for offline security and speed.

Key Takeaways

• Monolithic prompts fail on local hardware, requiring micro-service architecture.
• Local LLMs eliminate corporate privacy risks and recurring cloud compute costs.
• Generating workshop questions prevents AI models from making unverified assumptions.

FAQ

• Why transition away from cloud-based models like ChatGPT? Local models keep proprietary corporate data behind firewalls and avoid cloud API expenses.

Figure 5 From One Assistant to a Service Framework

Figure 6 From One Assistant to Service Framework: Early Experiments

Figure 7 Early Experiments -> Rethink the Work

What is the ELM AI Data Modelling Framework?

What is the architectural design of the ELM AI Data Modelling Framework?

The ELM AI Data Modelling Framework uses a standardised JSON schema as a universal exchange protocol between modular services. This architecture treats LLMs as junior data modelling assistants, separating model logic from technical visualisation and publishing layers.

Multiple LLM families evaluate the same business case independently. Their outputs are compared and consolidated before human stakeholders reach consensus, as business consensus cannot be automated.

Architectural Components

• JSON Schema Protocol: Standardises business concepts and relationships across all services.
• Separation of Concerns: Keeps business reasoning distinct from visual publishing tools.
• Multi-Model Evaluation: Uses multiple LLM families (e.g., Llama, Qwen, Mistral) to generate diverse model candidate options.

Key Takeaways

• JSON schemas act as the single source of truth connecting all framework services.
• AI provides candidate models, but human business stakeholders make all consensus approvals.
• Modular design allows easy swapping of LLMs and publishing tools.

FAQ

• Can AI automatically establish business data consensus? No, business consensus requires human stakeholder discussion and cannot be automated.

Figure 8 From One Assistant to Service Framework

Figure 9 The Problem: AI Can Model Fast, but Meaning Needs Governance

Figure 10 The Answer: Separate the Work into Services

Figure 11 Service Architecture in One Slide

Figure 12 The Structure Business Model is the Shared Memory

Figure 13 The Modelling Lifecycle: from Language to Communication

How does Discover Service Extract and Classify Terms?

How does the Discover Service extract business terms and prepare workshop questions?

The Discover Service analyses business documentation or transcripts to extract core terms and classify them into categories. Rather than guessing ambiguous business logic, the service generates targeted workshop questions for human clarification.

Terms are categorised using ELM classifications (Event, Person, Place, Thing, or Other) or alternative frameworks like John Gorman’s 16 facets. This structured extraction establishes a clean, business-defined vocabulary.

Core Extraction Steps

• Term Extraction: Identifies business concepts directly from meeting transcripts or documentation.
• Concept Classification: Categorises concepts into structural categories (e.g., Events vs. Persons).
• Question Generation: Flags ambiguities and outputs precise workshop review questions.

Key Takeaways

• The Discover Service categorises business vocabulary using established modelling facets.
• AI generates workshop questions instead of making risky guesses about business logic.
• Business definitions remain human homework, with AI offering draft suggestions.

FAQ

• How does the Discover Service handle raw meeting recordings? It processes verbatim transcripts from tools like Fireflies to extract classified vocabulary and workshop questions.

Figure 14 Discover

What are the Services of the ELM Framework?

What are the core services of the ELM AI Data Modelling Framework?

The ELM AI Data Modelling Framework consists of six specialised services designed to transition business concepts into validated models. Each service performs a distinct task within an iterative modelling pipeline.

Service Breakdown

• Connect: Identifies relationships triggered by business events, transactions, or natural hierarchies.
• Review: Evaluates model quality, identifying assumptions, strengths, and debate points.
• Facilitate: Runs live workshop sessions in Validation mode (querying logic) or Evolution mode (incorporating approved updates).
• Publish: Converts JSON schemas into external formats like CaseTalk, Mermaid diagrams, or DDL.
• Compare & Consolidate: Merges outputs from multiple LLMs or teams into a unified consensus model.

Key Takeaways

• The Facilitate Service is the only service authorised to modify an established model.
• Connect links terms based on transaction events and structural hierarchies.
• Compare & Consolidate synthesises multi-model perspectives into a single refined schema.

FAQ

• What is the difference between Validation and Evolution modes in Facilitate? Validation checks if the model can answer business queries; Evolution incorporates new approved business knowledge.

Figure 15 Connect

Figure 16 Review

Figure 17 Facilitate

Figure 18 Publish

Figure 19 Compare & Consolidate

How is the ELM AI Data Modelling Framework executed?

How is the ELM AI Data Modelling Framework executed in a local environment?

The framework operates locally using Open WebUI and lightweight open-source models (such as Llama 3.2, Phi, and Gemma) running on mini-PCs or laptops. System prompts define specialised skills and intern-like constraints to execute each service cleanly.

During execution, raw business cases pass through Discover and Connect prompts, producing structured JSON schemas. The Facilitate Service tests business scenarios, such as distinguishing car-sharing reservation roles, and exports verified models to CaseTalk or Mermaid.

Implementation Highlights

• Local UI & Models: Runs offline via Open WebUI using small, efficient LLMs.
• Token Optimisation: Manages context memory by passing focused JSON payloads between services.
• Export Integrations: Ingests schemas directly into CaseTalk for fact-based modelling.

Key Takeaways

• Open WebUI manages offline model prompts and specialised domain skills.
• Token memory limits on small hardware require modular JSON data transfers.
• Automated exports bridge AI-generated models directly into CaseTalk and visual diagrams.

FAQ

• Can this local setup export directly to commercial data modelling tools? Yes, schemas publish directly to tools like CaseTalk, Ellie, and Mermaid code.

Figure 20 Demo

Figure 21 GA Local AI Modelling Assistance

Figure 22 Models

Figure 23 Knowledge

Figure 24 Skills

Figure 25 CaseTalk: CBC-List & NBR Canvas

Figure 26 CBC List – Core Business Concepts

How does the Framework Maintain Air-gapped Security?

How does the framework maintain air-gapped security and data governance compliance?

The ELM AI Data Modelling Framework operates completely air-gapped behind enterprise firewalls, ensuring zero external network transmission of sensitive business data. Local server deployments allow organisations to maintain strict data governance and PII compliance.

The framework integrates directly with mandated enterprise business glossaries, automatically mapping business aliases (such as “client” or “Kund”) to approved terms (like “Customer”). It explicitly functions as a conceptual modelling bridge, not a physical database builder or prompt engineering trick.

Security & Boundaries

• Air-Gapped Security: Local hardware deployment prevents external API data leaks.
• Glossary Enforcement: Mandates approved enterprise terms while preserving alias mappings.
• Defined Boundaries: Excludes physical database DDL creation, ETL pipelines, and automated prompts.

Key Takeaways

• Air-gapped execution ensures total protection of sensitive PII and corporate data.
• Enterprise business glossaries enforce consistent nomenclature across AI outputs.
• Physical pipeline execution is handed off to specialised tools like Data Vault Builder or CaseTalk.

FAQ

• Does the framework replace physical data modelling tools? No, it produces conceptual/logical schemas and hands off physical implementation to downstream data tools.

Figure 27 GA Local AI Modelling Assistance Framework

Figure 28 North Star – Data Town Plan

Figure 29 What the Framework is Not

Figure 30 Best Practises: Make Modelling Collaborative

How can Readers Apply the ELM AI Framework?

How can readers purchase the book and apply the framework to legacy systems?

The ELM AI Data Modelling Framework is available globally on Amazon in print (colour, as well as black-and-white for Australian distribution) and Kindle, or directly as a PDF from the author.

During the final Q&A, Broekmans explained how the framework applies to brownfield and hybrid systems by extracting glossaries from existing data marts and comparing them against new business requirements using the Compare & Consolidate service. He also pointed to integration potential with graph ontologies and FCO-IM fact-based modelling.

Availability & Applications

• Publishing Formats: Amazon print/Kindle editions and direct author PDF distribution.
• Brownfield Integration: Comparing legacy data mart schemas with newly discovered models.
• Methodology Support: Expanding framework services to support FCO-IM fact modelling and graph ontologies.

Key Takeaways

• Print and digital formats are accessible worldwide with direct PDF options from the author.
• The framework evaluates legacy brownfield systems by running Compare & Consolidate across old and new models.
• Future service extensions will integrate Fact-Based Oriented Modelling (FCO-IM) and formal graph verification.

FAQs

• Can the framework be used if an organisation already has legacy databases? Yes, existing glossaries can be ingested and compared against new models using the Compare & Consolidate service.

Figure 31 Q&A