This guide helps you, as a CDO or head of data engineering, decide how to structure enterprise data models and which framework fits your lakehouse.
An enterprise data modeling framework is the structured approach an enterprise follows to design and govern its data models.
Key Takeaways
- An enterprise data model documents information requirements independent of any implementation, and a framework standardizes how those models are designed and governed
- Frameworks range from architecture frameworks such as Zachman, TOGAF and DAMA-DMBOK to warehousing methodologies such as Inmon, Kimball and Data Vault
- On Databricks, Data Vault and third-normal-form models typically sit in silver and Kimball star schemas in gold
- Schema designs take weeks and code generation is manual, while generic AI assistants invent plausible columns and joins
- Agentic EDM answers with five commitments, from measuring the data first to shipping the operating kit, and an engineer approves every phase
- Four workflows and 31 reviewable phases run inside a Databricks workspace, and the roadmap has three themes
What Is an Enterprise Data Modeling Framework?
An enterprise data modeling framework is the structured approach an enterprise follows to design and govern its data models, from logical and physical models to code and documentation.
An enterprise data model is a data model independent of any implementation, documenting the information requirements of an enterprise in canonical terms (Source). A framework builds on that model and standardizes how it is designed and governed.
Agentic EDM is LatentView’s agentic AI enterprise data modeling framework. It automates end-to-end enterprise data modeling processes, including four areas:
- LDM and PDM generation
- ETL/DDL/DML code creation
- Confluence documentation
- Governance workflows
A unified, GUI-driven experience lets both technical and business users create analytics-ready data assets while ensuring standardization and compliance.
Each workflow starts from the live catalog and ends at reviewed, executable, documented assets. The agent drafts and checks, and the engineer decides at every phase, on evidence.
Enterprise data architecture is a wider scope: data models define structure within systems, while enterprise data architecture defines structure across the organization. Our enterprise data architecture guide covers it in full.
The Evolution of Enterprise Data Modeling Frameworks
Enterprise data modeling frameworks now span Inmon and Kimball warehousing methodologies, the more recent Data Vault pattern and the layered medallion architecture of the lakehouse.
Architecture frameworks such as Zachman, TOGAF and DAMA-DMBOK are covered in our enterprise data architecture guide, so this article focuses on how models are built.
Kimball’s dimensional modeling is a bottom-up approach that denormalizes business data into dimensions and facts, with different subject areas connected through conformed dimensions. The star schema is its most common form. Data Vault is a more recent pattern, used to build data warehouses for enterprise-scale analytics, that organizes data into hubs, links and satellites.
On the lakehouse, both styles have a place. Normalized Data Vault models are write-optimized and denormalized dimensional models are read-optimized: hubs and satellites in silver load the dimensions in the star schema, and link tables drive the facts. Databricks notes that different projects on a lakehouse can apply different modeling techniques.
On the delivery side, schema designs take weeks and code generation is manual, while documentation lags.
Types of Enterprise Data Modeling Frameworks
Enterprise data modeling frameworks include architecture frameworks, warehousing methodologies, lakehouse design patterns and agentic AI frameworks, each covered in this guide.
| Type | Examples | What It Covers |
|---|---|---|
| Architecture frameworks | Zachman, TOGAF, DAMA-DMBOK | Frameworks that a data architecture can draw from |
| Warehousing methodologies | Inmon’s Corporate Information Factory, Kimball’s dimensional modeling, Data Vault 2.0 | Three leading data warehousing approaches described by Databricks |
| Lakehouse design patterns | Bronze, silver and gold layers | The layers through which lakehouse data is curated |
| Agentic AI frameworks | Agentic EDM | Automation of end-to-end enterprise data modeling processes |
The Zachman Framework was developed by John Zachman at IBM in 1987 and uses a matrix of six layers mapped against six questions. TOGAF was developed by The Open Group in 1995 (Source).
Databricks describes Inmon’s Corporate Information Factory, Kimball’s dimensional modeling and Data Vault 2.0 as three leading data warehousing approaches, and shows how each one maps to the medallion architecture (Source).
Agentic EDM is LatentView’s agentic AI framework, and the sections that follow cover how it works.
Why Traditional Enterprise Data Modeling Breaks Down
Traditional enterprise data modeling breaks down because schema designs take weeks and code generation is manual, while documentation lags and generic AI assistants guess.
The key challenges of traditional enterprise data modeling are:
- Bronze data arrives with duplicate copies and regional splits, along with JSON payloads and undeclared keys, and someone has to profile it before anyone can model it
- Schema designs take weeks, and code generation is manual
- The diagram, the SQL, the checks and the documentation are written separately, and they stop agreeing within a sprint
- Fan-out joins and summed ratios produce numbers that look right and are not, and nothing flags the problem before a dashboard does
- Documentation lags the build, and governance is an afterthought
- Generic AI assistants invent plausible columns and joins, and the reviewer inherits the risk of finding them
Migration is judgment work, and most tools guess. AI projects also depend on AI-ready data: Gartner predicts that through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data (Source).
Core Pillars of an Agentic Data Modeling Framework
An agentic data modeling framework works from five commitments that tie each model to measured data and to an engineer’s approval at every reviewable phase.
Agentic EDM states its approach as five commitments:
| Commitment | What It Means |
|---|---|
| Measures first | Profiles, joins and keys are queried before any model is written |
| One approved contract | Every artifact renders the same approved spec |
| Checks before people | Structural validation and rubric scoring run before a person reviews |
| A human on every phase | A person approves, or rejects with feedback |
| Ships the operating kit | Pipelines, checks, documentation and catalog comments come with the build |
These commitments answer four problems: messy bronze data, models that drift from code, KPIs that quietly double-count and generic AI that guesses.
Our guide to agentic enterprise data modeling covers each commitment in depth, from the five evidence steps to the six verification layers.
The Four Core Workflows of Agentic EDM
Agentic EDM runs four core workflows across 31 reviewable phases, and each one starts from the live catalog and ends at reviewed, executable, documented assets.
- Lakehouse: Runs bronze to silver in ten phases and silver to gold in ten more, taking requirements to profiled, joined and reconciled silver tables and then to KPI tables that reconcile
- Data Modeling: Moves from the logical model to the physical model, DDL, ingestion and documentation in five phases, with design approved before tables exist
- Semantic Layer: Turns a plain-English question into an EXPLAIN-checked query in three phases, with a query-plan diagram and documentation
- Auto Docs: Documents an existing schema in three phases, with an overview, a whole-schema ERD and a per-table reference published as one Confluence page
Eight of the 31 phases never call a model, and all four workflows share one checked cycle of context, generation, validation, evaluation and review. Our guide to agentic enterprise data modeling walks through the phases and the use case for each workflow.
Enterprise Data Modeling Architecture on Databricks
On Databricks, enterprise data modeling maps to bronze, silver and gold layers, with Unity Catalog and Jobs holding the tables, comments and pipelines.
Databricks describes how different modeling techniques fit in each layer of the lakehouse:
- Bronze: The landing zone for source data, where table structures correspond to the source systems as-is, with history, lineage and the ability to reprocess
- Silver: The enterprise view of key business entities, where data is matched, merged, conformed and cleaned just enough, with third-normal-form or Data Vault models and enterprise-level rules
- Gold: The presentation layer, where denormalized, read-optimized models such as Kimball star schemas or Inmon-style marts fit, with project-specific rules applied as data moves up from silver
The Lakehouse workflow follows the same layers: ten phases take bronze data to silver, and ten more take silver to gold. Incremental pipelines are deployed paused and are idempotent by name, dimensions load before facts, and a redeploy updates and never duplicates.
For a leading US health insurer with no central data hub, whose data scientists worked around slow on-premises infrastructure, we built a governed Member 360 foundation. The team moved a Member DNA feature repository onto Azure Databricks Unity Catalog and deployed a medallion architecture with rigorous ETL quality testing.
Technical Frameworks in the Agentic EDM Stack
The Agentic EDM stack is built on LangGraph for orchestration, FastAPI for hosting, Angular for the studio and MLflow with Postgres for tracing and recovery.
- Orchestration with LangGraph: One graph per workflow is built as a state machine, so each phase stays small enough to verify on its own
- Hosting with FastAPI: One service on Databricks Apps hosts every workflow and streams progress to the browser by server-sent events
- The studio in Angular: Requirement input, progress and review all happen in one interface
- Tracing and recovery with MLflow and Postgres: MLflow records one run per workflow, with a span for each node and each model call, and the Postgres store checkpoints gates so they survive restarts
Behavior is held in 24 versioned prompt files, and the model endpoint is configuration. The platform is Databricks-native: compute, model serving, catalog, jobs and tracing all come from the workspace. The full component list is in our guide to agentic enterprise data modeling.
Governance and Human Review in Enterprise Data Modeling
Governance in agentic enterprise data modeling sits inside the workflow: each approval gate matches the decision, and writes happen only on an explicit click.
It is not one big approve button: the gate matches what is being decided.
- Table selection: A checkbox grid shows every table, why it matched, its rows, key candidates and redundancy notes, and the selection is validated against the profile
- Statement by statement: Each target’s CTAS is approved, executed or rejected on its own, and rejecting one regenerates only that one
- Confirm the facts: Grain, additivity and join-safety hypotheses are each confirmed or corrected before any KPI is modeled, and corrections become feedback
- Execute is explicit: Validation is read-only, using EXPLAIN and LIMIT 0 probes. Writes happen only when a person presses Execute or Deploy, and replacements need a confirmation
Every run is answerable. One MLflow run per workflow records a span for each node and each model call, and tokens, latency and quality score are recorded per iteration, so retries show as a trend. The per-phase score, reason and reviewer feedback are kept as an audit list in the UI. Every artifact can answer what it cost, how long it took, what it scored and who approved it.
Profiling, search, join proofs, EXPLAIN and investigation are read-only, and tables, comments, pipelines and pages are written only on an explicit click. Gold investigation returns aggregates only, so no row values reach the model, and diagram rendering makes no network calls.
A stray approval with no gate armed is refused, so the run cannot be advanced by accident. The ingestion phase scores as its worst target.
Evaluation Checklist: How to Choose the Right Enterprise Data Modeling Framework
Evaluate an enterprise data modeling framework against an eight-point checklist, then test the answers on one domain before committing to a wider rollout across the enterprise.
Use these eight questions to test any framework, and compare the answers with how Agentic EDM responds:
| Checkpoint | Question to Ask | How Agentic EDM Answers |
|---|---|---|
| Platform fit | Do the layers map to your lakehouse? | Bronze, silver and gold layers, with tables, comments and pipelines held in Unity Catalog and Jobs |
| Evidence base | What are the models built from? | Queried statistics, proven joins and confirmed keys |
| Consistency | How do artifacts stay in agreement? | Every artifact renders the same approved spec |
| Verification | What is checked before a person reviews? | Structural validation, schema probes, EXPLAIN and rubric scoring |
| Human control | Where do approvals happen? | Gates matched to each decision, with writes on an explicit click |
| Traceability | Can each run be audited? | One MLflow run per workflow, with scores, reasons and reviewer feedback kept |
| Privacy | What does the model see? | Aggregates only from gold investigation, and diagrams rendered in-process |
| Documentation | What ships with the build? | A migration record, comments in the catalog, and Confluence or Markdown output |
Databricks places Data Vault in silver and star schemas in gold, so methodology can be matched to each layer.
A test on one domain makes the checklist concrete:
- Choose one domain
- Ask for the evidence for each model, including the profiles, the joins and the key checks
- Review the approval record and the checks for each phase
- Compare the totals with baselines captured before the build
How to Implement an Enterprise Data Modeling Framework
Implement an enterprise data modeling framework as a governed run in two stages, a silver stage that profiles and reconciles data and a gold stage that builds KPI tables, with every phase approved.
Agentic EDM runs on the workspace you already have, using its compute, model serving, catalog, jobs and tracing, and an engineer approves each phase. Documentation can be published to Confluence or downloaded as Markdown.
Run the Silver Stage
The silver stage runs ten phases: Require, Profile, Discover, Analyse, Logical, Physical, Load, Verify, Pipeline and Document. The main steps are:
- Require: Paste text or upload one of 11 file types, and review the extracted text shown before the run starts. Each rule quotes your words verbatim
- Discover: Review the checkbox grid, where each table shows why it matched, and approve the selection
- Load: Each target’s load statement is approved, executed or rejected on its own, and EXPLAIN checks every statement
- Verify: Reconciliation checks run one per declared quality rule, each reported as PASS, FAIL or ERROR
- Pipeline: Incremental pipelines are deployed paused, so nothing runs until you resume them
- Document: An eight-section migration record, written from what the run actually did, including failed checks, is published to Confluence or downloaded as Markdown
Run the Gold Stage
The gold stage runs ten phases: Require, Discover, Investigate, Confirm, Model, Diagram, Build, Verify, Pipeline and Document. Across them:
- Ten standing questions are asked of every model, and investigation tools return aggregates only
- Grain, additivity and join-safety hypotheses are confirmed or corrected before any KPI is modeled
- Requests become metrics, dimensions and targets, with ratios kept as numerator and denominator and anything unsettled labeled an assumption
- A model is rejected if it sums across a fan-out join or sums or averages a ratio
- Additive measures are compared with baselines captured before the build, and grain uniqueness is re-checked on the built tables
- Refresh uses a trailing window of three periods, or a full rebuild when there is no time grain
- The stage closes with a six-section silver to gold hand-off
Each phase ends at a human gate, where a person approves or rejects with feedback.
Future Roadmap for Enterprise Data Modeling
Databricks now offers an LLM agent that generates governed Silver-layer data models in Unity Catalog, and the Agentic EDM roadmap has three themes that build on workflows running today.
Direction on Databricks
In July 2026, Databricks introduced Vibe Data Modeling, a Databricks-native LLM agent that produces the analytical Silver-layer business model from a plain-English description of a business (Source). Databricks describes these characteristics:
- Every plain-English change produces a new versioned model, and no version is overwritten
- Models are validated against 251 enforceable rules and reviewed by two architect personas
- One artifact, model.json, is the source for every downstream artifact, including the Unity Catalog deployment, the metric views and the sample data
- Foreign keys are applied as informational constraints, and metric views are ready for AI/BI dashboards and Genie
- The same logical model can be deployed as one catalog, a catalog per division or a catalog per domain
Agentic EDM Roadmap
The Agentic EDM roadmap has three themes, each building on the four workflows and 31 reviewable phases that run in a Databricks workspace today.
Enterprise Readiness covers rollout across teams:
- Single sign-on with per-user data access
- Usage, time and cost per run in the studio
- Audit trail persisted to Delta
- Run history to reopen and re-export past runs
Lakehouse Depth moves toward one pipeline, bronze to gold:
- Gold stage delivered through the Lakehouse pipeline
- Primary and foreign-key constraints applied from the approved spec
- Source files read directly in Data Modeling
Databricks documents primary and foreign keys on Unity Catalog tables as informational constraints that are not enforced, and constraints can be added to existing tables (Source).
Reach goes beyond one platform:
- Second platform through a warehouse adapter and dialect module
- Per-workflow model, quality threshold and retry budget
How LatentView Approaches Enterprise Data Modeling
Choosing a framework is one decision, and running it on the lakehouse you already have is another. LatentView built Agentic EDM to run on your Databricks workspace, using its compute, model serving, catalog, jobs and tracing.
As an end-to-end analytics partner, LatentView is platform-fluent, with 20+ years of enterprise analytics experience and 50+ Fortune 500 clients. Our Databricks Center of Excellence unites 400 data engineers, and we hold Elite C&SI status with Databricks. Start with one domain: talk to our data engineering team about Agentic EDM on your Databricks workspace.
Frequently Asked Questions
1. What is an enterprise data modeling framework?
A structured approach for designing and governing the data models an enterprise uses, from logical and physical models to code and documentation.
2. What are the main types of enterprise data modeling frameworks?
Architecture frameworks such as Zachman, TOGAF and DAMA-DMBOK, warehousing methodologies such as Inmon, Kimball and Data Vault, and lakehouse design patterns such as medallion.
3. Which modeling approach fits a Databricks lakehouse?
Databricks places Data Vault or third-normal-form models in silver and Kimball star schemas in gold, and both styles have a place.
4. How is an enterprise data modeling framework different from enterprise data architecture?
Data models define structure within systems, while enterprise data architecture defines structure across the organization.
5. How should a team implement an enterprise data modeling framework on Databricks?
Run two stages: silver profiles and reconciles data, gold agrees the facts and builds KPI tables, and an engineer approves every phase.