Blog
- From Data Vault to an AI-ready silver layer: same build order
Vault steps build an AI-ready silver layer's meaning layer: unhashed business keys, persistent staging audit, data contract, not point-in-time or bridge tables.
- You don't need to unlearn Data Vault to model for AI agents
In data modeling for AI, Ensemble Logical Modeling becomes a business ontology: the meaning layer of an AI-ready data architecture. It doesn't need the tables.
- AI for modeling, or modeling for AI?
AI-assisted data modeling guesses. Data modeling for AI writes business context for AI once as metadata in a context layer: AI proposes, a person approves.
- Who is the data model for?
A data model for AI isn't a new table shape. The model did two jobs: a diagram for people and a build instruction for machines. AI retires the second.
- Seven steps to a meaning layer
Where to start on a silver layer an AI agent can read: one business area, seven steps in order, and a check you can run at the end of each one.
- What a hub becomes when an agent reads the model
Data Vault and third normal form beside AIMS, construct by construct: where each hub, link and same-as link goes when an agent reads the model.
- Every method stops at the concept
Silver layer modeling methods say which business thing a table is about. Only the costliest says what each column means, and that's where AI answers break.
- Integration by exception
Silver layer integration is the last step, not an entry fee. AIMS adds a shared key and a golden record only when a business requirement needs one.
- What the agent reads
An AI agent's answer is only as good as what it reads, and every platform keeps its semantic layer for AI agents inside its own walls. Write the meaning first.
- You don't need Data Vault anymore
You don't need Data Vault or third normal form in your silver layer. AI reads business meaning from an ontology: the deciding stays, the rebuilding goes.
- Modeling for Machines That Can Read
Once a machine can read a sentence about your business, every benefit Data Vault's tables bought has a cheaper home. That's data modeling for AI.
- The economics of meaning flipped
Stated meaning now costs a sentence, not a table. What an ontology adds to AI accuracy on the benchmarks, and where the AI still has to stop.
- Context for AI, context by AI: metadata needs a confidence score
Why AI-generated metadata needs a confidence score: business context for AI comes from people building the conceptual data model; a local LLM still guesses.
- Keep AI-generated metadata marked after you accept it
Accepted AI-generated metadata keeps its data provenance: AI badge, confidence score and rationale stay in the data catalog so data governance sees each guess.
- A Model You Can Diff: Metadata Version Control for Data Teams
Pipeline code gets code review and an audit trail; metadata doesn't. Five metadata version control practices bring naming conventions under data governance.
- Six steps to your integration layer: from bronze to trusted gold
Bronze shipped and gold still isn't trusted. Here are six steps to the integration layer your medallion architecture is missing.
- Integration Layer: Pattern, Not Project
Data warehouse automation makes the integration layer a pattern, not a project: a catalog rendered through a Jinja template under one naming convention.
- Top-down and bottom-up data modeling: cart before the horse
Top-down and bottom-up data modeling: join the conceptual data model and business glossary to source system discovery, so staging tables carry data governance.
- The model is portable. The code is native.
Metadata-driven code generation should emit native code, not portable SQL. One governed model becomes a Databricks Asset Bundle; Snowflake and Fabric are next.
- This is why we need data engineers
Will AI replace data engineers? We built Ask AI into DeltaVault, then recorded it losing a race to a data engineer building one view by hand.
- Keep the model stable, change the template
Keep the model stable, change the template: metadata-driven code generation turns one governed model into native Databricks code, with other platforms next.
- No silver layer, no lakehouse
A medallion architecture with no defined silver layer is a dump house, not a lakehouse. Here is how to tell which one you built.
- The Mapping Is the Deliverable
The source to target mapping that turns a source system into the business model is usually never written down. Record it with who decided, when, and why.
- AI drafts the tedious part. You keep the decision.
AI-assisted data modeling in DeltaVault: skills draft keys, joins, views and descriptions as proposals you read first. Nothing writes until you accept.
- Meaning Leaks at Every Handoff
Business definitions in one tool, the technical data model in another, lineage in a third: meaning leaks at every handoff. DeltaVault keeps them in one model.
- The template engine is yours
DeltaVault's data warehouse automation templates are plain Jinja, versioned and team-owned, so your standards live in templates you control.
- AI that explains how reality became data
AI governance for data platforms starts with one loop. Every DeltaVault AI action follows five steps: context, content, describe the outcome, review, refine.
- You already own the stack
You already own the lakehouse, storage, and orchestration. The missing layer is data warehouse automation: business meaning in, Data Vault and workflows out.
- Edit the template, watch the code change
Jinja template code generation you can see: open DeltaVault's template playground, edit the template, and watch the generated code update live.
- Manage data, not plumbing
Declare table metadata once; DeltaVault generates create-table, merge and workflow code for Databricks, Snowflake or Fabric. That is data warehouse automation.
- The Medallion Three-Body Problem
Without an agreed business model, bronze, silver, and gold drift apart: three suns pulling the medallion architecture into a three-body problem.
- Should Software Be Free?
Free software still burns hosting, AI tokens, and human hours. A working position on honest software pricing and a free tier for the Data Vault community.
- Data Vault Needs New Rules for Lakehouse
Columnar storage, generated views, and AUTO CDC took over every load reason for Data Vault satellite splitting. What is left is business meaning.
- The Lakehouse Ended One Satellite Per Source
One satellite per source is a row-store habit. On Delta Lake width is cheap and joins are what you pay for: consolidate Data Vault satellites by meaning.
- Show Me Your Silver Layer
Any AI tool can land data in your bronze layer. The medallion architecture's real test is an integrated silver layer, and that is a people problem.
- Your Database Is Not Your Business
AI can infer a semantic layer from your technical exhaust. It cannot define your business ontology. Why a system of record is not a system of meaning.
- No Human, No Loop
Human in the loop puts a person at the end of an AI process as a checkpoint. In DeltaVault the human starts the loop, so there is no loop without one.
- The Model Says Yes. The Source Says No.
Your conceptual model says delivery address. Your source says address type. Describe the view in one sentence and keep column-level lineage on the result.
- Stop Asking Data Engineers to Paint
A conceptual data model breaks on first contact with a real source system. The fix isn't better artists. It's both sides recorded as metadata.
- The Meeting Is the Model
Turn a Teams meeting transcript into a conceptual data model in one workshop run, with every proposal traceable back to what was said in the room.
- How raw is your Data Vault?
In a raw vault, raw describes the data, not the structure. Why source-shaped vaults never integrate, and how to rebuild yours without starting over.
- AI Data Modeling: Enough Talk. Let's See It in Action
The AI business modeling workshop drafts concepts, relationships, and facts from your own content, shows its reasons, and writes nothing until you apply.
- How to stop the AI Guessing
The AI business modeling workshop checks your stories before it proposes a single concept. Here are the rules it applies.
- Where, Oh Where, Has Integration Gone?
An AI can model every source you own and still leave your warehouse un-integrated. Data integration is a decision about business keys, and AI can draft it.
- AI does not fix bad data models
AI data modeling aimed at your source schema returns a faster, shinier source schema. Business meaning comes from your own material, and sources map into it.
- Give AI Content, Not More Prompts
Upload, consolidate, and research business content into DeltaVault's AI knowledge base, then review every draft before it shapes your business model.
- Built, Not Painted
When the source does not hold the model's shape, AI helps build the bridge: discovered relationships, drafted views, and column-level lineage end to end.
- AI-native, not AI-added
What separates an AI-native data catalog from a bolted-on chat panel: governed writes, and AI skills you can open, edit, and extend.
- Build or buy: the cost AI hides
Build or buy a data platform? With AI-assisted coding, building your own looks cheap. The code was never the expensive part: owning it is.
- Business context is not an AI prompt
Business context for AI in DeltaVault is not what you pasted into the chat box. It is a governed setting every AI skill reads, written once in your own words.
- Meaning before machinery
AI made pipelines cheap to build. The scarce thing now is business context for AI: what your business means. DeltaVault captures it once and builds from it.