MyriadeAI on enterprise data: do you need a semantic layer?
Benjamin Derville · Updated November 10, 2026 · 8 min readThis guide describes what needs to be in place for an AI (assistant, agent) to answer correctly on a company’s data: what works out of the box, why it fails, the possible levels of governance, and which level is needed for which use.
01What works without governance
An AI connected directly to a database works well when:
- there are few sources and a readable schema (explicit table and column names)
- users can check a SQL query (IT / data team)
- questions are exploratory: a wrong number has no consequence
The criterion is not company size. A small company on a poorly configured ERP can be harder than a mid-size company with a well-modeled warehouse.
02Why AI fails on data
| Cause | Symptom | What addresses it |
|---|---|---|
| Complex or messy data | Wrong tables, wrong joins | Catalog |
| Poor data quality | Wrong answer given with confidence | Catalog (quality, freshness) |
| Company-specific vocabulary | The AI doesn’t understand “active customer”, “net revenue” | Glossary |
| Company-specific KPIs | Different tools, different numbers, loss of trust | Semantic layer (models, metrics) |
| Access rights | Sensitive data leaks | Database-level permissions |
| Uncontrolled cost | Warehouse bill grows out of control at scale | Pre-aggregated models, cost monitoring |
03The levels of governance
Each level adds a layer of knowledge the AI can rely on. You rarely need all six.
| Level | Content | Form | Maintained by |
|---|---|---|---|
| 1Context | The company, its activities, its users | Text | – |
| 2Catalog | Table and column descriptions, quality, freshness, lineage | Text + metadata | IT / data |
| 3Glossary | Business terms and their meaning | Text | Business |
| 4Access rights | Who can see which data | Permissions | IT / data |
| 5Semantic layer (models, metrics) | Formalized, standardized calculations | Model, code | IT / data |
| 6Ontology | Relationships between entities, possible actions | Model, code | IT / data |
1Context
What the company does, its business lines, who asks the questions and why. Without context, the AI misses the point even with the right tables.
2Catalog
What each table is for, what each column contains, whether the data is reliable and up to date. This is the level that prevents wrong joins and stale tables. Example tools: Unity Catalog, DataHub, native warehouse catalogs.
3Glossary
The company’s vocabulary, in plain language: “active customer = ordered in the last 12 months”. It aligns the AI with how the company talks. It does not lock in the calculation.
4Access rights
Access rights come into play when the AI is opened to business teams. Depending on the company, this happens before metrics (business users work from the glossary) or after (only once the key metrics are formalized).
Set them at the database level (tables, columns, rows). Permissions defined only in a semantic layer can be bypassed as soon as the AI writes its own SQL.
5Semantic layer (models, metrics)
The calculation is written once, in code, and everyone uses it. This is what is usually called a “semantic layer”. Example tools: dbt / MetricFlow, Cube, LookML, Snowflake semantic views, Databricks metric views.
6Ontology
The formalization of business objects (customer, order, store), their relationships and the actions that can be taken on them. Useful when the AI does not just read but acts. Example: Palantir Foundry’s ontology.
04Which level for which use
| Use | 1Context | 2Catalog | 3Glossary | 4Access rights | 5Semantic layer | 6Ontology |
|---|---|---|---|---|---|---|
| Exploration by the IT / data team | ||||||
| Self-service for business users (read) | ||||||
| Agents that act (write, trigger) |
NeededDepending on complexity
Agents that act also need explicit permissions on their actions.
05Where to start
- Connect the AI to the data and use it internally (data team).
- Document the catalog: most-used tables first.
- Write the glossary with business teams.
- Set up access rights at the database level, before opening to business teams.
- Formalize as metrics only the indicators that must be identical everywhere.
06Lexicon
- Catalog
- Documented inventory of tables and columns.
- Semantic layer
- The set of formalized business definitions (broad sense: glossary + metrics; strict sense: metrics in code).
- Glossary
- Definitions of business terms in plain language.
- Lineage
- Where a piece of data comes from and what transformations it went through.
- Metric
- Formalized calculation of an indicator.
- Ontology
- Formal model of business objects, their relationships and the possible actions.
07How Myriade does it
Most of this work used to be done by hand, over months. Myriade uses AI to build the governance; people validate it.
| Level | Myriade’s AI | Your team |
|---|---|---|
| Context | Captures the company context and the questions users ask | Can review |
| Catalog | Documents tables and columns autonomously by exploring the data; asks questions when there are ambiguities or missing information. | Answer & validate |
| Glossary | Proposes business terms and their definitions, from the data and from usage | Validate with business teams |
| Metrics | Drafts the semantic models for the indicators that must be identical everywhere | Approve |
| Access rights | Proposes database-level permissions for each role in the organization | Adapt & validate |
This means that in a few hours, we can lay solid foundations for using AI on your data, in any AI assistant.