The Semantic Layer AI Agents Actually Need
Onibex Team
Last updated: August 2026
What a semantic layer for AI agents is
A semantic layer for AI agents is a machine-readable description of enterprise data that tells an agent what each entity means, how tables relate, and which definition answers a business question. It lets an agent turn a natural-language question into governed, deterministic SQL, rather than guessing table names, field meanings, and join paths from a raw schema. This is the missing context that decides whether an agent returns a correct number or a confident wrong one.
The category term is worth stating plainly, because most teams building agents over enterprise data reach for a better model when the problem is a missing semantic layer.
Why do LLMs fail on real enterprise schemas?
Large language models write SQL well and chain steps well. They are unreliable at three things that a real enterprise schema demands, and none of the three is a code-generation problem.
The first is knowing which table answers which business question. A large ERP has thousands of tables, and the one a user means is rarely the one whose name looks closest to the question.
The second is knowing what a cryptic field means. A field named MATNR is a material number. Nine different date fields on one object mean nine different things. The model cannot recover that meaning from the name.
The third is knowing which join path is correct or cheapest. When several paths connect two tables, the model picks a plausible one, and plausible is not the same as correct. The result is SQL that runs and returns a wrong number.
Wrong numbers delivered confidently are worse than no answer, and the business notices. Because the failure is a semantics problem, not a code problem, a larger or newer model does not fix it. A semantic layer does.
What does the layer have to describe?
A semantic layer earns its name by formalizing the business meaning an agent cannot infer. At minimum it describes six things:
- Entities: the business objects, such as Customer, Product, and Sales Order, and what each one is.
- Grains: the level a fact is recorded at, so a measure is never summed across the wrong dimension.
- Measures: the quantities and how they are calculated, defined once rather than re-derived per query.
- Statuses: what a coded state actually means in business terms.
- Relationships: the correct join paths between entities, so the agent does not guess.
- Intent: the business questions a definition is built to answer.
Onibex ASK (Agentic Semantic Knowledge) formalizes exactly these into a layered contract that any agent runtime can consume. The point is not the list, it is that the list is written down once, in one governed place, instead of living in the heads of the people who wrote the last report.
How do the Gold, Silver, and Bronze layers work?
ASK organizes definitions into a medallion model with an explicit resolution priority, which is what makes agent behavior predictable.
| Layer | What it is | Agent visibility |
|---|---|---|
| Gold | A business definition, pre-joined and semantically resolved, such as an Open Sales Order Tracker | Preferred first |
| Silver | A reusable enterprise artifact such as Customer, Product, or Sales Order, composed from Bronze | Fallback when no Gold definition fits |
| Bronze | A raw source table, largely uninterpreted | Avoided, kept for lineage, not used as agent context |
An agent prefers Gold, falls back to Silver, and avoids Bronze. This ordering is the difference between an agent that answers from curated business definitions and one that improvises over raw tables.
Why is deterministic generation better than probabilistic guessing?
The reliability argument comes down to how the SQL is produced. A model that writes SQL freely is guessing, and it guesses differently each time. A semantic layer changes the task: the agent resolves the question against governed definitions, and the SQL is generated from that resolved plan rather than invented token by token.
ASK does this through three internal stages: Intent Resolution, which maps the question to the right definition; Graph Semantics, which resolves entities and the correct join paths; and SQL Generation, which produces the query from that resolved plan. The output is deterministic because it follows the semantic plan, not the model's improvisation.
What may an AI agent read, and what may it never change?
Enterprise governance for agentic AI converges on one rule: agents read governed facts and definitions, and business-changing actions enter only through the owning system's sanctioned interfaces. Agents never manipulate databases directly and never publish enterprise facts on their own.
A semantic layer implements the reading half of that rule cleanly. The agent resolves a question against the published layer, generates SQL against governed data products, and returns an answer. It does not touch source systems. An agent's action becomes a fact only after the owning system executes it. This is what lets a governance lead approve agent access without opening the database.
What is the ASK specification, and how do you adopt it?
ASK is published as two complementary things in one repository: the specification, a YAML contract that describes what a data product means, and the Onibex ASK Platform, the product that implements the standard end to end. The repository is source-available under a dual license, PolyForm Strict 1.0.0 or PolyForm Free Trial 1.0.0: noncommercial use, research, and evaluation are permitted, and production or commercial use requires a license from Onibex.
The separation matters for adoption. An enterprise can read and evaluate the specification openly, use it to describe how it structures AI-ready data products, and evaluate the platform separately. The specification lives in the repository at github.com/Onibex/agentic-semantic-knowledge-ask.
The platform surfaces are ASK Studio for authoring the semantic layer, ASK Setup for wiring databases, models, the LLM, OpenSearch, SAP, and the Model Context Protocol (MCP), and ASK Chat for business users. ASK runs on Kyma Runtime within SAP BTP.
A worked example: an open sales order question
A user asks an agent for the value of open sales orders for a customer this quarter.
Without a semantic layer, the agent has to find the sales order tables, know which status codes mean open, pick the join from order to customer, choose the right date field for the quarter, and compute the value. Any one wrong guess returns a confident wrong number.
With ASK, the agent resolves the question to a Gold definition, an Open Sales Order Tracker that already encodes the open statuses, the joins, the date semantics, and the value measure. Intent Resolution selects the definition, Graph Semantics confirms the entities and joins, and SQL Generation produces the query from that plan. The answer is governed and repeatable, and it is the same answer every time the question is asked.
Frequently asked questions
What is a semantic layer for AI agents?
It is a machine-readable description of enterprise data that tells an agent what entities mean, how they relate, and which definition answers a question, so the agent generates governed, deterministic SQL instead of guessing from a raw schema.
Why does text-to-SQL hallucinate on enterprise data?
Because the failure is a semantics problem, not a code problem. The model cannot reliably know which table answers a question, what a cryptic field such as MATNR means, or which join path is correct. A better model does not fix a missing-context problem; a semantic layer does.
What are the Bronze, Silver, and Gold layers in ASK?
Gold holds pre-joined, semantically resolved business definitions and is preferred first. Silver holds reusable enterprise artifacts such as Customer and Sales Order, used as a fallback. Bronze holds raw source tables, kept for lineage and avoided as agent context.
Can ASK be used without Onibex products?
You can read and evaluate the ASK specification, a YAML contract, from the source-available repository, which is dual-licensed under PolyForm Strict 1.0.0 or PolyForm Free Trial 1.0.0. Noncommercial use, research, and evaluation are permitted; production or commercial use of the platform requires a license from Onibex.
What ASK does not do
ASK grounds agents for reading and reasoning; it does not let an agent change source systems. Business-changing actions go through the owning system's sanctioned interfaces, and an agent's action becomes a fact only after that system executes it. ASK is not a fine-tuned model and does not depend on one; it supplies the business context a model lacks. The specification is source-neutral, so it is not limited to SAP, though the SAP depth is a differentiator. ASK does not claim to use LangGraph, OpenMetadata, MetricFlow, or any library not in its stack.
About Onibex
Onibex is an SAP Silver Partner and Confluent Gold Partner building real-time SAP data hyperconnectivity for the enterprise. Its flagship platform, Onibex OneConnect, streams SAP ECC and SAP S/4HANA business events to Apache Kafka, Confluent Cloud, Databricks, Snowflake, and other destinations in real time. Onibex also maintains ASK (Agentic Semantic Knowledge), a semantic-layer specification and platform that grounds AI agents in governed, business-ready SAP data. Onibex is headquartered in The Woodlands, Texas.