Concept guide, Platform

What is a semantic layer?

Author: PureField
Published: 30 October 2026
Updated: 30 October 2026

A semantic layer is the layer of definitions that translates raw data into business language. Which record counts as a customer, how a quote connects to a unit and how win rate is calculated are written down once. Reports and AI agents answer questions from these definitions, so a word means the same thing on every screen.

Why is everyone talking about semantic layers now?

Because AI agents have started asking questions of company data, and a language model uses your company's concepts correctly only once they are defined for it. Is an "active customer" one who ordered in the last 90 days, or one with a running maintenance contract? Is a quote won on the day it is signed, or when the order is opened? People settle these differences in a meeting; an agent works with the definition it is given. Without one, it picks the most likely reading and can give two different answers to the same question on two different days.

That is why large software vendors are investing in semantic layers. Salesforce describes Tableau Semantics as a layer that translates your data into your business language and grounds the answers of its Agentforce agents in it. Analyst firm ISG writes that semantic modelling gives data its business meaning and provides the context both BI and AI agents need.

What does a semantic layer define?

A semantic layer defines five things: record types, the links between records, measures, the company's own terms and who can see what.

Definition	What it describes	Example from a company that sells and services equipment
Record types	What each record is and what its fields mean	Equipment record: serial number, install date, warranty end
Relationships	How records connect to each other	Quote line → equipment record → work order → service report → invoice
Measures	Formula, period and which records count	Win rate: accepted quotes / (accepted + rejected)
Company terms	The words the team uses and their synonyms	"Unit", "machine" and "device" describe the same record type
Permissions	Who sees which records and measures	A technician sees their own work orders, the service manager the whole region

What happens when definitions are scattered?

The same question gets two different answers, and the team argues about where a number came from instead of what it means. The sales manager sees a 32% win rate on the dashboard; finance's spreadsheet says 27%. The difference is in the formula: one counted expired quotes in the denominator, the other did not. Both numbers are right on their own terms; the problem is that the definition was written twice, in two places. Ask an AI agent the same question and it has to choose which definition to use.

A semantic layer moves that argument to one place: the definition is written once, and the dashboard, the report and the agent all read it.

Where does a semantic layer live?

A semantic layer is built either on top of the warehouse where you collect your data or inside the business software where the data is born.

On top of a data warehouse	Where the data is born
Where the data comes from	Copied from several programs	Created as a record when the work is done
When definitions are written	After the data is moved, in a separate modelling tool	Record types and links come with the record; company terms and measures are added
Whose job it is	Data engineers and analysts	The team that sets up the software and the company's managers
Keeping it current	The model is updated for every new program and field	Links stay current as records change
Best fit	Sales, service and finance run in separate programs	Sales, service and finance run on one platform

Which one is right depends on where your data is born. If the work runs in ten separate programs, a layer on top of the warehouse brings them into one language. If the work runs on one platform, the shortest path is to keep meaning where the record is born: the moment a quote is accepted, it is clear which units it created, and the moment a work order is opened, it is clear which unit it belongs to.

How do AI agents use a semantic layer?

The agent maps the words in a question to definitions, builds the query from those definitions and answers together with the records it relied on. For "Which customers have units coming off warranty this quarter?" the flow looks like this:

The agent maps "units" to the equipment record, "coming off warranty" to the warranty end date and "this quarter" to a date range.
It follows the link from the equipment record to the customer record and groups the list by customer.
For each customer it checks the latest service report and drafts a maintenance contract quote.
In the answer it shows which records it relied on; an authorised person approves or corrects the draft.

With definitions in place, the agent takes these steps in the same order every time; even when the question is phrased differently, the same concept leads to the same record.

How does the semantic layer work in PureField?

In PureField, the semantic layer sits where the data is born: the moment a record is created, its type, the meaning of its fields and its links to other records are known.

Records are born connected. Accepted quote lines become equipment records with serial numbers; work orders are opened from the equipment record, service reports link to the work order, and the invoice is issued from the completed service report.
Agents read from a record catalogue. Copilot and digital employees read from a catalogue that defines record types such as customer, equipment, quote, work order and invoice together with their field descriptions. When you ask about a customer, they follow the quote, equipment, service and invoice links through the relationships in that catalogue.
Copilot is briefed on your company's language. Copilot is given the context of your business: whether equipment means units installed at customer sites or your own plant inventory, what a work order means for you and which rules its answers follow. Copilot uses this profile in every answer.
Numbers rest on records. In Copilot's management meeting, the sales, finance and service directors look at the records in their own areas; every number in the answer rests on a record, and any amount without a source is flagged.
You set the write rules. Complete create and update plans are saved directly; deletes, missing fields and content read from outside sources go to approval.

Copilot and digital employees run on PAIOS (PureField AI Operating System). See what each agent does in which module on the AI agents page.

Evidence

Tableau Semantics: Salesforce's product page. It describes the semantic layer as the place where data is translated into business language, where models are defined once and used everywhere, and where Agentforce agents ground their answers. Source: tableau.com/products/tableau-semantics
ISG, Salesforce Adds Knowledge to Tableau for Agentic Analytics: Matt Aslett's analyst perspective of 3 June 2026. It writes that semantic modelling gives data its business meaning and provides the context BI and AI agents need. Source: research.isg-one.com

Try a Copilot that speaks the language of your records in a PureField demo. Book a demo or try it free for 7 days.

Frequently asked questions

Is a semantic layer the same as a data dictionary?

A data dictionary lists what fields mean. A semantic layer turns those definitions into measures and links between records, and queries are built from them. A data dictionary is a document you read; a semantic layer is the layer reports and agents run on.

What is the difference between a semantic layer and a company brain?

A semantic layer answers "what does this number mean, and what is this record linked to?". A company brain answers "what do we do in this situation, and who approves it?". An AI agent uses both: first it reads what the records mean, then the company's procedure.

Do you need a data team to build a semantic layer?

If the layer is built on top of a data warehouse, you usually need data engineers and analysts. If sales, service and finance run on one platform, record types and links come with the software, and company-specific terms and measures are added together with your managers.

Does a semantic layer stop AI from giving wrong answers?

It cuts the risk considerably: the agent reads terms from definitions instead of guessing, and takes the same path to the same question every time. Checking which record an answer rests on is still a good habit; in PureField, the numbers Copilot gives rest on records and any amount without a source is flagged.