Semantic layer: unified definition of metrics for both humans and AI
The semantic layer is the part of the data warehouse that once and for all defines metrics and dimensions — business name, description, calculation, and relationships between data. Thanks to it, AI and humans calculate sales, margin or active customers in the same way.
What it contains
The semantic layer is the part of the data warehouse where the agreed definitions of metrics and dimensions live. Not in a table on disk and not in an analyst's head — right in the data model, where humans and AI read it.
In addition to the names and descriptions, mainly what is otherwise imagined:
- The name and description of the metric as named by the shop, not the technician.
- Calculation: with or without VAT, with or without transport, before returns or after deduction, according to orders or invoices, including unpaid or only paid.
- Links between data: which product belongs to which category, which orders came from which channel.
Example: how many versions of the truth does one sale have
For one online store sale, there are five decisions: VAT, shipping, returns, data source and whether to count outstanding payments as well. The switches below correspond to the demo at the beginning — same data, same decision. Switch and watch the "revenue" change.
At the same time, each variant is the correct answer to a different definition — that is why the definition must be agreed upon within the company.
What was our revenue in September?
4,850,000 CZK
Difference between highest and lowest version: 1,394,000 CZK
Illustrative data of a fictitious online store.
The semantic layer makes this decision once, permanently and for all. Each AI then answers the same.
Why is conversation analytics on it?
AI chat without a semantic layer makes up its own definitions. About the export, she answers confidently and with a specific number, but a number according to the definition she chose herself. Above the semantic layer, it will respond with the number you agreed on and tell how it came about.
The sales version demo at the beginning shows how one question gives five decisions and 32 different numbers. AI no longer thinks about the semantic layer: every definition is written and every answer can be explained.
Passersby will respond equally to you and your competition. Only you have your own analyst.
Management determines the definition, not the technician
What is a margin or an active customer is a management decision, not a technical detail. It takes time — and it's better to get it done before the AI starts responding.
No one knows what AI will be the best in a year. But each will need your context — and its core is the semantic layer.
And since the definitions stay with the company, they will also be carried over to the new tool. Whoever has the context ready will switch the AI tool within days.
Analyst knowledge, written once
Do you have your own analyst? The semantic layer is your analyst's knowledge, written once. Everyone gets answers according to his definitions — and it stops answering the same question for the fifth time.
Likewise, a new colleague: it doesn't read attachments in an email, but it asks — and gets an answer according to the same definitions as management.
That's why we build the semantic layer together with the warehouse: a warehouse without agreed definitions is just as inaccurate as an export.
Knowledge stays with the company, not the tool. Every new AI tool becomes your analyst right away.
Frequently asked questions
Won't AI soon be so smart that it can handle our data by itself?
Even the economics professor you meet on the street doesn't know your sales. Smarter models think better, but intelligence is not information: no model knows that your margin does not include transport unless it is written somewhere.
We have our own analyst. What is it to us?
The semantic layer is your analyst's knowledge, written once. Everyone gets answers according to his definitions — and it stops answering the same question for the fifth time.
We already have BI and dashboards. Do we need it?
Then you are half done. Dashboards answer pre-given questions, AI chats and those that no one expected — but only above the semantic layer. Your BI model is the starting point.
How do we know the AI is answering correctly?
Before handover, we test the AI on a set of real business questions with known answers and tune the semantic layer until it passes the test. You will see the results.
More on the topic
Make AI chat your analyst
Free and non-binding online consultation. We'll go through your systems and show you where to start.
Free consultation