Semantic Layer for AI
The Hopmann AI Foundation, Part 4
In short: The Semantic Layer for AI is a company’s dictionary. It defines in one central place what “revenue”, “conversion” or “customer lifetime value” actually mean, so that dashboards, ad-hoc queries and AI agents all compute the same number from the same raw data. For AI, this is not a nice-to-have. Without this shared language, the same agent returns a different answer to the same question every time it is asked, and nobody can tell which one is correct.
In the earlier posts of this series we showed why a solid measurement strategy and reliable data quality are the first two layers of the Hopmann AI Foundation. With those in place, it is clear what is being measured and that the underlying data can be trusted. What is missing is an intermediate layer that guarantees people and machines interpret the same numbers in the same way. That is what the Semantic Layer does.
Why a clean number is not yet an unambiguous answer
Ask any marketing team what a “lead” is and you will get at least three answers. Sales means a qualified contact with budget and need, performance marketing counts every form submission, the CRM flags a specific status change. Each definition is correct in its own context. For a single report, it can be negotiated.
An AI agent does not hesitate. It takes the first definition it finds, fluently formulates an explanation and puts a number on the table. Ask the same question ten minutes later in slightly different wording and a different number comes out, because the model followed a different interpretation. Both answers sound equally convincing, and neither is wrong in itself, they are just not the same.
A human analyst would ask back: which customers do you mean, active ones or everyone ever registered, with or without test accounts, net or gross? An agent does not do that on its own. It decides and delivers a plausible, cleanly formatted number. Once that number lands in a presentation, the choice of definition is no longer visible. Over time, discussions in the team get longer, not shorter, because every number first has to be checked against its implicit definition.
What a Semantic Layer actually is
A Semantic Layer sits between the raw data and everything that reads that data. It holds in one place what metrics and dimensions mean and how they are calculated. A handful of tables turn into a set of clearly named business objects: customer, order, campaign. A calculation rule turns into a metric with a formula, filters, granularity and time logic. A field name turns into a description that both a person and a language model can understand.
What makes it different today from three years ago is the dual audience. Traditionally the Semantic Layer was a BI convenience: it saved analysts the work of rebuilding a data model in every dashboard. Today AI agents read from the same definitions. The Semantic Layer has moved from a reporting convention to infrastructure for AI.
A Semantic Layer typically consists of these building blocks:
| BUILDING BLOCK | WHAT IT DELIVERS |
|---|---|
| Metrics | Defines formulas, filters and exclusions for KPIs like revenue, conversion or ROAS in one place. |
| Dimensions | Sets the attributes you can slice by, with clean data types and business-friendly names. |
| Entities | Maps customers, orders, products or campaigns to the physical tables that represent them. |
| Relationships | Describes joins, cardinalities and the grain of the tables so that analyses do not double-count. |
| Time logic | Fixes the fiscal calendar, the primary date field and the comparison logic for year-over-year or period-over-period values. |
| Descriptions and synonyms | Explain the meaning of a field to people and tell language models which words users actually use. |
The craft lies less in choosing the building blocks and more in the discipline of keeping every definition in only that one place. As soon as the same metric is maintained in parallel in an Excel formula, a Power BI measure and a CRM filter, the centralisation is diluted again, and the Semantic Layer loses the property that made it valuable.
Have the discussion once, not on every query
The real value of the Semantic Layer is not that it ends the debate about what an “active customer” means. It is that this debate happens exactly once and the answer then applies bindingly across all systems. What used to be a recurring argument in every reporting meeting becomes a one-off definitional decision, anchored in the data stack.
For AI this difference matters especially. A language model that freely decides each time how to compute a metric produces different answers to the same question and no basis for trusting any of them. A model that goes through the Semantic Layer does not recompute, it references a validated definition. That turns a generated answer into a reproducible one.
This also explains why a Semantic Layer delivers more than good documentation. A catalogue describes definitions, a Semantic Layer enforces them. A rule an agent only sees as a hint in the prompt is a rule it can reason its way around. A rule that is implemented as a metric in the Semantic Layer becomes the only technically available answer. The distinction sounds academic but decides in practice whether definitions are lived in day-to-day work or only exist in slides.
The payoff
The tangible effects of a well-maintained Semantic Layer show up in three places. Reports and dashboards deliver consistent numbers because they compute from the same source. New analyses are produced faster because the fundamentals do not have to be renegotiated with every ticket. And AI agents can, for the first time, reliably answer questions because they speak the same language as the people in the company.
Concretely: when an agent in a chat references [Umsatz], the number in the chat is the same one as in the dashboard. No newly invented calculation, no silent definition drift. For a team that used to spend an hour every week reconciling numbers between tools, that is a payoff felt every day, whether or not an agent is eventually put into production.
The payoff also scales with usage. A conventional report is opened perhaps ten times a month. An agent answers ten thousand questions in the same period. A definition captured correctly once therefore takes effect far more often than one that has to be guessed from context every time. The leverage of a clean central definition grows in direct proportion to how many AI applications go live.
What the Semantic Layer cannot do
Even the best central definition does not explain why the numbers look the way they do. A Semantic Layer holds the what, not the why. It knows how revenue is calculated, but not, for example, that the March figures are distorted by an IT migration. It describes the settled past, not the live present.
That is not a weakness, it is a deliberate assignment of roles. A Semantic Layer that also tries to carry the business context becomes unwieldy, slow, and almost impossible to maintain. That is exactly why the next layer of the Foundation exists.
What comes next
With the Semantic Layer in place, every system speaks the same language. What is missing is the frame in which those numbers are interpreted. A drop in revenue in December means something different in retail than in a subscription business, a strategy change in a campaign explains movements that would otherwise look like tracking errors, and an IT migration in March can devalue every comparison figure in the spring. That is the focus of the next post in this series: the Context Layer, the layer that puts the why alongside the Semantic Layer’s what.
Anyone who wants agentic AI in marketing to succeed does not start with the agent. They first make sure that everyone is talking about the same thing.
Want to find out what a Semantic Layer could look like for your data? Schedule a free 30-minute initial consultation with our team.
Follow Hopmann Marketing Analytics on LinkedIn for more insights on marketing analytics and AI.
Common questions we receive.
FAQ: Semantic Layer
What is a Semantic Layer in the context of AI and marketing?
A Semantic Layer is the central layer of a data stack that defines what metrics and dimensions mean and how they are calculated. Dashboards, ad-hoc queries and AI agents all draw on the same definitions. For AI applications this is the prerequisite for a question not producing a different answer every time it is repeated.
What is the difference between a Semantic Layer and a Data Warehouse?
A data warehouse stores the data itself and brings it into a structure suitable for analysis. The Semantic Layer sits on top of it and describes what the data means. Without a warehouse there is no reliable raw basis, without a Semantic Layer there is no consistent interpretation. Both layers address different questions and cannot be substituted for one another.
Why is good documentation not enough as a Semantic Layer?
Documentation describes definitions, a Semantic Layer enforces them. An agent that only sees a description in the prompt can deviate from it the moment a different interpretation seems more plausible. A metric implemented in the Semantic Layer, by contrast, becomes the only technically available answer. That distinction is what makes definitions reliable for AI.
How do the Semantic Layer and the Measurement Framework relate?
The Measurement Framework defines from a business perspective which KPIs a company steers by and how they are defined. The Semantic Layer implements those definitions technically and makes them binding in the data stack. Without a framework the Semantic Layer lacks the content; without a Semantic Layer the framework remains a document that every department can apply differently.
Does every company need a dedicated Semantic Layer?
Not necessarily as a standalone product. Most Power BI environments already contain a Semantic Layer in the form of their data models, dbt provides a variant with MetricFlow, and Snowflake, Databricks and other providers offer their own solutions. What matters is less the tool than the discipline of maintaining definitions in only one place and mirroring them from there into every consuming system.
How does Hopmann support companies in building a Semantic Layer?
We examine the existing data stack, clarify where definitions currently live and where they contradict one another, and decide together which layer will hold the metrics as binding going forward. The result is a robust Semantic Layer that is read identically by reporting tools and AI agents. That way the investments in measurement strategy and data quality become definitions that people and machines can equally rely on.