Three Departments, Three Answers to the Same Question
Ask "what was last month's revenue?" and it is surprisingly common to get three different answers. Sales reports KRW 1.2 billion in gross orders by order date. Finance reports KRW 1.144 billion after excluding 3% in cancellations and refunds (KRW 36 million) plus KRW 20 million in intercompany transactions. Marketing reports KRW 1.23 billion, having shifted late-previous-month orders into the current month on a settlement-date basis. None of the three calculations is wrong. What is wrong is that no one ever agreed on which of them is "revenue."
When multiple analytics tools each embed their own metric logic without a shared language or standard, these discrepancies accumulate by design. The result is a paradox: the more dashboards you build, the slower decisions get, because meeting time shifts to reconciling numbers. In an organization with 40 dashboards where "active customer" is defined slightly differently in each tool, those 40 screens are 40 competing claims.
The root cause is not the capability of your BI tool. It is that the definitions of "revenue" and "active customer" exist nowhere — neither in a document nor in code.
What a Semantic Layer Actually Does
A semantic layer is where business logic is defined once and reused across the organization. The formula for revenue, its exclusion rules, and its date basis are declared in one place, and BI screens, reports, APIs, and natural-language queries all reference that declaration. One definition yields one number.
The trend is visible in the platforms themselves: Snowflake's semantic views and Databricks' business semantics both bring a semantic layer into the data platform, and dedicated semantic layer products continue to attract investment. Metric definition is starting to be treated as an infrastructure concern rather than a reporting detail.
Putting a formula in a BI tool is not the same as defining it in a semantic layer. The former is valid only inside that tool, and duplicating a screen duplicates the definition. The latter lives outside the tool, so it can be version-controlled, code-reviewed, and tested — and it survives a tool migration.
What Belongs in a Metric Definition
A definition does not need to be a long specification document. One table per metric is enough. If any of the following is blank, that metric is not yet defined.
An even more important operating principle is to add dimensions, not metrics. Creating "mobile new customers" and "app new customers" as separate metrics splits the definition. Keeping a single "new customers" metric with channel, device, and region dimensions means 10 metrics and 8 dimensions can answer 80 different questions.
How Much Data Plumbing Do You Actually Need
Pointing reports directly at the production database is the fastest start, but it leads to incidents where heavy month-end aggregation queries run concurrently and slow down the live service. It also means the data shifts between queries, so opening the same screen twice gives different numbers.
A small organization can begin with a read replica and a few aggregate tables, no warehouse required. These signals, however, indicate it is time to separate an analytical store:
Load frequency should follow your decision cycle, not a technical preference. A once-daily load is fine for metrics reviewed in a weekly meeting, but if someone must act on today's inventory or ad spend, a figure up to 24 hours stale is unusable. Always display an "as of" timestamp on the screen so freshness is explicit.
What Has to Exist Before You Ask the AI
When a generative AI agent takes questions about internal data, the semantic layer's job is to give it curated knowledge and trusted context to work from. With definitions declared, the model retrieves the formula instead of inventing one — which reduces hallucination and makes the same question return the same answer.
Attach natural-language querying to an organization without definitions and you get the opposite: wrong numbers spreading faster and more convincingly. A human error gets caught in a meeting; an instant answer from an AI gets quoted without verification.
Permissions must travel down with it. If metric-level controls (labor cost metrics restricted above a certain grade) and row-level controls (a branch manager sees only their branch) are not enforced in the semantic layer, natural-language querying becomes a path around access control.
A Minimum Viable Setup in Three Months
Trying to complete company-wide data governance first means never starting. Narrowing the sequence is more realistic.
POLYGLOTSOFT builds internal systems and data integrations across ERP, payments, and CRM, and we handle the full path from writing metric definitions to implementing the calculation layer and the reporting screens. Because metric definitions are never finished — they change whenever the business changes — we recommend our subscription development model, which carries definition updates and screen changes forward month after month. If your meetings are currently spent reconciling numbers, send us the list of reports you use today and we will map out which metrics to define first.