Content and Authority for AI Answers

The Dashboard’s New Customer: How Semantics Are Reshaping Business Analytics

Graphwise’s Kent Stroker offers insights on the dashboard’s new customer and how semantics are reshaping business analytics. This article originally appeared in Insight Jam, an enterprise IT community that enables human conversation on AI.

For thirty years, business intelligence has found and formatted data into reports, charts, and dashboards. The finance analyst knew that “revenue” on the sales dashboard excluded returns. The regional manager knew which “customer” count to trust. The meaning is in people’s heads, tucked away in tribal knowledge. BI software vendors have realized the same thing: semantics are emerging as a new center of analytics. The dashboard, on the other hand, has a new user and it isn’t human. 

When an AI agent queries the data asking “What was net revenue in the Northwest last quarter?”, nobody is in between to supply that context. The agent discovers a column called revenue, creates a join, and confidently spits out an answer. If three tables define revenue three different ways, the agent has no way of knowing which one is actually needed. 

Large language models are convincingly confident,which is not the same as being factually grounded. Grounding requires both human and machine-readable definitions such as what things are, how they relate, and the rules that govern how they interact. This is where semantics fits in, and it’s why the analytics world is [re]discovering taxonomy and ontology. 

Three Layers Challenge 

“Semantic” has become a weaponized marketing term ,but really it revolves around three distinct aspects: 

Taxonomy aka the vocabulary. A taxonomy is a controlled vocabulary, arranged in a hierarchical order. It unifies what we call things, which synonyms mean the same thing, and how concepts nest. An “enterprise customer” is a kind of “customer,” and “client” and “account holder” are alternate labels for it. Standards such as SKOS make this machine-readable. Most BI users never fully build one, and the vocabulary is implied by table, column, and report names. 

Semantic layer aka the metrics. This is what Power BI calls a semantic model, and Tableau calls a semantic data model. This layer includes measures, dimensions, hierarchies, and join paths. It describes how data is shaped and calculated. It answers the question “How do I compute revenue from these tables?” 

Ontology aka the meaning. An ontology formally models concepts, relationships, and constraints. It informs both humans and machines what a customer “is” and well-known standards such as RDF, OWL, and SHACL provide the definition, classification, and validation layers. 

It is important to understand these differences For instance, a Power BI semantic model knows that the Sales table joins to the Customer table on CustomerID. It doesn’t know what a customer is. A star schema knows how tables join. But it can’t understand the business meaning. 

Current Directions and Various Approaches  

Microsoft is making Power BI semantic models a key part of Fabric IQ. Microsoft describes these in three layers OneLake for data, Power BI semantic models for business intelligence, and ontologies for what it calls operational intelligence. The ontology can be generated from an existing semantic model and exposed to agents. Many years of modeling work in Power BI has suddenly become AI infrastructure. 

Salesforce has introduced Tableau Semantics as a governed semantic layer. The query generator translates business questions from dashboards, apps, and agents into SQL, and Tableau Next uses it to ground Agentforce. 

The Interoperability Project, a group of vendors, has championed a different method. OSI is an open, vendor-neutral format for datasets, metrics, dimensions, and relationships. These are shared amongst the various AI and other tools. The project has since entered the Apache Incubator as Apache Ossie. 

What’s the Catch, and What Can You Do Right Now? 

First, most of what’s shipping is metrics, not meaning. A portable definition of “net margin” is valuable, but it isn’t an ontology. Standardized measures tell an agent how to calculate, but they don’t tell it what the things being calculated are or how they relate. The ontology features are the newest and least mature parts of every vendor’s stack. Unfortunately, the approach jumps straight to the ontology and lacks any tangible work for controlled vocabularies and taxonomies which is of concern. 

Second, beware of  the automation shortcut. Generating an ontology from a star schema is fast, but what you get is the schema drawn as a graph. The dimension tables become classes, foreign keys become relationships, and every wrong bit of data in the warehouse is now baked into the  “knowledge.” Applying real modeling informs questions is something a schema alone is incapable of answering. Those are business decisions that cannot, and should not, be left to a probabilistic, black-box ontology generator. 

Third, lock-in is moving up the stack. In addition to the realities of vendor lock-in for data repositories, it now includes where your definitions live. When meaning exists only inside one vendor’s semantic layer, switching is expensive. As a result, organizations should consider doing the following now: 

  • Inventory your definitions. Find every place “customer,” “revenue,” and “active” are defined such as DAX measures, Tableau calculations, SQL views, and spreadsheets. The disagreements are the real challenge. You do not need more technology; you need to foster agreement on terms across all your departments and data systems. 
  • Build the controlled vocabulary first. It is cheap, fast, readable by humans, and the foundation for everything above it and it can be done in a shared spreadsheet. This, in turn, drives the taxonomy creation. 
  • Treat semantics like code. Treat the semantic model like code by assigning ownership, running Scrum sprints, using CI/CD pipelines, and using Git-like versioning. 
  • Build the ontology thoughtfully. Start small with a single domain. Focus on where the agents are working. Yes, you can use LLM-generated drafts as a starting point but remember, LLMs are great editors and terrible authors. 
  • Stick to the standards. Use the existing, well-tested open standards like SKOS, OWL, OSI, so the semantics outlive any single tool, the next reorg and the next “really great idea”. 

Business intelligence spent the last three decades surfacing the data. The next evolution is about making it understandable to both people and machines. Investing in unglamorous work such as clean models and shared definitions will be the most strategic asset an organization can possess. Organizations who decide not to make that investment and instead reach for the Easy Button will soon discover that their agents are confidently, and fluently wrong. 

Share This

Related Posts

Solutions Review Expert Network Ad

Insight Jam Ad