If you’re exploring conversational AI solutions, here’s a simple test: ask the same question five times, in five sessions, and see if you get the same answer five times. I do this regularly as part of my research. Too often, I get different answers, and once in a while, data that should never have been returned.
Ask why that happens, and you get a vocabulary lesson. If your algorithm is anything like mine, you’re being fed a constant stream of definitions for semantics, ontologies, knowledge graphs, and context. What they are, how they differ, and which fits inside which.
I get it…the terms are confusing, but if you have to explain a product category before someone can decide whether they need it, you’re starting in the wrong place.
The more useful question for a data leader isn’t “What is a context layer?” It’s “What problem are you trying to solve?”
Vendor diagrams follow the product. Semantic vendors put semantics at the center. Metadata and catalog vendors have evolved from inventory to context, placing it at the center. The ontology folks, who have been preaching formal knowledge representation for decades, draw the most rigorous pictures of all, are cited by everyone, and are implemented by almost no one.
Everyone draws the biggest box around the thing they build.
Inconsistent meanings. And here I’ll pause for a public service message.
Sufficiently confusing alternatives to the requisite “Revenue” example:
A person doesn’t have much trouble figuring out which meaning applies to which situations…machines need instructions.
In EMA Data and Analytics Research (September 2026), more than 60% of organizations reported that missing, incomplete, or inconsistent context has caused an AI model or agent to misinterpret data or produce an inappropriate result at least occasionally.
Your humble data analyst has been reliably filling in the gaps for years. If two dashboards disagree, they catch it. They know what means what to whom, which numbers management needs, and not to expose an SSN just because the data is there.
AI systems don’t know any of that unless we tell them.
AI didn’t create the problem; it removed the human who had been compensating for it.
If two departments disagree about what a word means, no architecture fixes that. Someone still has to decide, and then determine where that meaning applies, how it’s calculated, who owns it, etc.
Technology can help find the conflicts. Increasingly, AI can even propose definitions and relationships. And our data reflects that dynamic: 76% say business context is authored by people either on their own (22%) or in combination with automated methods1. But at some point, someone has to decide which meaning is authoritative.
Establishing meaning is fundamental. Buying a semantic-layer product is optional. Relationships are fundamental. A knowledge-graph product is optional. Governance is fundamental. A standalone governance platform is optional.
Governance alone isn’t enough. A governed answer can still be wrong. Governance can tell you the source was approved, the AI system had access, and the action was logged. It can’t tell you the answer was correct.
Don’t start by shopping for a context platform. Agree on definitions, valid joins, and business rules. You need semantics.
Maybe they live in Power BI, Snowflake, Salesforce, catalogs, or application code. Now it’s about portability.
Semantic-layer products such as AtScale, Cube, dbt’s Semantic Layer, or Looker can be part of the answer, particularly if they already hold some of those definitions. But don’t recreate semantics in the next platform. Ask the vendor to show you how the product consumes what you’ve already built.
Don’t manufacture an architecture problem. If Databricks, Microsoft Fabric, Snowflake, or another primary platform can provide the semantics, governance, and context you need, use it. You don’t get extra credit for another tool.
A fragmented estate presents two problems that need different answers. Reaching the data is a federation problem, but definitions still have to be settled. Platforms such as Denodo, Dremio, and Starburst can help query across distributed sources without first moving everything into one place, and are increasingly adding their own semantic, governance, and context capabilities.
An alternative is a platform that brings semantics together with metadata, lineage, governance, and relationships across systems. Alation, Atlan, DataHub, and OpenMetadata/Collate are examples of that convergence, among a growing number of vendors heading in the same direction.
Either way, ask which data and which definitions reach which consumers. Federation without meaning just gets you inconsistent numbers faster.
Semantics alone is not enough. An agentic system needs to know what things mean, how they relate, who is asking, what they’re allowed to see, where the information came from, what rules apply, what you’re trying to accomplish, and what is true right now.
If those relationships and rules exist only in people’s heads, dashboard logic, or application code, the agent can’t reliably use them. Make the important ones explicit and machine-readable.
That may mean extending a semantic model. It may mean adding an ontology or knowledge graph. It may mean pulling identity, policy, lineage, and current state into the context the agent receives.
Estate shape and maturity determine architecture. Neither changes the foundations.
Not every business needs to centralize all context before it can put AI to work. Analytics platforms such as Hex, Sigma, and ThoughtSpot are increasingly bringing semantic models, business definitions, metadata, permissions, and other context into the environment where the users ask questions.
If the problem is helping an analyst use AI to interpret data correctly, assembling the necessary context at the point of consumption may be enough.
The limitation is scope. Context that works well inside an analytics application may not automatically govern an AI system operating across your estate.
We should all still care about the distinctions between semantics, ontology, knowledge graphs, and context. They describe different jobs.
But buyers don’t need another definition before they can act.
The vendors will keep putting more capabilities into broader platforms. Some will be very good.
Convergence can eliminate layers, but it can’t eliminate the work those layers were created to do.