What MIT got right about the semantic layer, and the part it left open.
When MIT's Center for Information Systems Research publishes a briefing on the semantic layer, a category that spent two decades as a footnote in BI documentation has officially arrived. The research is worth reading in full, and its central claim is one practitioners have been arriving at independently: the gating factor for getting value out of AI is not more data. It's meaning.
That conclusion deserves the institutional weight now behind it. It also leaves one question open, the question that decides whether a semantic layer actually changes outcomes or just becomes another artifact that drifts. The briefing defines the layer in terms of representation. Its own evidence points somewhere more demanding: operationalization.
What the research gets right
The most useful thing the briefing does is name the failure mode precisely. When data is pulled out of the application that created it, it loses the business meaning, the relationships, and the rules that were embedded at the source. Feed that decontextualized data to a generative model and you don't get an error. You get a confident, fluent, wrong answer. The model has no way to know that "active customer" excludes accounts in a 90-day grace period, or that two revenue figures can't be summed because one is gross and one is net. The context that would have caught the mistake was left behind in a system the model never sees.
This is the quiet danger of AI on ungoverned data. It rarely fails loudly. It fails plausibly.
The survey data makes the cost concrete. Only about one in five organizations rate their data curation practices as well-developed. The ones that do are roughly three times more likely to realize value from their AI initiatives and twice as likely to report a competitive advantage from them. Curation, the active and ongoing work of attaching meaning to data, is not hygiene. It is the differentiator. That is a finding worth quoting in any room where someone argues that the model is the hard part and the data will sort itself out.
Where the definition under-reaches
Here is the tension. The briefing describes the semantic layer as a system that maintains a consistent representation of data, sitting "conceptually between data and the user." Representation, conceptual position, between.
But look at the flagship example the research uses to prove the point. A data company onboarding product catalogs from thousands of suppliers built an engine that recognizes entities, applies ontologies and business rules, and maps messy inbound descriptions to standardized codes, automating roughly 80% of a process that used to be manual. That is not a representation sitting beside the data describing what it means. That is meaning executed inside the transformation itself. The same is true of the winning survey practice: what separates the leaders isn't having a semantic model, it's curating, the doing.
The headline definition is about describing. The evidence is about executing. And the gap between those two is exactly where most semantic layer efforts quietly fail.
A layer that sits "conceptually between data and the user" gets bypassed by every consumption path that doesn't route through it. This is the lesson the earliest semantic layers already taught us: meaning trapped inside one tool is meaning that the SQL, the notebooks, the downstream pipeline, and now the AI agents all ignore. They each re-derive their own version of "active customer," and the organization ends up with five definitions wearing one name. A semantic layer that isn't in the execution path isn't a source of truth. It's documentation, and documentation drifts.
Meaning is dynamic, not a dictionary entry
The research offers a small, perfect illustration without fully drawing it out: one supplier calls an item "Sterile Gauze Pad, 4x4" and another files the same thing as "SG-44-STR." The semantic work is reconciling those at load time, every time, as new suppliers and new conventions arrive.
That detail matters more than it looks. It tells you meaning is local, contested, and continuously renegotiated, not a fixed set of definitions you author once and freeze. Language in a real organization behaves like a working vocabulary that shifts as the business shifts: a new product line redraws "category," an acquisition introduces a second general ledger, a regulation changes what "compliant" requires. A semantic layer modeled as a static dictionary is obsolete the moment the business moves. The semantic work is never finished, which means it has to live somewhere that runs continuously, in the pipeline, not in a document that someone remembers to update.
Close the loop
The briefing's forward-looking step, use AI to generate metadata, is sound, with one caution practitioners should internalize. AI-generated metadata that isn't validated and routed back into executable transformation logic is just decontextualized description produced faster. More glossary entries that nobody enforces is not progress. The loop only closes when generated meaning becomes operational meaning: a definition the system applies, not a note the system stores.
From semantic layer to semantic intelligence
The shift the MIT research gestures toward, and that practitioners are already living, is a shift in the noun. A semantic layer represents meaning. What the AI era actually demands is a system that operationalizes and reasons over meaning, one that holds the definitions, applies the rules, reconciles the vocabularies, and does it consistently in the path the data travels, for humans and machines alike. That is the difference between a layer and intelligence.
This is the problem we built Coginiti to solve: a Semantic Intelligence Platform that makes meaning executable across SQL, BI, and AI agents from a single governed source, so that the definition you trust is the definition every consumer gets. MIT named the problem. Operationalizing the answer is the work.
If your AI initiatives are returning confident answers you can't fully trust, the gap is almost never the model. It's the meaning. We're happy to talk through where yours is leaking.
See Semantic Intelligence in Action
Coginiti operationalizes business meaning across your entire data estate.