Gartner released its Top Strategic Technology Trends for 2027 last week. One of them will look familiar to anyone who has spent the last two years arguing about semantic layers: AI Context Platforms.
Gartner describes these as the systems that give AI agents the contextual knowledge they need to interpret inputs, make decisions and act in line with business objectives. Gene Alvarez, Distinguished VP Analyst at Gartner, put the risk plainly: "Without a dedicated context layer, organizations face severe operational risks, including overprivileged agents, compliance failures, low trust and excessive token consumption."
The number that will get quoted is Gartner's prediction that by 2027, "organizations that prioritize semantics in AI-ready data will increase agentic AI accuracy by up to 80%," and reduce costs by up to 60%.
That sentence doesn't reward buying a context platform or installing a semantic layer. It rewards organizations that prioritize semantics. That is a statement about how you treat meaning, not about which product holds it.
It's also the second time this year Gartner has made the point. In March, at its Data & Analytics Summit in Orlando, Gartner predicted that "by 2030, universal semantic layers will be treated as critical infrastructure, alongside data platforms and cybersecurity," and said developing one "is now a must-do for D&A leaders either leading or supporting AI." Seven months later, the same idea has moved off the data and analytics agenda and onto Gartner's list of strategic technology trends for the whole enterprise.
Context is not a document
The industry already has a habit with context: write more of it. Prompt instructions, README files for agents, a markdown file of business definitions dropped into the context window. It works in a demo. It even works in production for a while.
The trouble is that context written this way has no owner, no version, no test and no review. Nobody can tell you whether the definition of "active customer" in an agent's instructions matches the one Finance approved last quarter. When it drifts, nothing fails. The agent keeps answering with confidence.
We made this point in When the Model Eats the Stack: generating context is getting cheap, and governing it is not. The Berkeley and Stanford study of data agents behind that post found that for the strongest model tested, more than 60% of failures were semantic or environmental: the wrong metric, the wrong source, a misread concept, a misused join key. Better models didn't remove that category of failure. It became the majority.
Prioritizing semantics means treating meaning the way you treat any other production system. Definitions are:
- Structured, so an agent queries them instead of paraphrasing them
- Tested, so a metric that drifts outside its expected range fails loudly
- Versioned, so you can say which definition was in production when an answer was given
- Reviewed, so a change to revenue gets a second pair of eyes before an agent uses it
- Traceable, so you can see which sources and transformations a number came from
Where the 80% and the 60% come from
Most wrong answers from data agents aren't hallucinations in the usual sense. The SQL runs. The number looks plausible. It is still wrong: an order-level amount summed across order lines, a definition from three versions ago, a filter that excludes refunds in one report and not another.
An agent that goes through a governed semantic layer doesn't have to make those decisions. It asks for revenue. The layer knows the grain, the joins and the approved definition, and it catches grain errors like fan traps and chasm traps instead of returning a silently inflated number. That is where accuracy comes from: fewer decisions left to the model.
The cost side runs on the same mechanism — note that "excessive token consumption" made Gartner's own list of risks. An agent pointed at raw tables explores. It lists schemas, samples tables, guesses at joins and retries. That bills twice, once in tokens and once in warehouse compute. Customers who route agents through Coginiti's semantic layer have seen token spend fall by as much as 70%, and the size of the saving tracks how messy the source schema was to begin with. We covered the economics in Token Costs Now Rival Developer Salaries.
Gartner's 60% and our 70% come from different measurements, but they point at the same mechanism. An agent that knows what things mean doesn't have to work it out on every request.
The context layer is the waterline
A few weeks ago we wrote that the semantic layer is only the tip of the iceberg. Above the waterline is everything people see: agents, copilots, natural language analytics, dashboards, APIs. At the waterline sit metrics, entities, relationships and governed definitions. Below it is the work that makes those definitions worth trusting.
A context platform has the same shape. In his Big Data LDN talk, The Context Stack, our CEO Richard Hall separates three things that get sold as interchangeable. Semantic models define business meaning. Ontologies capture relationships rich enough to reason across. Context engineering delivers the right information to the AI at the moment it's needed. Gartner's trend covers all three, and none of them is trustworthy on its own. Each one inherits its quality from what sits underneath.
| Below the waterline | Includes | What it gives an agent's context |
|---|---|---|
| Produce | Data modeling, transformation, analytics engineering | Clean, well-understood structures for definitions to sit on |
| Validate | Data quality tests, tests on semantic definitions, lineage | Evidence that a definition still returns what it should |
| Govern | Versioning, review, access control, audit | A record of who changed meaning and who approved it |
| Maintain | Scheduling, deployment, monitoring | Confidence that the context in production is the context you reviewed |
The risk with a new category name is that the market builds up from it. We watched it happen with semantic layers. Vendors invested in natural language interfaces, agents, visualization and APIs, all above the waterline, and assumed transformation, testing, lineage and deployment happened somewhere else. Context platforms will face the same pull. Retrieval, routing and memory are visible, and they demo well. Ownership, testing and versioning of what gets retrieved don't.
Alvarez's warning about overprivileged agents is an access problem, and it's real. Next to it sits a quieter one: the under-informed agent, fully authorized, reading context nobody has checked. A query can be authorized and still be wrong.
Six questions for any context platform
If AI Context Platforms are on your 2027 plan, start with the questions an auditor would ask about any answer an agent gives:
- Is this the approved definition?
- What transformations produced it, and were they tested?
- What source systems contributed?
- Who changed the logic, and was the change reviewed?
- Which version is in production?
- What other metrics depend on it?
A platform that can't answer these is handing agents the same guesses with better formatting.
Gartner is right that context is becoming infrastructure. Infrastructure earns trust the unglamorous way: tests, versions, reviews, deployments and monitoring. That is the part of the iceberg Coginiti was built for. Our Semantic Intelligence Platform puts the semantic layer in the same platform as the work beneath it, producing, validating, governing and maintaining business meaning, so the context your agents receive is context you can stand behind.
See Semantic Intelligence in Action
Coginiti operationalizes business meaning across your entire data estate.