Logo - Keyrus
  • Playbook
  • Services
    Data advisory & consulting
    Data & analytics solutions
    Artificial Intelligence (AI)
    Enterprise Performance Management (EPM)
    Digital & multi-experience
  • Insights
  • Partners
  • Careers
  • About us
    What sets us apart
    Company purpose
    Innovation & Technologies
    Committed Keyrus
    Regulatory compliance
    Investors
    Management team
    Brands
    Locations
  • Contact UsJoin us
Expert opinion

9

The Missing Layer: Semantics, Ontologies, and Why Now?

Javier Riera, International Executive Director, Keyrus Iberia

The problem is not the data 

There's a scene I see more and more in conversations with customers. The company already has its AI pilot, the agent answers quickly and articulately, and at the first steering committee someone asks where the number of active customers comes from. Finance has one, sales another and the agent a third.

And it's not new. Who hasn't been to a meeting where marketing, finance, and sales each show up with a different revenue figure? Half an hour discussing which number is right instead of what to do with it. The origin is the same: no one has written down what each thing means. AI hasn't created the problem; it makes it more visible and scales it.

For years, people solved this: the analyst who knew which table to use and which exception not to touch. It worked because whoever consumed the data had context. An agent acting on its own has none of that context unless we provide it. 

What that layer is 

Between the data (ERP, CRM, data lake, contracts, emails, external sources) and the people who use it, there's a missing layer that says what each thing means and how it relates to the rest. Each vendor gives it a different name, but in the end there are three levels: 

  • Semantic layer: the definitions. Margin, active customer or churn, calculated once and served consistently to everyone. 

  • Ontology: the business model. What entities exist (customer, contract, policy, asset), how they are related, and what rules govern them. 

  • Knowledge graph: that ontology with real data inside, including unstructured data. The PDF contract linked to the customer record in the CRM and to the open ticket in the ticketing system. 

None of this is new for those of us who have been in data for years. What is new is the consumer. A Power BI semantic model or a BusinessObjects universe served a tool and people. This layer has to serve all tools and, above all, machines that reason and act. 

Where the market stands 

Ontologies have lived in academia and a few niche sectors for more than twenty years. What has changed is that now there is someone who really needs them, and in just over a year all the major vendors have moved in the same direction: 

  • Snowflake, Salesforce, dbt Labs, and RelationalAI launched the Open Semantic Interchange, an open standard for sharing semantics between tools, in September 2025. The first version came out in January 2026, with Databricks and Qlik among those joining, and in July 2026 the project moved to the Apache Software Foundation as Apache Ossie (incubating).  

  • Microsoft introduced Fabric IQ at the end of 2025: an ontology that lets agents understand business entities, not just tables. 

  • Databricks builds it with Unity Catalog metric views. Google grounds its Looker agents in the Looker semantic layer, and goes further with vector search in BigQuery, the Agent Development Kit (ADK) and the Agent2Agent (A2A) protocol for building agents and letting them work together. 

  • Glean comes at it from unstructured data: its knowledge graph is built from documents and activity, respecting the permissions of each user. 

  • Palantir has been selling exactly this under a different name for years, and Veezoo (a Keyrus partner) was doing it before it became a trend: its Knowledge Graph is what keeps definitions, permissions, and rules consistent behind each answer. 

Analysts agree. Gartner published in May that the lack of semantics causes inaccurate agents and wasted spending, and predicts that by 2027 those who prioritize it will improve the accuracy of their agents by up to 80% and cut costs by up to 60%. In September, BARC dedicated an entire study to context engineering (with support from DataHub). 

My reading: there is consensus on the what and none on the how. Each platform has its version, the open standard is only now maturing as an Apache project, and portability between platforms doesn't yet exist in practice. 

How I'd approach it 

Many data governance programs, business glossaries, and MDMs of the last decade got stuck for the same reason: wanting to define the entire company before delivering anything. I would do the opposite: 

  1. Start with a decision, not a domain. Which customers are at risk and why, which assets need maintenance this week. The decision determines which entities and metrics matter. 

  2. Define 20 concepts, not 2,000. The ones that actually get debated in steering committees, each with a business owner. AI can propose definitions; a person decides. 

  3. Bring in unstructured data from day one. Contracts, emails, tickets. That's the context that dashboards never had. 

  4. Put it in front of a real agent. Expose it through MCP or the semantic layer's APIs, so the agent queries the definitions instead of guessing them. If the answers don't improve with the layer, the layer isn't done well. There's no better test. 

  5. Version it and stay vendor-neutral. Definitions with a history of changes, in a format that can be taken to another platform. 

The risks (and there are some) 

  • Semantic lock-in. Migrating data is expensive; migrating the meaning of the business is even more so. If the ontology lives inside a single platform, changing platforms means rewriting it. 

  • Treating it as an IT project. Definitions are business decisions. If they're made only by the data team, no one will use them. 

  • Nobody owns maintenance. The initial build is the easy part. The business changes (products, pricing, sales structure, acquisitions) and a layer that is not updated ages in silence: the agent keeps answering confidently about a reality that no longer exists. Before starting, you have to be clear about who owns each definition, how a change is approved, and what triggers a review. If the answer is "the project", there is no answer: projects end. 

  • Stopping at definitions. Knowing what an active customer is is not enough. The agent also needs the rules: when an exception is approved, who can do what. 

  • Beautiful semantics on bad data. The layer doesn't fix quality, it makes it more visible. And an agent with good context and bad data gets it wrong with more confidence. 

  • Name hype. Semantic layer, ontology, context layer, knowledge graph. It's worth asking what each thing does, not what it is called. 

Now or wait? 

I understand those who'd rather wait. The standard is not mature, platforms change every quarter, and no one wants to build twice. For the technology, fair enough. For the work, no. Agreeing on what an active customer is, what a contract at risk is, or how the real margin is calculated does not depend on any standard. It is the slow part, the one that forces business, finance, and data to sit at the same table, and the one that any future tool will need. Platforms are already starting to generate ontologies automatically; what none of them will do for you is decide which of the three figures is the right one. My recommendation: start now, start small and avoid lock-in. A use case with value, portable definitions and an agent that uses them from the first month. Whoever does it in the next 12 months will have an AI that understands their business. Whoever waits will have the same platform as everyone else and the same problem as always. We are all going to have access to the same models. The difference will be in the context that only your company can give them. 

How are you approaching it? 

Contact Our Experts
Continue reading
  • Success story

    EcoVadis Platinum: Recognition of Keyrus’ ESG Maturity

  • Blog post

    Being AI-Ready: Data, Governance, and Agentic Analytics to Successfully Scale AI

  • White Paper

    Democratize access to business information with Qlik Sense Cloud

  • Blog post

    30 years of Keyrus: Turning data into intelligence

  • Expert opinion

    Snowflake Cortex AI Explained: Capabilities, Pricing, and Real-World Use Cases

Logo - Keyrus
Lisbon

Avenida Defensores de Chaves,nº 4, 4º Andar 1000-117 Lisboa

Phone:+351 913 083 380