The productivity gains from AI are no longer in question. Emails get polished in seconds, code gets drafted and commented, presentations assemble themselves, and a quick question to an assistant replaces a meeting with a subject matter expert. The open question is what happens when those same probabilistic systems are asked to inform operational decisions, because a model that confidently misstates the state of the business carries a much higher price than a clumsy email. The answer, increasingly, is being built between the raw data and the model: a governed semantic layer that carries the business meaning AI cannot infer on its own.
Bertin Nono is a data engineering manager in decision science, with more than 15 years architecting cloud data platforms, real-time pipelines, and ML-enabled systems across industries, and he also advises founders as a fractional chief data and AI officer. He spoke in a personal capacity, and the views here are his own rather than his employer's. His starting point is blunt about where the leverage sits.
“At the end of the day, the data has to be right, properly curated, available near real time, and in the right format. Data engineering without question has a pivotal role in this AI revolution,” says Nono
Nono draws a clear distinction between what large language models do well and what enterprises actually need from them. Drafting, summarizing, and assisting are largely solved problems. Computing the business is not.
“These LLMs are not mathematical systems. They’re not really made to do calculus or compute business logic.”
For a CIO, the implication is straightforward: enterprise decision-making depends on deterministic calculations, consistent KPI definitions, and traceable lineage. These are not properties of probabilistic systems.
This is why generative business intelligence (BI) tools often struggle in production environments. The issue is not model capability - it is missing or inconsistent business context.
Nono frames the shift this way:
“We’re not just building pipelines to feed dashboards. We’re building data as a product, with a roadmap and a business outcome.”
That productization is the foundation decision intelligence depends on, but it still leaves a critical gap: shared meaning across the enterprise.
Nono’s answer to this gap is ontology—a graph-based semantic model that defines how business entities are structured and how they relate to one another. At its core, ontology establishes shared meaning across systems by clearly defining entities, attributes, and relationships. He compares it to a library index: the system that tells you exactly where information lives and how it is organized.
In practice, ontologies already power many real-world systems. Healthcare uses them to standardize diseases and treatments, e‑commerce platforms use them to connect products, customers, and transactions, and search engines rely on them to link entities and interpret intent. Within the enterprise, the same principle applies: concepts like Customer, Flight, Revenue, or Booking are defined once and reused consistently across systems.
His clearest example is an airline. “One way to define an ontology is a flight. A flight has crew, passengers, and food. The crew includes pilots and flight attendants. A pilot has qualifications and may operate specific fleets.” An AI system built on this graph understands what “tell me everything about the flight” actually entails. That shared structure is why ontologies measurably reduce hallucination and enable systems across an enterprise to operate with a common language.
He credits Palantir with recognizing this early, while pointing out the key challenge: building ontologies remains largely manual work. The automation of ontology creation, in his view, is the critical missing capability that would accelerate adoption—aligning with broader efforts to establish semantics as the foundation for BI and AI. Architecturally, he is clear: the semantic layer belongs in a graph store, not flattened into vectors, because large language models interpret structured relationships more effectively when business context is preserved.
A semantic layer is not a one-time build. Regulations shift, policies change, and a KPI definition from six months ago may quietly stop being true. Nono treats this as the decisive challenge, and it is organizational before it is technical.
"You can't just build the ontology and let it run by itself. It has to be governed in a way that when things change, it keeps changing too," he says. Versioning is the easy part, since a graph node can carry its version as an attribute. The hard part is culture: a centralized governance board with stakeholders from every business unit, syncing as definitions move, backed by investment from the top.
He sees a labor market reshaping around exactly this work, with traditional BI roles giving way to jobs governing semantic layers, labeling data, and validating generative BI outputs, the kind of discipline that keeps AI data debt from compounding. Pushed on where this ends up, he endorses the idea of semantic transactions, where ontology changes, KPI definitions, lineage, and AI context are governed together as one operational system. "It becomes a transactional system sitting on top of a graph database."
The payoff only materializes at scale. If each business unit defines its own semantics, fragmentation persists, and every AI system inherits that misalignment.
“For a company to run well, every single business unit has to do its piece. When you want to build an AI system, it has to have a full understanding, because everything connects.”
For CIOs, the takeaway is clear. The competitive advantage in enterprise AI is shifting:
Not toward better models, but toward better-defined, better-governed business meaning.
Organizations that treat semantics as infrastructure will build AI systems that are trusted, consistent, and scalable. Those that do not will continue to struggle - not with capability, but with alignment.
The views and opinions expressed are those of Bertin Nono and do not represent the official policy or position of any organization.
