Despite $30–40 billion in enterprise investment into GenAI, this report uncovers a surprising result in that 95% of organizations are getting zero return.

–MIT NANDA

In 2025, a sobering study from MIT’s Project NANDA, The GenAI Divide: State of AI in Business 2025, revealed that approximately 95% of enterprise generative AI pilots fail to deliver measurable business value or scale beyond experimentation. Rather than indicting the capabilities of generative models themselves, the research points to deeper organizational issues: flawed integration into workflows, misaligned priorities, lack of measurable outcomes, and poor alignment with core business processes as the principal culprits behind this high failure rate.

The study finds that the vast majority of pilots yield little to no profit-and-loss impact because companies often treat AI tools as bolt-ons or novelty features instead of embedding them meaningfully into systemic workflows and strategic initiatives. Leaders tend to apply generative AI to broad, ill-defined problems, prioritize function over value, and underestimate the complexities of integrating AI with existing systems and processes. This clearly leads to initiatives that stall, deliver superficial results, or never move beyond pilot mode.

This gap between technical capability and business impact suggests that the challenge isn’t a lack of powerful models, but a lack of context, alignment, and semantic coherence. Successful enterprise AI isn’t merely about deploying a sophisticated model; it’s about ensuring the AI understands the true meaning of business concepts, data relationships, and processes in the context of a specific organization. Without that foundation, even the most advanced generative tools struggle to produce trustworthy, accurate, or actionable results.

Here, we posit that ontologies — structured, evolving semantic models that codify how an organization conceptualizes its data and processes — serve as the critical semantic substrate that bridges this divide. When organizations build a semantic layer that explicitly maps out concepts, relationships, and meaning, they don’t just prepare data for AI, but they prepare AI for meaning. A thoughtful ontology exposes where high-impact questions reside, where AI can add genuine value (such as automating back-office tasks or answering complex enterprise questions), and where ambiguity or data disconnects are likely to derail pilot efforts.

Moreover, developing such a semantic fabric is itself a forcing function for solving the very integration and alignment challenges the MIT research calls out. It surfaces domain meaning, harmonizes disparate data sources, connects business logic to technical artifacts, and highlights dependencies that typically get missed in early AI initiatives. By investing in a semantic layer early, organizations not only streamline pilots, but they accelerate ROI, reduce friction across data domains, and lower barriers for future AI use cases by providing a common, machine-interpretable source of truth that underpins both human and AI understanding.

Before we dive into how ontologies supercharge generative AI, it’s useful to map out the layers that make up an organization’s data ecosystem; from the raw stuff machines store all the way up to the meanings and relationships humans think about.

Think of these layers as a journey from “what data is” to “what data means,” and why the latter truly matters when you start talking to AI.

At the base of everything is the physical layer. This is the gritty, real-world stuff:

• Tables and columns in relational databases.
• Files in JSON, XML, Parquet, or other formats.
• Unstructured text like documents, emails, transcripts, and logs.

It’s where your data resides and gets stored. But in this form? It’s like having a library full of books with no catalog; rich with data points, but hard to navigate.

This layer alone doesn’t tell us much about meaning; it’s purely about structure and storage. And in most organizations, it isn’t a single, unified thing at all. It’s a patchwork of data silos: HR systems, CRM platforms, finance and payroll tools, ticketing systems, document repositories, and bespoke line-of-business databases; each optimized for its own purpose, each with its own schema, terminology, and assumptions.

Above that we find the logical layer. This is not a physical thing you can touch, but more like the mental map the business uses to talk about its data:

• “What do we mean by customer?”
• “How do we define active subscription?”
• “What counts as revenue recognized?”

These ideas show up in reports, dashboards, and business glossaries. Teams verbally agree on these definitions during meetings and strategy sessions. They become part of how the organization conceptualizes its world.

But here’s the catch: those definitions often live in people’s heads, in spreadsheets, or in tribal knowledge. They’re rarely implemented in a way that systems, analytics tools, or AI models can use consistently. Over time, teams develop their own dialects; using the same words to describe subtly different concepts or inventing different terminology for what is effectively the same idea. These inconsistencies are usually manageable for humans, who rely on context and experience to bridge the gaps, but they become a serious obstacle when organizations try to scale analytics or introduce AI into the mix.

You can visualize the logical layer, talk about it, even build dashboards to reflect it — but until it’s made machine-interpretable, it’s not truly actionable at scale.

This is where things get interesting.

The semantic layer is the practical, machine-ready implementation of the logical layer.  This is the part that makes meaning usable across your organization. It’s what turns abstract business concepts into something both humans and AI systems can operate against with confidence.

A semantic layer (defined all over the place: AtScale, Atlan, DataCamp, DataGalaxy):

• Translates raw, messy data into familiar business terms and concepts.
• Connects and harmonizes definitions across systems so everyone (and every tool) can agree on what things mean.
• Enforces consistent business logic, relationships, and rules so that “revenue” means the same thing in a dashboard as it does in an AI query.

You can think of it like the Rosetta Stone for enterprise data; bridging the gap between technical storage and meaningful interpretation.

Importantly, the semantic layer doesn’t displace the logical ideas people talk about — it instantiates them in a form machines can use. Whereas the logical layer is the abstract conversation, the semantic layer is the machine executable version of that conversation.

In short: the physical layer is the data, the logical layer is the intent, and the semantic layer makes that intent real for machines. When ontologies drive the semantic layer, they ensure that AI doesn’t just retrieve data, it understands and distills meaning.

Generative AI is remarkably good at language. It can summarize, explain, rephrase, and reason across text with impressive fluency. But out of the box, LLMs don’t actually understand your organization; not its data, not its structure, and not the way its concepts relate to one another.

That gap isn’t a flaw in the models themselves. It’s a consequence of how they’re trained. Large language models learn patterns from vast amounts of public and semi-public text. They know how words tend to be used, but they don’t know what your organization means by customer, case, asset, or risk, nor how those ideas connect across systems. This is exactly where Ontologies come in!

Ontologies provide the grounded semantic context that generative AI lacks by default. They explicitly model the real-world entities an organization cares about, the relationships between them, and the rules that govern how those entities behave. Instead of forcing an AI model to infer meaning from raw tables, documents, or loosely connected APIs, an ontology gives it a structured map of how the organization understands its world.  In practical terms, that means an AI system no longer has to guess. It can navigate.

Because ontologies implement the organization’s logical layer in a machine-interpretable way, they enable consistent interpretation across teams, tools, and use cases. The same concept means the same thing whether it’s referenced in a dashboard, a workflow, or a natural-language question posed to an AI assistant. This foundation unlocks several critical capabilities for generative AI:

• Semantic grounding of responses
  AI outputs can be tied directly to real, well-defined entities and relationships, not just plausible sounding text. Answers are anchored in organization reality rather than probabilistic guesswork.
• Disambiguation and context refinement
  When the same term can mean different things in different contexts, ontologies make the distinction explicit. This dramatically improves relevance and helps reduce hallucinations caused by ambiguity.
• Cross-domain understanding and reasoning
  Many of the hardest and most valuable enterprise questions span data silos. Finance, Operations, HR, Sales, Compliance, etc.… Ontologies provide the connective tissue that allows AI to reliably reason across domains.

Seen this way, ontologies aren’t just a data modeling artifact, they’re a force multiplier for generative AI. They turn language-first systems into meaning-aware systems, capable of operating within the nuances, constraints, and intent of a real organization. Critically, they shift AI initiatives away from brittle, one-off integrations toward a durable foundation where new use cases become easier to implement over time.

At this point, the role of a semantic layer should be clear. Ontologies give generative AI the grounding it needs. They implement the organization’s logical understanding of its data in a way machines can actually use. And when done well, they turn fragmented data landscapes into a coherent, navigable fabric of meaning.

The remaining question is a practical one…

How do teams actually build, manage, and evolve that semantic fabric over time without it becoming brittle, siloed, or locked inside a single team’s head?

This is where Mobi comes into the picture.

Mobi is a collaborative semantic knowledge graph platform designed specifically to treat ontologies, vocabularies, and related semantic artifacts as living organizational assets as opposed to one-off modeling exercises. It provides a shared environment where teams can define meaning together, evolve it as the business changes, and connect it directly to the underlying data landscape that AI systems need to navigate.

Rather than treating semantics as static documentation or an upfront modeling hurdle, Mobi makes it part of the organization’s ongoing operational workflow.

Most organizations pursuing semantic AI already rely on, or are actively evaluating, enterprise knowledge graph platforms such as Altair Graph Studio, Stardog, GraphDB, or similar solutions. These platforms excel at transforming, storing, querying, reasoning over, and serving knowledge graphs at scale. They are the runtime engines where semantic models are executed and applied.

But as teams quickly discover, operating semantics over time is a different problem than hosting a graph.

Ontologies evolve. Definitions change. New concepts emerge. Use cases multiply. Different teams need to collaborate, review, approve, and version semantic artifacts without breaking downstream systems. This is where many initiatives struggle: not because the graph platform is insufficient, but because semantic change management and governance are left as ad hoc processes.

Mobi is designed to fill that gap.

Rather than competing with knowledge graph platforms, Mobi complements them by providing the semantic operations (SemOps) layer that organizations need to manage ontologies, vocabularies, and constraints as living artifacts within a broader semantic ecosystem.

Concretely, Mobi provides:

• Ontology management across platforms
  A collaborative environment for defining, reviewing, versioning, and evolving ontologies; independent of where they are ultimately deployed or executed.

• SHACL constraints and semantic validation
  Guardrails that help teams maintain quality, consistency, and trust as semantic models evolve and are applied to real data.

• Vocabulary control and harmonization
  Tools to manage terminology drift across teams and domains, aligning language without forcing rigid upfront standardization.

• Semantic integration and lifecycle support
  The ability to manage mappings, transformations, and semantic artifacts that bridge raw data and downstream graph platforms, analytics tools, and AI systems.

Seen this way, Mobi acts as the control plane for semantics, enabling teams to operate, govern, and evolve meaning over time, while existing knowledge graph platforms continue to serve as the execution layer for querying, reasoning, and AI integration.

This separation of concerns is intentional. It allows organizations to invest in best-of-breed graph technologies while avoiding the brittleness that comes from unmanaged semantic sprawl. This ensures that as generative AI use cases expand, the semantic foundation they rely on remains coherent, trusted, and adaptable rather than becoming yet another source of technical debt.