How semantic structure, not smarter models, separates AI pilots that stall from programs that produce real business impact.

Most enterprise AI failures share a common origin story. The technology performed exactly as advertised. The model was capable, the demo was convincing, and the proof of concept delivered on its promises. Then the organization tried to scale it, and the initiative stalled.

This pattern repeats across industries. According to the RAND Corporation, more than 80% of AI projects fail, which is twice the rate of failure for IT projects that do not involve AI.[1] Gartner has predicted that at least 30% of generative AI projects will be abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs, and unclear business value.[2] S&P Global’s 2025 survey found that 42% of companies abandoned most of their AI initiatives during the year, a dramatic spike from 17% in 2024.[3] Executives naturally look for technology explanations: the wrong model, insufficient compute, poor vendor selection. But in case after case, the root cause is more fundamental. The organization lacked the semantic infrastructure to make AI accurate and reliable across the complexity of real enterprise operations.

As we have argued consistently in these pages, there is no AI without IA. Information architecture, the semantic structures that give meaning to enterprise content, is the foundation on which every successful AI deployment rests. This article explains what that foundation consists of, why it matters for AI performance, how organizations can identify the highest-value opportunities for building it, and what executives should ask before investing further in AI scale.

The Real Enterprise AI Problem

The enterprise AI conversation remains overly model-centric. But the hard part is not generating language. The hard part is consistently surfacing the right knowledge (authoritative, current, and applicable) to the moment of decision. Most AI failures in production are not intelligence failures. They are application failures: wrong source, wrong version, wrong context, unclear authority, and no traceability.

In business terms, the goal is not “answers.” The goal is outcomes: fewer escalations, faster cycle times, consistent compliance, reduced rework, and lower audit friction. Those outcomes depend on whether the system can answer the question that matters: not “what is likely,” but “what applies here, now, for this customer, product, and jurisdiction, based on an authoritative source.”

If your knowledge is fragmented (multiple definitions, conflicting policies, unclear ownership, weak versioning), generative AI does not fix it. It accelerates it. You get speed without control: fluent output paired with operational drag.

Why “Better Model” Thinking Disappoints in Production

Benchmarks reward general capability. Enterprises operate in constraints. The difference shows up immediately in production: the system answers quickly, but subject matter experts spend their time correcting it. Citations look impressive but do not hold up under scrutiny. Slight changes in how a question is phrased flip the recommendation entirely.

This is why AI risk programs are moving beyond model safety into system safety, encompassing governance, traceability, and managed lifecycle. If you cannot reconstruct how the system reached an answer, you do not have a scalable enterprise capability. You have a liability.

A model is a commodity. Application is differentiation. Enterprises succeed when they make meaning executable: common language, stable identities, explicit relationships, provenance, and lifecycle governance. That semantic backbone is what turns any capable model into an asset that can survive audit, regulation, and operational edge cases.

What Semantic Architecture Actually Means

Large language models retrieve content based on linguistic similarity, not semantic understanding. An LLM does not know what your terminology means, how your concepts relate to each other, or which rules apply in which business contexts. It matches patterns in language.

Consider what happens when your product taxonomy calls the same item three different things in three different systems. An LLM does not resolve that confusion. It inherits it. It may retrieve outdated documentation, pull instructions for the wrong product variant, or blend content from incompatible regulatory jurisdictions. The result is answers that sound authoritative but create liability, erode trust, and stall adoption.

Semantic architecture is the structured foundation that prevents these failures. It consists of four interlocking components, each of which plays a specific role in grounding AI for enterprise use.

Taxonomies: The System for Organizing Meaning

A taxonomy is more than a hierarchy. It is a controlled system for naming things, grouping related concepts, separating distinct concepts, and enforcing consistency. For AI systems, taxonomy determines which content belongs together and which content must remain separate. It governs how variants are grouped and how retrieval paths are narrowed.

In a field service environment, for instance, taxonomy distinguishes product family from model, generation from variant, subsystem from component, and failure mode from symptom. Without these distinctions, retrieval-augmented generation (RAG) systems confuse procedures across similar versions of equipment, delivering instructions that are close but dangerously wrong.

Ontologies: The Logic of Relationships

Ontology describes the semantic structure of the enterprise: how concepts relate to each other in ways that reflect real operational meaning. Ontologies define “see also” relationships, cause-and-effect relationships, dependency chains, and rules for context interpretation. This product addresses this problem; this troubleshooting guide is for this error code. Things that are conceptually related, relationships curated by human subject matter experts.

For AI, ontology is essential because it expresses the logic behind concepts, keeps related ideas connected, and prevents unrelated ideas from merging. In an insurance context, ontology clarifies that a claim relates to coverage, coverage depends on policy type, policy type varies by jurisdiction, and jurisdiction determines regulatory requirements. These relationships control retrieval and prevent model misconceptions. The ontology is the knowledge structure for the enterprise.

Metadata: The Engine of Retrieval

Metadata drives retrieval more than any other factor. Well-designed metadata ensures AI never pulls outdated content, unapproved drafts, content for the wrong product variant, or contradictory definitions. When a manufacturing procedure is tagged with model, variant, revision, and applicability attributes, the AI retrieves precisely the right instructions rather than everything that looks linguistically similar.

Without metadata, retrieval accuracy cannot be governed. And without governed retrieval, every additional use case multiplies the risk of wrong answers at scale.

Knowledge Graphs and Controlled Vocabularies: Connecting the System

Knowledge graphs connect these components to the operational data of the enterprise. This allows the connection of structured data to unstructured content, the “what happened” from analytics to “why did it happen” from a customer analyst. They link categories to content and data. Map controlled vocabulary terms across operational systems and encode the pathways that AI follows during information retrieval. Controlled vocabularies ensure that synonyms, abbreviations, and department-specific variations all resolve to the same canonical meaning.

When “deviation,” “nonconformance,” and “exception” carry different meanings in a quality environment, vocabulary governance prevents the AI from treating them as interchangeable. This matters enormously in regulated industries where precision of language directly affects compliance.

How Application Architecture Turns Capability into Performance

Applying a model is not connecting it to a document repository. It is building a decision system around it, one that can assemble context and enforce constraints. At a minimum, applied enterprise AI must do five things consistently.

Identify the entity: the correct customer, product, policy, asset, or configuration, not a look-alike. This is entity resolution, and it is foundational to reliable context assembly.

Select authoritative sources: approved policies and procedures outrank training decks, drafts, and tribal knowledge. The system must enforce source precedence.

Determine applicability: jurisdiction, effective date, product variant, customer segment, role, and channel all affect which information applies. These dimensions must be computable, not guessed.

Provide evidence: cite the governing clause, version, and provenance so a human can validate quickly. In high-consequence environments, provenance is not optional.

Fail safely: refuse or escalate when evidence is insufficient, rather than improvising. A reliable system knows when it does not know.

These five capabilities are architectural. They sit above the model. They determine whether the model’s general capability translates into specific business performance.

Enterprise Information Metabolism: Where Competitive Advantage Lives

Every business process is a continuous loop of sensing, interpreting, and acting. Organizations capture data and signals from operations (sense), apply semantic structure and domain knowledge (interpret), and deliver insights that enable decisions (act). Your organization’s speed and competitive advantage depend on how fast and reliably these loops execute.

LLMs and agentic AI can dramatically accelerate this metabolism, but only when the information they process is structured, trustworthy, and findable. Without semantic architecture, AI amplifies fragmentation rather than resolving it.

At every point in the cycle, friction slows the flow. Data scattered across siloed systems means days spent chasing the right source. Research confirms the pattern at a granular level: enterprise employees spend an average of three hours every day searching for information they need to do their jobs, with 47% reporting that fragmented knowledge remains their biggest obstacle to productivity.[4] The same item described differently across multiple ERPs means no shared vocabulary for AI to use. Manual handoffs and approval bottlenecks slow every decision. Teams routinely spend two to three days chasing down data sources and verifying accuracy before they can even begin analysis.

You cannot layer AI on top of friction and expect the friction to disappear. AI accelerates whatever it touches, including dysfunction.

Information Leverage Points: Finding Disproportionate Impact

The strategic question is not where to apply AI broadly. It is where to find information leverage points: the specific friction points where one targeted fix has a disproportionate downstream impact.

When you identify a bottleneck in the information metabolism, the approach follows three steps. First, map your information flows and find the bottlenecks with the biggest downstream effect. Second, design targeted interventions: standardize vocabulary, automate handoffs, and deploy agents where processes are well understood. Third, measure. Baseline before, measure after. If the intervention works, scale it. If not, that is a signal to look deeper at the underlying information architecture.

Agentic AI can remove entire manual steps from your processes, but only when the process is understood, and the semantic infrastructure supports it. The foundations make the leverage possible.

The Economic Case: Model Costs Fall, Knowledge Costs Compound

Model capability will keep improving, and inference costs tend to fall. But knowledge costs compound when meaning is not managed: rework loops, escalations, audit remediation, duplicated assistants, and integration debt accumulate across the enterprise.

McKinsey estimates that generative AI could add the equivalent of $2.6 trillion to $4.4 trillion annually to the global economy, with total economic benefits (including broader productivity gains) reaching $6.1 trillion to $7.9 trillion per year.[5] But that value is not unlocked by purchasing a smarter model. It is unlocked by operationalizing knowledge: making it reusable, governed, and applicable across contexts. The same semantic backbone that powers a customer service assistant can ground an internal compliance tool, an engineering knowledge base, and an analytics pipeline. That reuse is where ROI compounds.

McKinsey’s own research confirms this pattern: organizations reporting significant financial returns from AI are twice as likely to have redesigned end-to-end workflows before selecting modeling techniques.[6] Meanwhile, MIT’s 2025 study on the “GenAI Divide” found that roughly 95% of enterprise generative AI pilots produce no measurable profit-and-loss impact.[7] The gap between potential and reality is not a technology gap. It is an architecture gap.

Measuring and Building Readiness

You cannot improve what you do not measure. You cannot automate what you don’t understand. Organizations that scale AI successfully build their foundational capabilities in a specific sequence. And, they assess where they stand before investing in expansion.

Building Foundations in Sequence

The five foundational capabilities for enterprise AI (governance, information architecture, knowledge engineering, data and content management, and operational readiness) must be built in order. Each layer enables and stabilizes the next.

Governance comes first, and it does not have to be heavy. A lightweight decision-making framework is sufficient to start. One organization implemented governance by requiring that every new initiative identify which taxonomies and controlled vocabularies it would use. That single requirement prevented inconsistent terminology from proliferating across AI projects. That was a gating factor managed by governance processes.

Information architecture follows, providing the vocabulary your AI uses to understand your business. Knowledge engineering transforms documents into machine-ready knowledge. Data and content management ensures that knowledge remains accurate, current, and retrievable. Operational readiness embeds AI into real workflows with real measurement.

Each of these layers is a minimum viable program, not a heavy initiative. The key is starting lightweight, building on what works, and investing where leverage is highest.

The AI Readiness Maturity Model

The AI Readiness Maturity Model evaluates organizational readiness across four interconnected domains: Knowledge Readiness, Operational Readiness, Technical Readiness, and Governance. Across these domains, the model that my firm applies examines 17 factors through 74 diagnostic questions, providing a comprehensive view of where an organization stands and what must be built next.

Each factor is scored on a five-level maturity scale from ad hoc to optimized, revealing the specific gaps that prevent scaling and the sequence in which they should be addressed. In our experience, Knowledge Readiness is consistently the weakest domain, which is precisely why so many AI initiatives stall when they move from controlled pilots into messy enterprise reality.

What Executives Should Ask Before Scaling AI

Before approving expansion of AI initiatives, leaders should be able to answer seven questions about the systems they are building.

First: What decisions will the system influence, and what is the cost of being wrong?

Second: What is the authoritative source of truth for this domain, and what outranks what?

Third: Do we have effective dates, versioning, and jurisdiction overlays that the system can enforce?

Fourth: Can we resolve identity for the core entities (customer, product, provider, asset)?

Fifth: Can the system show evidence (clause, version, provenance) for every high-consequence answer?

Sixth: What happens when evidence is insufficient: does the system refuse, escalate, or improvise?

Seventh: How will we measure drift and reconstruct outcomes for audit and incident response?

If the answers to these questions are unclear, the organization is not ready to scale. It is ready to build the foundation that makes scaling possible.

From Velocity to Value

Enterprise AI performance is not a function of model intelligence alone. It is a function of whether you can apply the model inside a disciplined knowledge environment, one that enforces authority, applicability, evidence, and lifecycle governance. That is information architecture as an operating model. And it is why there is no AI without IA.

Organizations with faster information metabolism out-decide and out-execute slower ones. In regulated industries such as financial services and life sciences, wrong answers create liability, not inconvenience. Ungoverned AI in these environments is a safety and legal risk. Even in less regulated sectors, the cost of lost trust and abandoned adoption quietly erodes every dollar invested in AI capability.

The organizations that build semantic foundations will be the ones that capture enterprise-scale AI value. Not because they chose the right model, but because they built the architecture that makes any model effective. The path from pilot to production runs through information architecture. It starts with governance, builds through semantic structure, and scales through measurement. The investment is not optional. It is the prerequisite.

Notes

[1]Ryseff, J., De Bruhl, B., & Newberry, S.J., The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, RAND Corporation, RR-A2680-1, 2024. https://www.rand.org/pubs/research_reports/RRA2680-1.html

[2]Gartner, Inc., “Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025,” July 29, 2024. https://www.gartner.com/en/newsroom/press-releases/2024-07-29-gartner-predicts-30-percent-of-generative-ai-projects-will-be-abandoned-after-proof-of-concept-by-end-of-2025

[3]S&P Global Market Intelligence, “The Big Picture 2025: Generative AI,” 2025. S&P found 42% of companies abandoned most AI initiatives in 2025, up from 17% in 2024.

[4]Coveo, “The Search for Relevance: Can AI Connect Employees to What Matters?” EX Relevance Report, April 2025. Survey of 4,000 U.S. and U.K. knowledge workers at enterprises with 5,000+ employees, conducted in partnership with Arlington Research. The report found employees spend an average of three hours per day searching for information, with 47% citing fragmented knowledge as their biggest productivity obstacle. See also: APQC, “Knowledge Workers Lose 25% of Time to Productivity Drains,” November 2021.

[5]McKinsey & Company, “The Economic Potential of Generative AI: The Next Productivity Frontier,” June 2023. https://www.mckinsey.com/capabilities/tech-and-ai/our-insights/the-economic-potential-of-generative-ai-the-next-productivity-frontier

[6]McKinsey Global Survey on AI, 2025. Organizations reporting significant financial returns were twice as likely to have redesigned end-to-end workflows before selecting modeling techniques.

[7]MIT NANDA Initiative, “The GenAI Divide: State of AI in Business 2025.” The study found that approximately 95% of generative AI pilots produce no measurable profit-and-loss impact.