Organizations understand what they want from enterprise systems: customer focus, operational responsiveness, digital agility. They want to deliver precisely what employees and customers need at the moment they need it. The data exists. The technology is available. Yet artificial intelligence investments consistently fail to deliver promised value.
The fundamental problem executives must grasp: AI implemented without disciplined foundations fails reliably. However, when AI operates with genuine understanding of business context, it transcends incremental improvement. It becomes central to organizational efficiency and effectiveness.
Understanding the Problem Through Example
Business insurance presents substantial complexity compared to home or auto coverage. It protects against diverse risks—legal and physical—across countless business types, from hair salons to retail establishments to aerospace component manufacturers. This creates challenges for the ten thousand small insurance agencies distributed across the United States who must sell products generating innumerable questions. Sales agents at these companies frequently contacted support lines seeking answers.
This call volume eroded profitability for one major insurer. Each representative required sixteen weeks of training to reach productivity, but high turnover meant continuous investment in new training. Answers lacked consistency, creating additional problems—agents calling back repeatedly, hoping to receive their preferred responses.
This seemed ideal for AI application. The challenge spanned complex domains, required consistency, and depended on extensive information contained in policy documents. Executives launched a project creating a virtual expert—an AI resource intended to generate consistent, scalable answers for insurance agents with policy questions.
In theory, dumping all relevant documents into a system and instructing AI to retrieve needed information seems straightforward. But this ignores a critical step. Any AI consuming that information requires an internal model of the business—a model understanding that "clients" meant the same as "customers," that banks are financial institutions while car dealers aren't, that financial advisors occupy different categories. The model must include knowledge about insurance product types, risk categories, and regulations varying by state or city—and how these variables relate and interact.
The Role of Taxonomies and Ontology
Creating the system so it could understand this complexity required developing classification structures called taxonomies—taxonomies of business types and risk types, for example. These taxonomies linked together into a structure called an ontology. The ontology models the business, enabling the system to ingest policy documents or other information, represent them internally as valuable content, and surface that content when applicable to agent questions.
The system took a year to build. Once launched, it became essential infrastructure. Call center volumes decreased 10% as agents realized they could obtain answers quickly from the online system. Callbacks plummeted because agents received consistent, correct answers on first contact. Call center staff themselves began using it as a resource and reached productivity in twelve weeks instead of sixteen. The president at the time described it as a gigantic success factor. Eventually, the company made it available online where anyone—including prospective customers—could access answers directly.
Why Information Architecture Precedes AI
AI cannot start from nothing. It builds on information structures and architecture. Artificial intelligence begins with information architecture. There is no AI without IA.
AI functions only when it understands organizational essence. Ontology is the key to that understanding. It represents what matters within the company and what makes it unique: products, services, solutions, processes, customer descriptors, organizational structures, methods, and every type of data and content. When built correctly and applied appropriately, it makes the difference between raw AI promise and sustainable delivery on that promise.
An ontology is a consistent representation of data and data relationships that can inform and power AI technologies. It includes elements previously described with terms like data models, content models, information models, data/content/information architecture, master data, or metadata. But it exceeds each of these individually.
Building Ontology: The Process
You cannot purchase an ontology from technology vendors because it's unique to your industry and company. Building it requires systematic process of the kind undertaken for the insurance example—classifying and organizing all information within the company.
Building ontology is step-by-step work that starts with observing how the company solves problems, imagining better and more productive ways to organize those solutions, identifying who uses those solutions, and building detailed use cases. With that work completed—and for any decent-sized company, it's substantial effort—you can develop organizing principles for data and content.
Alternatively, taking a bottom-up, data-centric approach involves observing all data the company uses and organizing it into master sets of taxonomies that, together, comprise the ontology basis.
Top-Down Approach
The top-down method begins with business problems and use cases. How do agents currently get answers to questions? What questions do they ask most frequently? Which answers are most difficult to provide consistently? Who needs access to this information? What contexts determine which answers apply?
This approach maps directly to business value because it starts with problems you're trying to solve rather than data you happen to have. It ensures ontology structure supports actual usage patterns rather than theoretical organizational schemes.
The challenge with top-down approaches is ensuring comprehensive coverage. Starting with specific problems may miss important data relationships that aren't immediately obvious from use case analysis. You may need to expand scope multiple times as you discover additional concepts and relationships.
Bottom-Up Approach
The bottom-up method examines existing data systematically. What data types exist across the organization? How is information currently structured? What terms and categories are used in different systems? Where do definitions conflict or overlap?
This approach ensures comprehensive coverage of existing information landscape. It reveals where different departments use different terminology for the same concepts or same terminology for different concepts. It exposes data quality problems and inconsistencies that must be addressed.
The challenge with bottom-up approaches is connecting data comprehensiveness to business value. You may create perfectly organized taxonomies that don't actually support priority use cases because they're optimized for data structure rather than business problems.
Hybrid Reality
Most successful ontology projects combine both approaches. Start with priority use cases to ensure you're solving actual problems. Validate against existing data to ensure comprehensive coverage. Iterate between business requirements and data reality until the ontology serves both effectively.
The Multiple Returns on Investment
Whatever approach you take, you're creating an asset that pays off in multiple ways. As the insurance company found, a system for organizing and understanding all data becomes valuable to multiple audiences. You can evolve it to encompass more problems and solutions. With ontology to build on, you can apply AI in ways that remain sustainable as technology improves and data becomes richer.
The benefits extend beyond initial use cases:
Consistency across systems: When multiple applications use the same ontology, they share common understanding. This eliminates translation problems between systems and enables meaningful integration.
Faster development: New AI applications can leverage existing ontology rather than starting from scratch. Development time decreases because foundational structures already exist.
Improved accuracy: AI systems grounded in ontology make fewer errors because they understand relationships between concepts. They know which distinctions matter and which don't.
Easier maintenance: When business changes, updating ontology propagates changes to all dependent systems. You don't need to update each application individually.
Knowledge preservation: Ontology captures domain knowledge explicitly rather than leaving it embedded implicitly in individual systems or people's heads. This knowledge remains accessible even when people leave.
The Failure Pattern of Point Solutions
It's certainly possible to apply AI without this foundational work. Organizations do it constantly. But AI point solutions nearly always fail—sometimes immediately, and sometimes later when it becomes clear they're adding complexity layers to already complex problems.
Point solution failures follow predictable patterns:
Limited scope: The solution works for the specific narrow case it was designed for but can't extend to related problems. Each new use case requires starting over.
Integration difficulties: The solution doesn't share common understanding with other systems, making integration expensive and fragile.
Maintenance burden: Business changes require updating the point solution independently of other systems, multiplying maintenance costs.
Knowledge loss: When the people who built the point solution leave, understanding of how it works and why it makes specific decisions leaves with them.
Scaling problems: What worked for small data volumes or simple cases fails when applied more broadly.
You cannot build modern structures on rotten foundations. And you cannot genuinely leverage AI power unless it's built on ontology foundations modeling what matters in the business.
Getting Started
Organizations ready to build ontology foundations should begin with clear scope definition. Don't attempt to model the entire enterprise at once. Start with specific domain that supports priority use cases.
Assemble the right team combining domain experts who understand the business, information architects who understand how to model knowledge, and technical implementers who understand how AI systems will use the ontology. All three perspectives are necessary.
Plan for iteration. First attempts at ontology will be wrong in important ways. That's normal. The process of building ontology reveals gaps and inconsistencies in organizational understanding. Expect to revise significantly based on what you learn.
Establish governance from the beginning. Ontology requires ongoing maintenance as business evolves. Someone must own decisions about what gets added, changed, or deprecated. Without governance, ontology will degrade over time.
Connect ontology work to business value throughout. Don't let it become abstract modeling exercise. Keep asking: how does this support specific business problems we're trying to solve? What decisions will this enable? How will we measure whether it's working?
The Path Forward
The gap between AI promise and AI performance isn't primarily about technology limitations. It's about foundational capabilities organizations haven't built. The most sophisticated AI models cannot compensate for absence of structured business knowledge they need to function effectively.
Organizations investing in ontology foundations position their AI initiatives for sustainable success. Those attempting AI deployment without these foundations position themselves for predictable disappointment regardless of how much they spend on technology.
The work isn't glamorous. Building taxonomies and ontologies doesn't generate the excitement that AI demonstrations do. But it's the prerequisite for AI that actually delivers on its promises rather than just demonstrating what's theoretically possible.
The question isn't whether to build ontology foundations. The question is whether you'll build them deliberately before AI deployment or reactively after discovering AI doesn't work without them. Deliberate investment is less expensive and more effective. Reactive investment often arrives too late to save failing initiatives.
Start with understanding that AI begins with information architecture. Then commit to building the foundations that enable AI to understand your business as deeply as your best employees do. That's when AI transforms from disappointing technology expense into genuine competitive advantage.
This article was originally published on CEOWorld Magazine and has been revised for Earley.com.