Over the last 23 years running Taxonomy Strategies, I have watched tech vendors promise endless “magic bullets.”

A director at a major global firm recently made a striking observation: despite millions invested in enterprise AI licenses, he could not point to a single agentic AI application delivering end-to-end, deterministic business utility.

Enterprise taxonomy and metadata projects often face frustrating administrative delays. Contract approvals stall, budget clearance takes weeks, and teams are left navigating market uncertainty.

In the late 1990s and early 2000s, web search arrived with massive promise. Before Google standardized search, we had directory trees and early search boxes like AltaVista. Back then, finding the right document was an exercise in “type-and-pray.”  If you didn’t guess the exact keyword the author used, you got zero results or endless noise. Today, we are watching enterprise …

We are seeing a quiet crisis unfold across enterprise teams: brilliant professionals are feeling trapped between AI inputs and AI outputs. They are handed unread, prompt-generated summaries, asked to review them at lightning speed, and then told to spit out another automated deliverable.

A common frustration among enterprise product managers and marketing directors is publishing high-value technical assets—white papers, research videos, workshop documentation—only to hear total silence. The issue rarely lies in the quality of the content. It lies in a fragmented distribution architecture. To maximize the reach of technical assets, organizations must build an integrated content ecosystem: Algorithmic platforms over static hosting: …

Where does generative AI deliver immediate, low-risk ROI for enterprise knowledge workers? It isn’t in drafting final strategic documents or writing creative copy. It is in breaking computational bottlenecks and structural blocks. When transforming a dense, 40-page technical specification or white paper into a structured executive presentation, AI excels at digesting complex text into preliminary outlines. Outline extraction: Reversing finished, …

When American tech firms plan global AI initiatives or international conferences, they often assume data and people flow seamlessly across borders. The reality is starkly different. From international visa restrictions that force major global summits to relocate out of the U.S., to multi-jurisdiction data privacy laws, physical and administrative boundaries dictate tech execution. Building global data systems requires adherence to …

I will be speaking at Taxonomy Boot Camp in November. Here’s an Exclusive Taxonomy Boot Camp Discount Code to get $100 off your registration: SPKR26. 

Prompting an enterprise LLM to search the entire global internet—or an unstructured multi-terabyte data lake—every time an employee asks a basic question is an operational failure.

In academic and professional publishing, a major shift is underway regarding AI usage disclosures.

I frequently observe a curious dynamic in enterprise software development: highly talented data scientists and coders routinely build applications that solve zero practical business problems.

The most expensive mistake an enterprise can make in information architecture is starting from a blank page.

In modern enterprise tech discourse, human-in-the-loop has become a popular buzzword.

If you only document the inputs you fed into an AI model and the outputs it generated, you are missing the most critical piece of business data: Why were you doing it?

Generative tools are exceptionally fast at identifying gaps in large datasets.

The fundamental problem with modern AI workflows isn’t the model—it’s the lack of explainability.

Enterprise AI development is currently trapped in a familiar historical loop.

When faced with massive, unstructured corporate repositories, systems architects frequently overcomplicate the initial organization of information.

Enterprise AI strategies built strictly inside regional tech echo chambers routinely fail when deployed across global operations.

Relying on massive, unstructured “zero-shot” prompts to drive complex enterprise workflows is an architectural dead end.

The narrative that Generative AI represents a complete break from historical IT disciplines is fundamentally false.

What happens to your custom enterprise AI application when the lead developer who engineered the prompt stack leaves the company?

Twenty-five years ago, assigning a permanent, web-addressable Unique Resource Identifier (URI) to every individual concept in a business vocabulary was considered radical overkill.

When managing data systems across multi-jurisdictional frameworks or independent business units, leadership faces a classic architectural standoff: global standardization versus local autonomy.

Enterprise IT departments routinely view data management failures as software infrastructure problems rather than structural perspective problems.

The fastest way to stall an enterprise taxonomy project is attempting to model your entire corporate organization on Day One. Monolithic, top-down metadata initiatives frequently collapse under their own weight. Successful information architecture requires pragmatic problem-solving—focusing directly on the operational squeaky wheels. To deliver immediate business ROI, target the high-friction pain points that keep department heads up at night: ▶️ …

Relying purely on sequential random sampling to validate enterprise AI models creates dangerous operational blind spots.

If the engineer writing your AI prompts already knows the exact answer key, your validation testing is fundamentally compromised.

As enterprise IT transitions toward agentic AI, organizations face an urgent architectural challenge: component proliferation.

Generative AI is not autonomous intelligence. It is a high-speed mirror reflecting your organization’s internal echo chambers and architectural flaws.

The most dangerous AI implementations in the enterprise are not being built by non-technical managers—they are being built by your data engineers. Highly skilled coders and data scientists understand algorithmic mechanics, but they frequently lack training and experience in information management and data stewardship. When technical teams build Generative AI applications without structured metadata guardrails, a predictable pattern occurs: ▶️ …

Data governance operates under the fundamental laws of physics: without active, structured intervention, information systems naturally degrade toward disorder. In enterprise IT, teams routinely treat taxonomy deployment as a “one and done” milestone. Once the initial classification is built, governance stops. This approach guarantees system failure. Taxonomy management functions exactly like garden care: ▶️ Routine cultivation: If you do not …

A persistent divide exists between academic research and commercial IT execution. Academic reviews demand strict formulas—hypotheses, methodologies, and endless theoretical justifications. Commercial marketing, on the other hand, frequently devolves into vague fluff, unsubstantiated hyperbole, and unstructured jargon.

The modern data landscape is too complex for any single vendor or proprietary platform to own end-to-end. Yet, many enterprise IT departments remain trapped in rigid vendor lock-in, attempting to build monolithic, proprietary data silos. This approach guarantees long-term technical debt.

Early internet search engines relied heavily on popularity heuristics to rank information. What was searched for most frequently was promoted to the top. This created the tyranny of the common—a system where popular content was assumed to be correct, regardless of its actual truth or accuracy.

Reorganizing service delivery across multi-jurisdictional public or private frameworks is one of the most complex tasks in information architecture. When organizations attempt to merge disparate regional operations into a single catalog, they typically default to creating a massive, centralized classification tree. These monolithic structures inevitably collapse under regional variations.

The single greatest myth in enterprise AI implementation is the assumption that technology deploys automatically without strategic planning. Organizations purchase expensive software platforms, distribute licenses, and expect operational transformation to magically occur. This unmanaged approach leads straight into the valley of despair—wasted compute budgets, redundant work, and zero ROI.

Many corporate data strategies fail because they rely on arbitrary, single-tree hierarchies. These monolithic classification structures attempt to force an entire operational domain into one deep, inflexible tree. They scale poorly, break when new data arrives, and create endless orphan terms.

Is a taxonomy merely a list of terms, or is it something larger? Many organizations assume that purchasing or generating a controlled vocabulary completes their metadata strategy. It does not.

Language models have an inherent behavioral flaw: systemic sycophancy. Because these systems are trained to deliver pleasing, natural responses, they tell users what they want to hear. By anthropomorphizing technology, teams risk treating AI like an agreeable pet rather than an objective analytical co-partner.

Natural language is inherently ambiguous. While conversational AI interfaces feel effortless, using natural language as a programming medium creates a severe structural problem: a lack of prompt reproducibility.

Organizations routinely struggle to document artificial intelligence applications effectively. In traditional IT, systems are governed by clear technical specifications. In the rush to deploy AI, teams rely on contextless prompts and informal notes, creating fragmented logic that guarantees failure.

Enterprise AI development is currently facing a discipline crisis: the rise of vibes-based coding.

How do you transform an unpredictable generative AI utility into a robust business application? You must move beyond simple single-shot prompts.  There is a significant business opportunity in building semantic technology frameworks that power specific AI applications. 🎯 Targeted Architecture: Developing custom, targeted point solutions for clients is more complex to frame but offers extremely high value. 🔁 Repeatability: Highly …

Building a controlled vocabulary does not mean you must invent a proprietary schema from scratch. To teach non-experts, it is best to borrow existing public domain schemes rather than constructing rigid, monolithic trees. Monolithic trees are inflexible, scale poorly, and are incredibly difficult to change as data evolves. Best practices for democratizing taxonomy include: 🕹️ Leverage Existing Standards: Borrowing resources …

Standard AI queries carry hidden financial costs that most enterprise leaders have failed to calculate. Unlike traditional web search, AI utilities charge by the token. This represents a return to the old time-share models of computing, where users paid strictly by the minute.  When your staff relies on unstructured queries, they are burning compute budget on statistical guesswork. 🧑‍🏫 The …

Unsupervised AI acts as a digital meat grinder, frequently producing poor-quality, automated content known as slop. This is less problematic in highly regulated fields like finance, medicine, and law because the consequences for errors are severe, forcing organizations to maintain strict guardrails.  However, in less regulated areas like consumer product descriptions, automated generation can lead to degraded systems and the …

The primary myth of Artificial Intelligence implementation is the belief that it occurs automatically without planning. Organizations frequently purchase expensive technology operating under the assumption that advanced processing will magically organize their unstructured data.  It will not.  Successful integration requires orchestration, which involves treating AI as a carefully managed component within a broader portfolio of methods and processes. Without this …

Specialized knowledge domains—such as medical informatics and cardiology—frequently suffer from acute information bottlenecks. This friction occurs because historical workflows rely entirely on manual content curation by internal subject matter experts (SMEs). SMEs are naturally protective of their specific data turf, often harboring an inherent suspicion of systematic, parametrically driven data approaches. They worry that algorithmic data structures will degrade accuracy …

When designing an enterprise metadata schema, determining what not to classify is just as critical as selecting your primary terms. This is the essential principle of negative space in information science. Organizations frequently design over-engineered, narrow, and deep taxonomy structures that collapse under the weight of their own complexity.  For example, during a recent agile infrastructure project for a client …

The unchecked output of unsupervised AI engines presents a severe compliance risk within highly regulated domains such as finance, medicine, and corporate law. In these environments, data inaccuracies are not merely minor errors; they carry immense financial penalties and legal liability. To eliminate this structural risk, enterprise architects should adopt the metadata logging principles perfected by the Wikipedia model. Wikipedia …

Standard enterprise search tools frequently break down because they are hardcoded for literal string matching rather than conceptual context. This is the root cause of the inexact match failure. Consider a practical domain scenario: if an online customer searches a food inventory system for “Buffalo Wings,” a precise semantic layer should comprehend that the user may accept “Spicy Cauliflower.” This …

Corporate customer self-service models are frequently undermined by the assumption that automated conversational bots operate via background wizardry. They do not. Enterprise information systems are currently transitioning from the legacy phone-tree design toward agentic AI frameworks. This shift is evolutionary rather than revolutionary.  To achieve true scale without generating massive operational chaos, data architects must systematically analyze, break down, and …

What problem are you trying to solve? Purchasing an expensive enterprise AI software license before defining your underlying business logic is like purchasing a nuclear weapon. It is an impressive display of capital expenditure, but it will not solve 99.9% of your daily tactical operational problems. Many organizations are currently caught in the E5 License Delusion—the mistaken belief that purchasing …

A taxonomy is a living business logic infrastructure, not an ad-hoc arrangement of labels.  Many digital product managers mistakenly believe that modern search engines or unsupervised AI tools can magically organize information systems without human curation. This ad-hoc approach inevitably leads to degraded system performance and fractured user experiences.  To maximize information value, data architects must adhere to Busch’s Golden …

An insurance policy is not merely a compliance document; it is a complex data asset that should be managed as a structured product catalog.  Most financial and healthcare enterprises fail to optimize information retrieval because their portfolios remain trapped in unstructured information silos.  For example, legacy systems hide the critical connections between coverage parameters, rules, and client demographics.  By transforming …

As AI tools have had the time to be fully implemented and iterated across industries, we’re noticing a shift in the conversation around the technology and how it’s deployed and used. The first thing is to clearly define the problem being solved by your AI tools. Buying AI without a clear purpose is risky at best and a loss at worst. Unless the goal is …

The unchecked proliferation of AI model adoption is driving a massive overbuilding of infrastructure that risks creating digital ghost towns.  We have witnessed this structural pattern before in zombie shopping malls and empty mining towns.  Today, the technology sector faces a cooling period as massive data centers are constructed to ingest uncurated data.  Unsupervised AI acts as a digital meat …

How does a semantic framework transform an unpredictable generative AI utility into a robust business application?  Building an enterprise chatbot is not a matter of deploying a simple single-shot prompt interface.  Unmanaged information systems cannot function effectively without structured metadata schemas.  To move beyond single-shot prompt iterations, organizations must build custom semantic technology frameworks that function as point solutions. Nested …

Standard generative AI queries are becoming fabulously expensive for enterprise environments that fail to implement targeted semantic controls.  Organizations frequently treat AI interfaces like standard web search, overlooking the structural cost mechanics. While web search is cost-free to the end-user, AI utilities charge by the token—a modern instantiation of the old time-sharing models where users paid strictly by the minute. …

I recently returned from London with a collection of B-Roll photos—planes, trains, automobiles, and the 17th-century pubs of Greenwich. On their own, they’re just files on a drive. They only become a story when they are placed into a framework—a timeline, a map, or a narrative of urban renewal. Your enterprise data is the same. AI vendors will sell you …

For a long time, building a taxonomy was seen as a specialized, permanent endeavor—the work of a few high priests of information science. We are entering the era of Democratized Taxonomy. As enterprises adopt localized language models, we need tools that allow teams to generate their own topic maps organically. This isn’t about replacing the expert; it’s about making the …

Documenting what you want to do is as important as doing it. In the rush to implement the latest AI tool, documentation is often treated as an afterthought—or worse, drudge work to be automated away. But documentation isn’t about bureaucracy; it’s about understanding. You need to understand the business logic as deeply as possible before you build the solution. At …

If you have a complex condition like fatty liver disease, you don’t just need a doctor. You need a specific type of practitioner that fits a tertiary care model—a niche within a niche. We recently helped the Centers for Medicare and Medicaid Services (CMS) solve this exact problem. The user knows their symptoms (the problem), but the system is organized …

I recently spent time in Greenwich, London—the home of the Prime Meridian. In the 18th century, the time ball at the observatory dropped at 1:00 PM every day so ships in the harbor could set their timepieces. Without that shared reference point, navigation was guesswork. In the modern enterprise, we are still trying to set our clocks. When different departments …

In the early days of the web—and even back to my work with NASA—the mission was clear: transform fragmented slop into structured intelligence. Today, the tools have changed, but the fundamental logic hasn’t. Many firms are bushwhacking through the AI hype cycle right now, feeling overwhelmed by vendor promises. They need more than a tool; they need a methodology that …

AI is becoming fabulously expensive to run at an enterprise level. If every employee in your organization is asking the same question to a Co-pilot—and you are paying the usage fee for that single-shot query every single time—that isn’t innovation. It’s a waste. We need to move back to the principle of Collection Management. Create a collection of good stuff. …

Standard search is a blunt instrument. If a customer looks for Buffalo Wings, a basic tool might miss the mark. But a precise taxonomy knows that if the user’s intent is spicy sauce, they might accept Spicy Cauliflower. This isn’t just about food; it’s about Interoperability. Whether you are managing Medical Informatics or Legal Archives, your specialized vocabulary needs to …

For a few months, the narrative was that GenAI would replace the drudge work of taxonomy and ontology. The tune is changing. As organizations fall into the valley of despair, they are realizing that unanchored AI is a liability, not an asset. Without a validated fact basis, you aren’t building a solution; you’re building a hallucination engine. Even Gartner is …

Buying an E5 license doesn’t automatically organize your company. I’m seeing a recurring pattern: organizations are spending millions on enterprise AI infrastructure, only to realize they’ve built a high-speed engine with no road to drive on. The Plug-and-Play mentality is a delusion. You cannot automate a process you haven’t defined, and you cannot scale an AI pilot that is grounded …

The current discourse surrounding Generative AI mirrors the existential panic of the early 2000s web search inflection.  In 2002, when Google began to dominate the market, critics predicted massive structural collapses.  Today, we are witnessing an inevitable cooling period as the business reality of AI infrastructure faces necessary economic scrutiny. The hype cycle: Unwarranted anxiety obscures the truth that technology …

There is a massive management failure happening right now: the belief that an E5 Microsoft license and a 25% staff cut equals “productivity.” Real transformation doesn’t come from the tool; it comes from modeling the processes the tool is meant to augment. Financial services get this because they are regulated—they must document the “how” to survive. Most other industries are …

We use the word “interoperability” as if it’s a setting you can toggle on in your tech stack. It’s not. In complex fields—like medicine or global logistics—the “coordination of care” (or data) fails because the specialties don’t speak the same language. The industry is currently obsessed with “Agents” to break down tasks. But agents are only as good as the …

Most organizations are treating Retrieval-Augmented Generation (RAG) as a plug-and-play upgrade. They assume that if they feed a Large Language Model (LLM) enough PDFs and internal docs, the “math” will eventually find the truth. It won’t. LLMs operate on statistical probability—predicting the next token based on patterns, not understanding. When you move from the vast, generic web to a specialized …

In the world of search, we used to focus on “exact matches.” In the world of AI and modern ontologies, the real value lies in inexact matching. Think about food science or product recommendations. If a customer wants “Buffalo Wings” but they aren’t available, what is the viable alternative? Is it fried shrimp? It depends on the context—is it the …