A question I’m hearing in nearly every governance forum lately: should we still be investing in manual taxonomy and ontology development when LLMs can generate plausible-looking knowledge structures in minutes? The vendor pitches are seductive. Point the AI at your document corpus, and watch a knowledge graph materialise. No taxonomists required.

The empirical picture is more complicated than either side of the argument suggests. AI-generated structures are useful in ways that hand-curated taxonomies weren’t, and they fail in ways that hand-curated taxonomies didn’t. The question for governance professionals isn’t whether to choose one over the other but how to combine them sensibly.

What AI Actually Does Well Here

Let’s start with what’s genuinely working. Modern LLMs are surprisingly competent at extracting entities, relationships, and hierarchical structures from unstructured text at scale. What used to take a team of analysts months to draft from a corpus can now be approximated in hours.

The “approximated” qualifier is doing a lot of work in that sentence, but the approximation is good enough to be useful for several governance functions.

The first is initial taxonomy drafting. Starting from a blank page is the hardest part of taxonomy work. An LLM-generated initial structure, even with significant errors, gives subject matter experts something to react to. Reaction-driven feedback produces better outcomes faster than blank-page generation.

The second is gap analysis on existing structures. Pointing an LLM at an existing taxonomy and a corpus of documents that’s supposedly classified using it produces useful identification of mismatches—terms that appear in content but not in the taxonomy, taxonomy terms that don’t appear in content, or relationships implied by content that the taxonomy doesn’t represent.

The third is ongoing maintenance. Taxonomies decay because the world changes faster than the curation cycle. AI-assisted monitoring of changes in terminology usage across the corpus can surface needed updates more reliably than scheduled review processes.

A study published in March by the Information Architecture Institute found that hybrid taxonomy programs—AI-assisted drafting with expert curation—produced higher-quality outputs faster than either pure-manual or pure-AI approaches. The result wasn’t surprising in retrospect but does have implications for how governance teams should structure their work.

What AI Doesn’t Do Well

The failure modes of AI-generated knowledge structures are predictable but consequential.

Definitional precision is the first problem. LLMs generate plausible-sounding term definitions that often gloss over the distinctions that matter in regulated or technical contexts. The difference between “customer” and “client” in a financial services context, or between “patient” and “consumer” in healthcare, is not the kind of distinction LLMs reliably preserve. These distinctions are exactly the ones taxonomies exist to enforce.

Consistency across the structure is the second problem. AI-generated taxonomies tend to use different levels of abstraction in different branches, mix faceted and hierarchical organisation inconsistently, and produce relationship types that aren’t applied uniformly. A human taxonomist enforces consistency through deliberate effort. An LLM doesn’t have the same kind of internal model of consistency.

Authority and provenance are the third problem. A taxonomy is partly a governance artifact—a statement of how the organisation has decided to describe its domain. AI-generated structures don’t carry that authority by default. They feel provisional, and stakeholders treat them that way unless deliberate work is done to validate and adopt them formally.

The fourth, less discussed problem is that AI-generated knowledge structures often reflect the corpus rather than the policy. If your training documents contain historical biases or inconsistencies, the AI structure will encode them as authoritative. A taxonomist would identify these as problems to be corrected; an LLM treats them as patterns to be represented.

The Operational Reality

In governance programs I’ve observed implementing AI-assisted approaches, the most successful patterns share some characteristics.

They use AI to accelerate specific phases of work rather than to replace the work entirely. Initial structure generation, candidate term identification, and corpus-against-taxonomy gap analysis are the strongest use cases. Final adjudication, definitional refinement, and structural consistency remain human work.

They maintain clear authority over the output. The AI-generated artifacts are inputs to human-governed processes, not authoritative outputs in themselves. Stakeholders understand that the taxonomy is owned by the governance team regardless of how it was drafted.

They invest in verification capacity. The shift from manual creation to AI-assisted creation doesn’t reduce the need for expert review—it changes its nature. Reviewers need different skills, oriented toward evaluating AI outputs critically rather than drafting from scratch.

The programs that have failed tend to share opposite characteristics: treating AI outputs as drafts that need light review, underinvesting in verification, and assuming that scale of output equals quality of output.

The Tool Vendor Landscape

The vendor market has bifurcated. Established taxonomy management vendors—PoolParty, TopBraid, Synaptica, and others—have added AI-assisted features to their existing platforms. These tend to be more conservative in their claims and better integrated with existing governance workflows.

A newer set of vendors is selling AI-first knowledge graph platforms, often with claims about replacing traditional taxonomy work. These products have impressive demos and varying real-world results. Several have failed to retain large customers after the initial enthusiasm cycle because the operational realities of maintaining knowledge structures didn’t match the demo experience.

If you’re evaluating tools, the questions worth asking include: how does the tool support human curation of its outputs, what’s the workflow for resolving conflicts between AI suggestions and existing curated content, and what evidence does the vendor have of long-term customer retention in programs comparable to yours.

A Practical Framework

For governance teams trying to figure out how to integrate AI into their work, I’d offer a rough framework.

Start by identifying the phases of your taxonomy and ontology work where AI augmentation is lowest-risk and highest-value. Initial drafting, gap analysis, and ongoing usage monitoring are reasonable starting points. Formal definition work, relationship modeling for high-stakes domains, and final approval are less suitable.

Invest in verification capability before investing in generation capability. The bottleneck in AI-assisted taxonomy work isn’t generating outputs—LLMs do that nearly for free. The bottleneck is evaluating outputs critically and rapidly. Build the human capacity for that first.

Don’t restructure your governance program around the AI capabilities. Restructure the AI capabilities around your governance program. The governance discipline is what produces trustworthy outputs; AI is a tool that supports that discipline when used well and undermines it when used badly.

Several organisations I’ve talked to have brought in outside expertise for the architectural decisions around how to integrate AI into governance workflows. We’ve recommended their .NET team for a couple of implementations where the underlying systems integration mattered as much as the AI capability itself. The right partner depends heavily on your stack and your governance maturity.

The Question Behind the Question

The original question—whether AI knowledge graphs are replacing traditional taxonomies—has an answer that’s both yes and no. The artifacts being produced look superficially similar; the governance disciplines that produce them are evolving rather than being replaced.

The professional taxonomists I respect most are the ones who’ve leaned into AI augmentation rather than treating it as an existential threat. The work is changing. The need for the discipline isn’t going away.