And you probably shouldn’t build your own taxonomy anyway — the vendors just want you to think so.
A few weeks ago, at the end of a post about pay transparency, I made a promise: to explain what a skill taxonomy, a skill ontology and a skill graph actually are — and why most companies have the first and rarely the third. This is that post. But I want to go further than definitions, because there are two things about skills that almost nobody says out loud: one about buying, and one about forgetting.
Why this conversation is happening now
The numbers behind the urgency are well established: research building on the OpenAI–OpenResearch–UPenn and Brookings work (summarised by Mercer) suggests generative AI touches roughly 80% of today’s jobs, transforms about half the tasks in a fifth of them, and carries a 10–30% productivity potential in AI-augmented work. The practical consequence is simple and uncomfortable: job content is now changing faster than HR structures can follow. If your hiring, development and pay decisions run on job descriptions written three years ago, they run on “fiction”.
That’s what pushes organisations from the old model — one person, one job, linear careers — toward skills ↔ work: tasks and projects matched to capabilities, wherever those capabilities sit. The published Standard Chartered / Mercer case shows what’s at the end of that road: 26,000 users on an internal talent marketplace, 1,700 internal assignments, 144,000 hours redeployed, $6.9M in productivity value.
But between here and there sits a language problem.
Taxonomy, ontology, graph
A skill taxonomy answers: what is what. A hierarchical catalogue of skills — the shared vocabulary. Without it you get what I call “skill soup”: lists of tens of thousands of entries, full of duplicates and synonyms, where “stakeholder management” and “managing stakeholders” are two different skills and none of the data is comparable.
A skill ontology answers: what relates to what. Which skills are neighbours, which cluster together, which map to which roles. This is where skills stop being a list and start being a structure. In our Hungarian Telekom job-architecture work, we distilled 700+ raw skills into skill domains — and it was the domains, not the raw list, that made similarity between jobs analysable.
A skill graph answers: what follows. Once skills and their relationships exist as connected data, inference becomes possible: gap prediction, career-path recommendation, matching, reskilling routes. This is the layer where AI actually earns its keep — and the layer almost nobody has, because it presupposes the two below it.
Many organisations have a taxonomy (usually a messy one). Some have fragments of an ontology. The graph is rare — and that’s fine, because:
You don’t have to eat the whole menu
The most useful framing I know comes from Jesuthasan and Boudreau (MIT Sloan Management Review, 2021): skills-based operation is a spectrum, not a switch. Fixed roles → hybrid (internal marketplace, jobs and skills planning) → flow-to-work (capability deployment, skills-based planning throughout). The goal for almost every organisation is not the end state — it’s the next stage. Most large organisations I meet sit at the fixed→hybrid transition, and the single most expensive mistake is buying a flow-to-work vision when what you need is a clean taxonomy and one working use case.
The part the vendors won’t tell you
Here’s my contrarian claim: you probably shouldn’t build your own skill taxonomy — and you shouldn’t buy a proprietary one either.
When a technology vendor tells you that your skills journey must start with their specialised, “AI-curated,” proprietary taxonomy, notice what that architecture does: it makes your skill data speak a language only their platform understands. That’s not a data model but rather a nicely packaged vendor lock-in.
The publicly available taxonomies — O*NET, ESCO — are, for most organisations, simply good enough. And “good enough” undersells them, because public taxonomies have properties no proprietary one can match:
- They’re free. The single most defensible line item in your business case.
- They’re continuously maintained — by institutions whose entire job is maintaining them, not by your HR team.
- An entire research ecosystem is built on them. Labour-market studies, AI-exposure analyses, benchmarks — when your skill data speaks O*NET, every one of those studies becomes directly applicable to your workforce. Your data stays comparable to the world’s.
This is why gAIn, our strategic workforce planning tool, runs on O*NET. Not because we couldn’t build a proprietary taxonomy — but because a workforce plan you can benchmark against published research is worth more than one that lives in a walled garden.
Can you build your own? Of course — and for genuinely firm-specific capability areas, a thin custom layer on top of a public spine makes sense. But a taxonomy is not a project; it’s a pet. It needs feeding forever. Every skill the market invents, every term that drifts, every merger — someone has to maintain it.
The question nobody’s tracking: which skills is AI draining?
And one more reason skill data has to be living infrastructure rather than a one-off inventory. Everyone measures the skills AI adds — prompt literacy, tool fluency. Almost nobody tracks the skills it quietly removes.
This is no longer a hypothesis. A multicentre study published in The Lancet Gastroenterology & Hepatology followed endoscopists at four centres after AI-assisted detection was introduced. Their detection rate without AI fell from 28.4% to 22.4% — a six-point absolute drop in a core professional skill, within months, in experienced clinicians. It’s the first real-world evidence of automation-induced deskilling tied to patient outcomes — and there is nothing colonoscopy-specific about the mechanism. Delegate a skill long enough, and the skill leaves.
For workforce planning, this changes the question. A static skill inventory tells you what your people had at the moment of measurement. What you need to know is direction and velocity: which skills are appreciating, which are quietly atrophying because a tool took them over, and which of those atrophying skills you’ll desperately need on the day the tool is wrong. That’s not a taxonomy question. That’s a graph question — and it’s the real reason the third layer matters.
Where to start
Not with a platform. With three questions: Do we have a shared skill language, or skill soup? Which single talent process — hiring, development, pay — would benefit first from structured skill data? And what’s our next stage on the spectrum — not the end state?
Public spine, thin custom layer, one use case, then the next stage. Skills-based organisation isn’t a transformation programme. It’s a sequence of defensible decisions.
Public spine, thin custom layer, one use case at a time.