The Vendors Will Sell You a Skills Graph. Build Your Own Anyway.

In my previous article I argued that HR’s deepest data problem is not quality but statics: annual cycles and document artifacts produce snapshots, and talent decisions run on those snapshots long after they’ve stopped being true. The structural answer, I argued, is a workforce knowledge graph — timestamped, signal-fed, with humans validating what inference proposes.

It isn’t science fiction. In fact, the big vendors are already building exactly this, at a scale no internal team can match. So why would any company build its own?

Let me take that question seriously — because the honest answer is not “everyone should build.” It’s “some organizations should, for reasons that have nothing to do with technology pride.”

Credit where due: this already exists

Microsoft’s People Skills infers skill profiles from Microsoft 365 activity signals — documents, emails, chats, meetings — using LLM-based inference over the Microsoft Graph, mapped to a taxonomy derived from LinkedIn’s skill library, and refreshes those inferences on a rolling basis. Workday’s Skills Cloud is a machine-learning-driven skills ontology built into the core of Workday HCM, using graph technology to map how skills relate and to suggest and infer skills across recruiting, learning, and talent processes. Eightfold has built an entire talent intelligence platform on deep learning over a global dataset of career trajectories, inferring not just current skills but adjacent skills and learnability.

This is precisely the architecture I described: continuous signals instead of assessment events, inference instead of self-declared documents, a living graph instead of a folder of JDs. The vendors got there first, with more data and more engineers than you will ever have.

And yet. Three structural problems don’t go away no matter how good the vendor’s model is — because they are not model problems. They are ownership problems.

1. The vendor’s graph sees the vendor’s world

An inference engine can only infer from the signals it observes. Microsoft’s engine, by its own documentation, currently draws on Microsoft 365 activity — it does not see your line-of-business applications. Which means it sees your workforce as reflected in documents, emails, and meetings.

Think about who that systematically favors. The people whose work produces M365 exhaust: managers, analysts, project leads, anyone whose job is coordination and documents. And who it systematically underweights: your field engineers, your plant operators, your developers living in Git and Jira, your customer service agents in the CCaaS platform, your sales team in the CRM. Their actual competence signals live in systems the vendor’s graph never touches.

This is not a criticism of Microsoft — it is a structural property of any platform-anchored graph, Workday’s and everyone else’s included. The vendor’s graph is anchored to the vendor’s ecosystem. Your workforce’s real signal set is scattered across your entire system landscape — and only you can see all of it. A graph that is meant to represent your organization has to be able to ingest from all of your organization, not from one vendor’s slice of it.

2. “The model said so” is not a defense

Here is the scenario that should worry you more than any technical limitation. An employee is passed over for an internal role. The works council asks why. The answer, traced back through your shiny skills-based process, is: the platform’s inferred skill profile ranked her lower.

Inferred from what, exactly? Which signals, weighted how, with what confidence, as of when? For most vendor inference engines, you cannot fully answer that — the inference logic is proprietary, the training data is aggregated across customers, and the output arrives as a conclusion, not as an auditable chain of evidence.

For low-stakes uses — suggesting a learning course, helping someone find a colleague who knows Kubernetes — this opacity is a perfectly acceptable trade. But the moment inferred skill data feeds consequential decisions — promotion, redeployment, selection for restructuring, pay — the standard changes. Under the EU AI Act, AI systems used in employment decisions sit in the high-risk category, with obligations around transparency, documentation, and human oversight. And independent of regulation, basic organizational justice demands that a consequential decision be explainable to the person it affects.

An owned graph lets you meet that standard by construction: every edge carries its source, its timestamp, and its confidence; every consequential decision records which human validated what, based on which evidence. AI suggests, human decides, every decision is documented — and documented in your system, auditable by you, not summarized in a vendor’s transparency note.

3. Your skill data is a strategic asset. Rent the tools, own the asset.

Workforce capability data compounds. Every project staffing, every learning completion, every internal move adds to a longitudinal picture that becomes more valuable every year — arguably one of the most valuable data assets a company can build about itself.

Now ask: where does that asset live, in what structure, mapped to whose taxonomy? If the answer is “inside the vendor’s platform, in the vendor’s ontology,” then your most strategic people data is structured in terms you don’t control and portable only to the extent your contract says so. Switch platforms in five years — companies do — and you discover how much of that accumulated intelligence actually transfers.

There is also a subtler taxonomy problem. Universal ontologies — LinkedIn’s tens of thousands of skills, Workday’s cross-customer ontology — are impressive general-purpose maps. But your talent decisions hang on your job architecture: your families, your levels, your career paths, the specific capability distinctions that matter in your business. A universal taxonomy laid over a specific organization always fits approximately. When the graph is the basis for consequential decisions, “approximately” is a liability.

So who should build?

Honestly: not everyone. If you are a 400-person company running a single-vendor HR stack, with no data engineering capability and no plans to feed skill data into consequential decisions — use the vendor’s graph, confirm the inferences, be glad it exists. Building would be vanity.

The case for building your own gets strong when three conditions stack:

Your decisions are consequential and contested. Skills data will feed promotion, redeployment, restructuring, or pay — decisions that employees, works councils, and regulators can and will challenge, and where you must own the evidence chain.

Your signal landscape is multi-system. Your workforce’s real activity is spread across an HRIS, dev tooling, operational systems, learning platforms, project tools — and no single vendor sees enough of it to represent your organization fairly.

Skill intelligence is strategic, not administrative. You intend to build durable capability advantage on this data over a decade, which means it must survive every platform migration you will ever make.

The pragmatic architecture: own the spine, rent the signals

“Build your own” does not mean rebuilding Workday in your basement. The sane architecture is narrower and cheaper: own the graph, treat vendor inferences as one signal source among several.

You own the spine — the job architecture, the skill taxonomy fitted to it, the entity model, the provenance layer where every edge carries source, timestamp, and confidence. Vendor-inferred skills flow into that graph as evidence: “Microsoft’s engine inferred X, on this date, from M365 activity” — a signal with a known scope and a known blind spot, weighted accordingly, sitting alongside project-assignment signals, learning data, and human validations. The vendor engines become sensors. The record of truth — and the accountability for decisions built on it — stays with you.

That is the actual dividing line. Not build versus buy, but: who owns the record of truth about your workforce’s capabilities, and who can defend the decisions made on it?

If the answer to both is “a vendor,” you haven’t built a skills-based organization. You’ve subscribed to one.

Own the record of truth about your workforce capabilities.

Explore gAIn →Explore Job Architect →