HR has spent a decade fighting for a seat at the data table. Dashboards, people analytics teams, “data-driven talent decisions” on every slide. And yet there is a problem hiding underneath all of it — one that better dashboards cannot fix.
Even if HR were perfectly data-informed tomorrow, it would still be managing a dynamic workforce with static data.
We talk a lot about the usual pathologies of HR data assets: disparate systems, inaccurate records, duplication, incoherent taxonomies, missing timeliness. All real. But there is a deeper, structural issue that persists even in organizations that have cleaned all of that up: the data is static by design, because the processes that produce it are static by design.
Static data is not an accident. It’s process design.
Look at how the core artifacts of talent management come into existence:
Performance is assessed once or twice a year. Between two assessment events, the organization officially knows nothing new about how anyone is performing — even though performance is generated continuously, every single day.
Job descriptions are written when a role is created or re-graded, then quietly drift away from reality. Ask any HR team what proportion of their JDs accurately describe the job as it is done today. The honest answers are sobering.
Skill profiles are captured at a point in time — a self-assessment campaign, a manager review, an assessment center. The moment the campaign closes, the data starts aging. Skills are acquired, applied, and forgotten continuously; the record is updated episodically, if at all.
Organizational structure is snapshotted for headcount reports, while the actual flow of work — who works with whom, on what, using which capabilities — reorganizes itself weekly.
None of this is negligence. It is the logical consequence of a document-centric, event-based operating model. The JD is a Word file. The assessment is a meeting. The skill inventory is a survey. Each produces a snapshot, the snapshot gets stored, and storage is where the data’s life ends rather than begins.
The result: talent decisions are systematically made on outdated data. An internal mobility decision draws on a skill profile captured eighteen months ago. A strategic workforce plan is built on an org snapshot that predates two reorganizations. A pay decision references job content that no longer matches the job. The organization moves; the data describing it stands still. The gap between the two is where bad decisions live.
“Assess more often” is not the answer
The instinctive fix is frequency: quarterly reviews instead of annual, skill surveys twice a year. This fails for two reasons.
First, it doesn’t scale. Assessment fatigue is real — you cannot survey your way to real-time data. Every additional campaign costs attention, and attention is the scarcest resource in any organization.
Second, and more fundamentally: a document-based model cannot be made dynamic by running it faster. A snapshot taken monthly is still a snapshot. You have not changed the nature of the data, only its refresh rate — at rapidly escalating cost.
The problem is not the interval. The problem is the architecture.
The knowledge graph — honestly
The structural answer, in my view, is a workforce knowledge graph: people, roles, skills, projects, learning events and organizational units as entities, connected by explicit relationships. But I want to be careful here, because “knowledge graph” has become one of those phrases that gets nodded at in conference rooms and then implemented as a slightly fancier database.
A knowledge graph does not solve anything by itself. If it is fed by the same annual cycles, you have simply stored your static data in a more elegant structure. The graph is the carrier, not the cure. Three conditions have to hold for the shift to be real:
1. Every edge carries a timestamp, a source, and a confidence
The unit of knowledge is not “Anna knows Python.” It is “Anna’s Python proficiency was inferred at level 4 in Q3 2025, from project deliverables, with 0.7 confidence — and that confidence decays over time unless refreshed by new signals.”
This sounds like pedantry. It is the entire point. Data with vintage can be reasoned about: you know how old it is, where it came from, and how much to trust it. Data without vintage silently pretends to be current, which is precisely how an eighteen-month-old skill profile ends up driving a mobility decision as if it were captured yesterday. Every data point carries its age; the system knows what it doesn’t know.
2. Continuous signal ingestion replaces assessment events
The graph stays alive not because someone fills in a form more often, but because the operational exhaust of work continuously updates it: project assignments, work artifacts, learning completions, internal moves, collaboration patterns. A skill stops being an assessed state and becomes an inferred state — derived from what people demonstrably do, refreshed as they do it.
This also answers the objection I hear most often: “We can’t even keep our JDs up to date — now you want us to maintain a graph?” It is exactly backwards. The document-centric model fails because it depends on manual authorship. A signal-fed graph reduces the maintenance burden, because it is built from what the organization already produces in the course of working, not from documents someone has to remember to rewrite.
3. Inference proposes; humans decide; decisions are documented
Inferred data is probabilistic, and probabilistic data must never masquerade as fact. The graph derives a current best estimate; a human validates it before it drives a consequential decision — a promotion, a redeployment, a pay adjustment — and that validation is logged. AI suggests, human decides, every decision is documented. This is not a compliance afterthought. It is what makes the whole architecture defensible when someone — an employee, a works council, a regulator — asks why a decision was made and on what data.
Without these three conditions, a knowledge graph is a static data asset with better UX. With them, it becomes something categorically different: a representation of the workforce that ages honestly, updates from reality, and knows the limits of its own knowledge.
What this looks like in practice
We recently completed a skill profiling exercise for a large Hungarian telecommunications company: skill profiles for roughly 1,500 distinct positions, mapped to a structured taxonomy of about 700 skills, in a machine-readable format designed to plug into downstream talent systems. Valuable work — and I will be the first to say that on the day of delivery, it is a snapshot.
Its actual value depends on what happens next: whether those profiles become nodes and edges that live — refreshed by hiring, by project staffing, by learning data — or whether they become this year’s version of the JD folder, admired at launch and quietly obsolete by next spring. Building the profiles is the entry ticket. The architecture that keeps them true is the game.
That is the practical sequence I would recommend to anyone starting this journey: do not attempt to build the whole graph at once. Start with the job architecture as the spine — it is the entity your reward, mobility and planning processes already hang on. Layer a skill taxonomy onto it. Then, and only then, start connecting the dynamic signals. A graph grown from a solid spine is defensible; a graph assembled from everything at once is a demo.
The uncomfortable conclusion
The static-data problem cannot be fixed inside the current process model, because the current process model is the problem. Annual cycles and document artifacts will keep producing snapshots no matter how good your BI layer is.
The organizations that get this right will not be the ones with the most sophisticated dashboards. They will be the ones that redesigned how workforce knowledge is produced — continuously, with provenance, with honest uncertainty, and with humans accountable for the decisions built on it.
Your workforce changed while you were reading this article. The question is whether your data noticed.
And no, this isn’t science fiction — Microsoft, Workday and others are already building exactly this into their platforms. Whether you should rely on theirs or build your own is a different question. That’s the next article.
Start with what your HR data can already support.