Pay Equity as a Data Engineering Problem: Preparing for Directive (EU) 2023/970
The EU Pay Transparency Directive is usually discussed as employment law. From where we sit, it is a data engineering brief with a legal enforcement layer. Employers must compute gender pay gaps overall, by pay component, by quartile, and — the hard part — by "categories of workers performing the same work or work of equal value", then defend every number in front of workers' representatives, and remediate any unexplained category gap of 5% or more on a six-month fuse. First reports for large employers are due by 7 June 2027, generally describing 2026 pay. Whoever builds the pipeline decides whether that process is calm or chaotic. Here is how we think about the build.
The Schema Is the Legal Argument
The directive's central abstraction — the category of workers of equal value — does not exist in most HR systems. Job titles are marketing; grades drift; the same role is banded differently across countries and acquisitions. Constructing categories means modelling work itself: skills, effort, responsibility, and working conditions as comparable attributes, applied consistently across functions. Two properties matter more than sophistication. Stability: categories that reshuffle every cycle make year-on-year gaps meaningless and look like gerrymandering. Explainability: every assignment needs a reason a non-statistician can follow, because the audience for this model is a works council, not a data science review. Our general position on grounded, explainable outputs applies with full force here — a category model that cannot show its work is a liability with a schema.
Assembling Pay Is Harder Than Analysing It
"Pay" in the directive includes complementary and variable components — bonuses, allowances, overtime, equity, benefits in kind. In a multinational, those live in a payroll system per country, a compensation tool, an equity platform, and a spreadsheet nobody admits to. The unglamorous 80% of this build is entity resolution and normalisation: one worker identity across systems, one currency-and-FTE-normalised view of each component, one defensible answer to "what did this person earn in the reference year?" Every simplification made here surfaces later as a works council question. The engineering disciplines are familiar: idempotent ingestion, lineage on every field, reconciliation totals against payroll, and versioned reference data — the same provenance thinking we apply to AI outputs, pointed at compensation data.
Analytics With Small-Group Discipline
The computations are elementary — means, medians, quartiles, category deltas. The traps are not. Small categories dominate multinationals' long tails, and a 5% threshold on a category of nine people is noise wearing a compliance trigger. The directive itself acknowledges privacy limits on small groups. So the pipeline needs minimum-group-size rules, suppression and aggregation policies, and uncertainty communicated honestly rather than false precision. It also needs monitoring, not annual archaeology: if the first time you compute a category gap is for the report, you have donated your remediation window to the calendar. Computing continuously turns a 5% breach from a discovered crisis into a managed backlog item.
Reproducibility Is the Compliance Property
A published gap figure is a claim the employer may need to defend years later, possibly in litigation where the burden of proof sits with the employer. That makes the run, not the number, the real artefact: inputs pinned, category model versioned, methodology documented, code and reference data recoverable. "Can you re-produce the 2026 report?" should be a checkout command, not a forensic project. We hold our compliance evidence to the same standard — continuous, generated from the systems themselves, and re-runnable — for exactly this reason.
Design for the Second Audience
Every artefact in this pipeline has two audiences: the regulator who receives the report, and the workers' representatives who consult on its methodology, confirm its accuracy, and — if a gap goes unexplained — co-conduct the joint pay assessment. Outputs designed only for the first audience fail at the second: representatives need category definitions in plain language, methodology they can interrogate, and views that answer their questions without leaking individual pay. The organisations that get this right will treat representative-facing reporting as a product surface, not an export button. The labour-relations side of that engagement — which bodies, which rights, which sequence — is analysed in depth on our sister site: the Pay Transparency Directive and works councils; the operational machinery for running those consultations is Graylark LRM.
The Shape of a Sensible 2026
Build the identity and pay-assembly layer first; it has the longest lead time and everything depends on it. Draft the category model with HR and reward, version it, and socialise it with representatives before it is load-bearing. Run the directive's full metric set on current data privately, with small-group discipline, and triage what you find. Wire the run into something reproducible. None of this requires waiting for national transposition laws — the directive's floor is known, and data foundations are transposition-proof. The employers who treat 2026 as the build year will spend 2027 publishing numbers they already understand.
For the broader platform context behind this work, visit Graylark Technologies.