In the coming months, pay transparency must move from analysis to operations

The Danish bill implementing the EU Pay Transparency Directive is expected to enter into force in January 2027. The final Danish legislation is not yet in place, but companies should use the coming months to get their master data in order, test specific AI applications and turn their ambitions into a plan that can work in day-to-day operations.

Pay transparency is a term that has occupied a great deal of attention for quite some time. And it is just around the corner from moving beyond something we discuss at a theoretical level to becoming a concrete part of reality. One important question is therefore whether your organisation can handle the specific questions, requirements and comparisons that come with it.

When employees are given the opportunity to ask about their pay in relation to comparable groups of employees, you must be able to provide a consistent answer. This does not necessarily require all processes to be perfect from day one. But it does require you to know which information the answer is based on, who is responsible, and how any uncertainties are handled.

The work should therefore now move from high-level analyses to practical execution. In this article, you can read about three areas that should be central: usable master data, well-defined AI applications and a specific plan leading up to the turn of the year.

Pay transparency is often described as a reporting issue. In practice, it is at least as much a question of data quality.

If employees’ jobs, levels, working hours, seniority, organisational placement and pay components are recorded differently, comparisons will be uncertain. And if HR, payroll, finance and managers use different definitions, two extracts from your systems may produce different answers.

The problem is rarely that the company lacks data. More often, the challenge is that there are too many local datasets, manual interpretations and unclear ownership structures.

For example, job titles may be created differently across the organisation. The same title may cover different roles, while different titles may cover work with comparable content. Pay components may also be defined differently, just as working hours, seniority and allowances are not always handled consistently in the systems that are intended to provide the basis for the analysis.

This makes master data more than an administrative discipline. Data becomes part of the company’s explanation of how pay is determined and develops over time.

However, you do not need to clean up all your data at once. The key is to prioritise the information that has the greatest significance for comparisons, employee queries and reporting. The first step should be to clarify which fields are critical, what they mean, which system is the master, and who owns the quality.

Nor is it enough simply to add more fields. If the organisation does not agree on what a job level, job function or pay component means, more registrations will not automatically create better data. Shared definitions are more important than the amount of data in itself.

A practical approach is therefore to identify the most critical errors, correct them first and, at the same time, establish a simple process for future changes. The aim is not a perfect dataset, but a data foundation that the company can explain, control and continuously improve.

Work relating to pay transparency contains many tasks that are manual, repetitive and data-intensive. This makes the area relevant for AI. But it does not make it risk-free.

AI can help you analyse, group and identify patterns. However, the technology cannot decide on its own what constitutes a comparable job or explain why a specific pay difference is justified. A good approach is therefore to begin with a few well-defined applications where the benefit is clear and the results can be controlled.

Here are three obvious areas in which to apply AI:

1. Quality control of master data

   AI can identify missing fields, duplicates, inconsistent job titles, illogical combinations of jobs and levels, and differences between HR, payroll and finance systems. AI can thereby help identify where a manual review should begin. However, it should not automatically change data without a clear approval process.

2. Assisted job mapping

   Based on job titles and job descriptions, a model can suggest relationships between jobs, job families and levels. This can be an efficient way to work through a large volume of material, but the suggestions must be validated by people with insight into the organisation and the specific roles. A model can recognise linguistic patterns, but it does not necessarily understand the actual difference in responsibilities, complexity and working conditions.

3. An internal knowledge assistant

   AI can support HR, managers and recruitment managers in finding answers in approved guidelines. This may concern, for example, what may be disclosed to candidates, how employee queries are handled, and when a case should be escalated to HR or Legal. The value does not lie only in providing faster answers. It also lies in creating more consistent practices across the organisation.

The general recommendation is therefore to use AI technology for analysis, assistance and quality control. However, the technology should not be given final responsibility for classifications, explanations or decisions that affect employees’ pay. You should therefore decide in advance which data sources the solution may access, who validates the results, and how its use is documented.

Nor does the technology eliminate problems in the data foundation. If job titles are used inconsistently, or historical data reflects inconsistent practices, AI may make the ambiguity faster and more widespread. AI must therefore be combined with clear definitions and human oversight.

You should use the coming months leading up to the turn of the year to create visible results. A realistic plan can be divided into four phases.

1. Create an overview

   This involves appointing a responsible sponsor, establishing a cross-functional team and mapping the most important data sources and processes. A good approach is to identify the greatest data-related risks and then select one or two AI applications that are relevant to test.

2. Prioritise the effort

   Establish key definitions, clarify ownership and correct the master data errors that have the greatest impact on analyses and employee queries. This is also where it may be beneficial to document which pay components are included in the comparison basis.

3. Test

   Conduct a pilot analysis of selected employee groups, test realistic employee queries and trial any AI solutions on a controlled dataset, with the results validated manually. The test should not only show whether your systems can produce a result, but also whether HR and managers can explain the result consistently.

4. Make the work part of operations

   Roles and responsibilities must be approved, relevant employees must be trained, and a simple process must be established for ongoing data control, updates and the handling of queries. At the same time, you should decide how to document the use of AI and how the solution will be developed after the turn of the year.

There are still legal details that may change before the Danish legislation is finally adopted. This does not change the fact that high-quality master data, shared definitions, clear ownership and practical processes will be necessary regardless of the final wording.

Pay transparency should not be treated as one large project that can only begin once all decisions have been made. It should be broken down into specific tasks that can be solved, tested and continuously improved.

Your next management discussion should therefore be based on four questions:

• Do we know which master data is essential to explaining pay differences?

• Have we assigned responsibility for data quality and employee queries?

• Have we selected specific AI applications where the benefit is clear and the risk can be controlled?

• Do we have a plan with responsible owners, deliverables and testing milestones leading up to the turn of the year?

If the answer to one or more questions is no, the next step should not be yet another high-level analysis. It should be to select the specific task that needs to be solved first.