Frame the workforce decision and its evidence · Lesson 4

Data fitness and operational privacy

12 min read

Lesson objectives

  • Profile missingness, coverage, lag, and access while deciding whether an analysis is fit for use.
Reading tools

An HR extract has a 'primary skill' field for every worker, but 40% of values are blank. Does blank mean the worker lacks the skill? No. A person may never have been invited to enter it, one branch may have moved to the new system earlier, or the skill may not yet have been assessed. Coding missing as 'no skill' can manufacture a talent gap and send training funds to the team with the weakest data-entry process rather than the greatest need. Data fitness means that the information is suitable for the question, decision, and time horizon. Size or apparent column completeness is insufficient. CIPD's practitioner guide also notes that personal data processing must meet the law applicable in the relevant jurisdiction. The analyst must join fitness to a legitimate purpose and controlled access. This is professional guidance, not a ruling about the law of any particular country.

Check four dimensions before modelling. Coverage asks which people, periods, and organizational units appear. If contractors provide half the service hours yet are absent from the employee system, a capacity measure is incomplete even if no cell is blank. Completeness asks what share of required records have valid values. Calculate it by team and month, not only in aggregate; a 90% overall average may hide a branch at 50%. Timeliness asks when an event happened and when it entered the warehouse. Do not announce this month's departure rate if the last week's departures commonly arrive ten days late. Validity asks whether values fall in sensible ranges, work fractions reconcile, and job-title codes changed. Combining these dimensions into one score does not remove the need to understand how each defect could alter the recommendation.

A narrower decision may still be possible. If the question concerns Branch A alone and its record is 90% complete, report an A-specific finding with the effect of the remaining 10% unknown. A comparison between branches from the current file is not fair. Investigate the missingness mechanism. If B's missing entries belong mainly to new hires awaiting assessment, missingness is associated with tenure and may be associated with the skill itself. Deleting missing records is therefore not neutral. If a short assessment is introduced, define the task and performance standard and allow people without a prior entry to participate. A data repair process should not quietly become an unjustified exclusion test. Keep interim validation data separate from individual employment decisions until its suitability has been reviewed.

Operational privacy is a design decision. State the purpose before extracting columns, use internal IDs rather than names in analysis, and give managers appropriate aggregates rather than a raw file of individual histories. A tiny group is not necessarily anonymous when the manager knows its members. Record access rights, retention, and correction routes; consult jurisdiction-specific expertise before sensitive processing or cross-border transfers. If a predictive model is involved, the NIST AI Risk Management Framework offers voluntary Govern, Map, Measure, and Manage functions for ongoing risk work. It is not legal permission to use particular people data. Removing a name alone does not prevent re-identification when distinctive details remain.

Who reads the skills report, and what do they need?

An operations manager wants to know whether a new shift can be covered. An analyst sends names, ratings, and training notes. The file is rich but reveals more than the coverage decision requires. Start with intended use: is the manager deciding how much qualified capacity exists, or assigning named people? Aggregated skill and availability may suffice for the first purpose. Named assignment needs different access and further context. Design two outputs instead of one broadly accessible file. The first presents aggregate capacity; a restricted second output supports authorized assignment. Privacy then becomes part of decision design rather than a small warning below an exposed workbook.

Map information from its source to its last recipient. Who assesses the skill, who corrects errors, and where is the planning-approved version kept? A coach's development suggestion can become a permanent worker judgment in a later file. Preserve the score's definition, collection purpose, and date. Separate verified evidence from self-report and unchecked fields. If a manager requests individual rankings, clarify authorization and consequences. Shift cover may require one specific work-task check, whereas individual ranking may need broader safeguards. Specialists determine applicable workplace rules; the analyst states what the professional evidence actually supports.

Data quality and privacy interact when records are missing. Some workers may avoid a skills survey because they fear performance-management use. Treating a blank as absence of skill could move development opportunities away from that group merely because of disclosure concerns. Investigate missingness through suitable evidence and report inference limits. The team might propose a short, clearly explained work check instead of an extensive survey, with a correction channel before assignment planning uses the register. These steps need coordination with workforce and data owners. Authorization to use information remains distinct from whether the assessment validly measures the required skill.

Design correction as carefully as extraction. When an obsolete skill appears in a report, workers or authorized managers need a clear review route. Keep the last verification date and approving function in the restricted record. On correction, identify dependent reports needing updates; fixing the source may leave an old roster plan in another team. Notify relevant owners of changed capacity without spreading sensitive individual reasons. Retain versions according to the organization's policy. The aim is a report fit for today's decision, rather than an expanding set of permanent files with no clear authoritative version.

Test the usefulness of a new field before adding it to a form. A planning team might request every individual's absence reason to forecast next month's capacity. It may only need unavailable hours and expected return dates, while detailed reasons remain with the authorized function. Ask which calculation changes with each column. Build a trial model without detailed reasons and compare the capacity estimate. If forecast hours stay the same, the extra field is unnecessary for that purpose. Where a legitimate need emerges, define it narrowly and review authorization. This shows that data minimization can preserve analytical usefulness rather than making privacy and accuracy permanent opponents.

Finish with three separate decisions in a short record. First is authorization for this use, owned by the relevant specialist. Second is whether the measure represents the skill, supported by task evidence and score definitions. Third is whether the output adequately supports coverage or budgeting. One can be resolved while another remains pending. State that directly so the manager knows what is awaited. Tomorrow, review one recurring report and identify a column that does not change its decision. Discuss removing or restricting it, then test that the reduced report still answers its original question.

Lesson summary

An HR extract has a 'primary skill' field for every worker, but 40% of values are blank.