Which HR metrics actually matter?

Name the decision it changes and the person who makes that decision. If neither can be named, the metric is decoration and it is costing somebody real effort every period. The second test is whether a movement would actually cause anyone to do something differently, and what that something is. The third is whether the number can be produced identically every period without manual reconciliation, because anything depending on a spreadsheet judgement will quietly change definition and nobody will notice. Applying all three to an existing HR pack usually shortens it, which is the point rather than a side effect.

Headcount against plan, by function, because it drives budget and hiring approvals and both finance and the business act on it. Open roles and how long they have been open sit alongside it. Span of control, meaning how many people report to each manager, is worth watching because both extremes cause trouble: too wide and nobody gets managed, too narrow and the structure carries cost it does not need. Contractor and temporary share belongs here too, since it drifts upward quietly during a hiring freeze and creates an exposure nobody chose. These are structure metrics, and they should be read at the level where somebody can change the structure.

As a split, never as a single figure. Voluntary against involuntary, and within voluntary the regretted share, with the trend shown against your own history rather than an external number. Add early attrition, meaning departures inside the first year, as its own line, since it points at a different set of causes from departures among long-tenured staff. Keep each view at the level of a decision: an organisation-wide figure supports a board conversation, a per-team figure supports a conversation with a manager, and mixing the two produces a chart nobody can act on. The method behind the splits belongs in a separate piece of work, not in a monthly pack.

Payroll accuracy, expressed as the count of corrections and off-cycle payments made after a run closed, because each one represents an employee paid wrongly and avoidable rework; payroll software that logs corrections against the run makes this countable rather than anecdotal. On-time submission of statutory filings, tracked as a simple met or missed, matters because the rules vary by state and change, so confirm the current position with a qualified advisor or the relevant authority rather than trusting last year's calendar. Data completeness in the employee record belongs here, since every downstream report inherits its gaps. So does query resolution time for the HR helpdesk. None of these describes the workforce; they describe whether the function is dependable.

Absenteeism trend, read as a direction on your own history and cut by team, because a rise concentrated in one place is usually about that place. Leave balances building up, which flags people who cannot take time off and a settlement liability accumulating quietly. Overtime landing repeatedly on the same names. Internal mobility, meaning the share of roles filled from inside, which tells you whether people can see a path. Time to productivity for new hires, which is the honest end of hiring quality and needs the manager to define what productive means for the role first. Several of these draw on attendance and leave data, so a connected attendance system removes most of the collection effort.

Because the definitions underneath them differ. Two organisations reporting absenteeism may count different absence types, treat part-day absence differently, and draw on workforces with different shift patterns, so the comparison is arithmetic without meaning. The same applies to engagement scores, attrition rates and cost per hire: published figures vary by industry, by region and by how each was collected, which makes them poor targets and worse justifications for a decision. Internal comparisons hold up because you control the definition on both sides. Compare this period with the last, this site with that one, this year's joining cohort with last year's, and be explicit whenever a definition changes.

Give every metric an owner and a review date, and retire it when it fails the decision test. Resist adding one because a system can produce it; reporting attention is finite and every extra line dilutes the rest. Once a quarter, read the pack as a reader rather than an author and ask which lines anybody acted on. Where an analytics layer makes a hundred charts available on demand, the discipline matters more rather than less, because the constraint that used to limit the report has gone. A short pack that gets read changes more than a comprehensive one that gets filed.