2 Salary Administration AI Agents, Live in Production.

Salary administration has an asymmetric error cost that most process design ignores. An error found before the payroll cut-off is a data correction taking two minutes. The same error found afterwards is someone’s rent arriving short, a difficult conversation, an off-cycle payment and a loss of trust that outlasts the fix. Yet the checking is usually done after submission, because before the cut-off is exactly when everyone is busiest. Pay review introduces a second, quieter problem: rates drift out of band through successive individual decisions, changes get applied without recorded authorisation, and differences accumulate between people doing identical work. Nobody intends any of it, and nobody audits for it, so it surfaces years later as a pay equity finding.

These two agents put the checks in front of the deadline. Payroll Input Validation checks what HR is sending to payroll before the cut-off — starters, leavers and changes — for exactly the reason above: the same error costs almost nothing here and a great deal afterwards. Salary Data Validation checks salary records before a pay review lands: that every rate sits in its band, that changes carry authorisation, and that the gaps between comparable people can be explained by something other than who they are. That last test is framed as a question rather than a verdict, because a gap can have a legitimate explanation — different tenure, a retention decision, a market premium — and the point is to make sure someone has actually stated it rather than to score the organisation.