The hook
Payroll is the one system where “the averages look fine” is a dangerous sentence. Retroactivity, allowances, multi-award rules, and casual/overtime edge cases mean that aggregate gross totals can reconcile perfectly while thousands of individual pays are wrong. Two of the most heavily audited payroll failures on record prove the point.
Queensland Health (2010)
Queensland Health’s SAP HR/WorkBrain payroll go-live is one of the most thoroughly audited public-sector SAP failures anywhere. The Queensland Auditor-General found that parallel testing compared individual pays for only a 10% sample of employees. The f irst parallel run surfaced a $1.2 million discrepancy — and after many later configuration changes, no further individual-pay comparisons were done at all. Theaudit flagged explicit “data conversion issues” and recommended that adequate time be allowed for data preparation and migration.
Around 78,000 employees were in scope. Costs exceeded $101.958 million by the end of March 2010, the go-live landed 18 months late, and the interim solution ran roughly 300% over its original budget.
Novopay (2012)
New Zealand’s Novopay payroll service for schools went live in August 2012 and triggered a ministerial inquiry. The official technical review states plainly that “data quality has been affected by system issues,” calling out the migration of initial information from the previous system. The service paid roughly 110,000 people over a year, with a significant minority not paid correctly or on time. Project costs rose $23.9 million above estimate to $56.8 million, and the fallout delayed school audits and affected the Ministry’s financial reporting.
The shared failure — and where TDE fits
Both cases show the same trap: too narrow a comparison population, and no repeatable way to re-run person-level validation after late changes. Test Data Express reframes the problem from “how big is the sample” to “which archetypes are covered”:
- Build risk-based payroll cohorts — casuals, overtime-heavy staff, multi-award employees, retro-pay cases, edge-case allowances — instead of a generic percentage sample.
- Create repeatable migrated-payroll packs so full comparison logic can be rerun after every configuration change, not just once.
- Isolate and refresh conversion-exception datasets for the employees most likely to break award interpretation.
- Capture a failing pay scenario once, mask it, preserve its dependencies, and make it available on demand to developers, testers, payroll specialists, and auditors.
Modeled benefit:
20–45% shorter parallel-run preparation, a large reduction in nonproduction data footprint through selective extraction, and — the real prize — nearcomplete coverage of high-risk payroll archetypes rather than a broad-but-shallow sample.
The takeaway
Raising a sample from 10% to 20% doesn’t help if the wrong 20% is chosen. Coverage of failure-prone archetypes beats raw population percentage every time — and that’s a testdata design problem, not a testing-effort problem.
Sources
- Queensland Health Payroll System Commission of Inquiry — final report, 2013 (covers the Auditor-General’s findings, the 10% sample, and the $1.2M discrepancy):
- Queensland Auditor-General — Report to Parliament No. 7 for 2010, “Information Systems Governance and Control, Including the Queensland Health Implementation of Continuity Project” (primary audit, referenced within the Inquiry report above)
- Novopay Technical Review (Deloitte, March 2013) — official technical review, “data quality has been affected by system issues,” ~110,000 payees (PDF):
- New Zealand Government — “Ministerial Inquiry report into Novopay released” (ministerial summary, $56.8M total cost):
- New Zealand Office of the Auditor-General — “Part 3: Novopay’s effect on school audits”:
