Parallel Running

Parallel running is expensive and unpopular, and it is the only reliable way to find a rule configured wrongly before someone is paid wrongly.

How long

Three periods, not one.

The first finds gross errors: missing people, wrong pay codes, an integration that dropped records.

The second finds rule errors, now that the gross noise is gone.

The third finds edge cases: the month with a public holiday, the person who changed contract, the shift crossing a period boundary.

Organisations that run one period go live with the rule errors intact, and find them in production where correction costs a re-run and an apology.

Compare line by line

Per person, per pay code, not totals.

Two errors in opposite directions produce a matching total, which is the reason total-level reconciliation gives false confidence.

Automate the comparison. Exporting both and diffing is an afternoon of work and turns a week of manual checking into a report.

Sort by absolute difference and work down. The distribution is usually a handful of large differences and a long tail of rounding, and the tail tells you about rounding configuration.

Include the awkward cases deliberately

The person on nights whose shift crosses midnight.

The one with two contracts or two pay groups.

The one who took an interrupted break.

The one who changed role mid-period.

The seasonal or agency worker.

A parallel run on a sample of ordinary employees proves the ordinary case, which was never in doubt.

What to do with differences

Classify each: old system wrong, new system wrong, or a deliberate policy change.

The third category is real and is frequently the reason for the project — a rule that was being applied incorrectly for years. Document it, because it has consequences beyond the system.

Fix the configuration, re-run the period, compare again. A difference explained but not re-tested is not closed.

Keep the comparison records. They are the evidence that the new system was verified, and an auditor will ask.

Signing off

By payroll, in writing, naming the periods compared.

With the remaining known differences listed and accepted, rather than an assertion that everything matches.

Before the old process stops, not concurrently with stopping it.

And with a date after which the old system stops being updated but remains readable, which is a different thing from being switched off.

Automate the comparison

An afternoon of work that turns a week of checking into a report.

Export both systems' results for the period.

Diff per person per pay code.

Sort by absolute difference and work down.

The distribution is usually a handful of large differences and a long tail of rounding, and the tail tells you about rounding configuration rather than about rules.

Keep the comparison files as the evidence that verification happened.

Exercise the clock event

Parallel running should include late starts, forgotten stops and manager amendments. Use this workflow example to shape the scenarios, then reconcile every resulting record with the old process.