Oriental Nippon Development
Back to blog
Guides

The Parallel Run: The Unglamorous Step That Makes Automation Safe

Invoices to 300+ facilities, payment reconciliation, monthly data redaction, every automation we trust went through the same boring step first. Here's why we never skip it.

Read our case studies back to back and one line keeps turning up in the delivery plans: run the new thing beside the old thing before you switch. It is the least exciting phase in any project, and the one we would defend hardest in a budget meeting.

This post explains what a parallel run is, what it cost on four real projects, and the rules we follow so it doesn't drag on forever.

What a parallel run actually is

For one full cycle (a billing month, a funding period, a run of weekly files) the automation runs and so does the manual process. Every output is compared, item by item. The manual version retires only when every difference is understood, not merely small.

What it cost on four projects

These are the parallel or verification phases from four of our revenue-cycle engagements, taken from their delivery plans.

ProjectWhat was comparedParallel phase
Invoicing for 300+ facilities~500 invoices a month, one by one4 weeks
Weekly insurance pipelineAutomated vs manual weekly output3–4 weeks
Reconciliation for 12 facilitiesSilent findings vs manual reconciliation2–3 weeks
Monthly data redactionOne live cycle across 20+ tables2 weeks

On a project of three to four months, that is roughly a fifth to a quarter of the timeline. It is a real cost, so it needs a real reason.

Billing is where we never skip the parallel run: every invoice in a full period is compared before switch-over.

Where it earned its keep

  • Invoicing. For the first live cycle, distribution pointed at an internal mailbox, not at hospitals. A full period was compared invoice by invoice before anything reached a facility. A misdirected invoice is a data incident, not just a billing error.
  • Reconciliation. The flows ran silently until the people receiving findings accepted the false-positive rate. Alerts nobody trusts get ignored within weeks.
  • Insurance data. Several weekly cycles side by side. A pipeline that quietly disagrees with the process it replaced is worse than no pipeline.
  • Redaction. This one cannot be reversed, so it was built and tested only on a restored copy, then one verified live month before the schedule was trusted.

The rules we follow

  1. Define "correct" before building. Six discrepancy categories for reconciliation, one definition per measure for reporting, all agreed with the business first. Without that, there is nothing to compare against.
  2. Compare everything, not a sample. Samples miss the one facility with an odd attachment rule.
  3. Write down every difference and its cause. The comparison record is handed to the client. Auditors like it.
  4. Switch over in slices. Monthly cadence before weekly. One facility, one depot or one branch first.
  5. Instrument every run. Each automation reports success or failure to a named owner, from day one.

When you can skip it

Rarely, but it happens. If the output is internal, easy to regenerate, and nobody outside the team acts on it, a short spot check can be enough. Billing, compliance and anything a client sees always gets the full cycle.

Why it is worth the time

A parallel run costs weeks. Losing the business's trust in an automation costs the automation.

Sources and further reading

  1. Parallel change, Martin Fowler
  2. Blue-green deployment, Martin Fowler
  3. Alerting on SLOs, Google SRE workbook

Related case study

Hands-Free Invoicing for 300+ Hospital Facilities