Begin with what the system still does well
An older system may contain years of business knowledge. Before replacing it, identify the work it reliably supports and the constraints that actually cause trouble: fragile exports, unsupported infrastructure, limited access or a queue of manual corrections.
Take an order system that works well for sales but cannot share stock or invoice status. A complete replacement would also change customer records, order entry and reporting. An integration may solve the immediate problem with a smaller operational change.
Draw a boundary around the first improvement
Choose a flow with a clear input, output and owner. For an order-to-invoice connection, agree which system owns customers, stock and invoice numbers. Use a shared order reference and document how quantities, currency and statuses map between systems.
Start by reading and reconciling data before allowing the integration to write it. Compare a representative sample, including returns and cancelled orders. Investigate mismatches rather than normalising them away.
Design failure as an ordinary state
Connections fail. A safe integration needs timeouts, retries and duplicate protection. A retry must not create a second invoice. Record the external reference and use an idempotency key so the same instruction can be recognised.
When stock is short or a record is incomplete, place it in a visible exception queue. Give someone responsibility for resolving it and preserve the original information. Do not silently substitute a value just to keep the flow moving.
Pilot, reconcile and keep a way back
Run a limited pilot with a team or an order category. Compare quantities, totals and record counts against the source systems. Agree the acceptable error rate, monitoring owner and conditions for stopping the rollout before starting.
Keep the previous route available until the pilot passes its checks. A rollback plan should explain how to stop new writes and reconcile anything already sent. Restoring an old application does not undo invoices created in another system.
Use evidence to choose the next stage
Expand when the first boundary is reliable and the team can operate it. Useful evidence includes fewer manual re-entries, shorter delays and exceptions resolved without developer intervention. Document what the pilot taught you about the wider system.
Sometimes the next step is another connection. Sometimes the underlying platform is no longer supportable and replacement becomes necessary. Make that decision with a clearer dependency map, tested data and a team that has already practised change.
