“We'll just hook it up to the ERP” is a sentence that makes any experienced project manager sit up. Integrations are the most underestimated part of enterprise development: 20 % of the budget and 60 % of the problems.
The five classic mistakes
1. Nobody looked at data quality — and the ERP holds thirty thousand duplicate customers. 2. A “once a night” sync for a process that needs data now. 3. No outage scenario: the ERP goes down for maintenance and orders fall into a black hole. 4. An integration without an owner — a code list changes and for three months nobody knows why prices do not match. 5. Point-to-point spaghetti: the fifth system connected to the previous four means twenty interfaces to maintain.
How to do it properly
An integration layer (API gateway or middleware) instead of spaghetti, a message queue for outage resilience, monitoring with alerts, and a contractually named owner for every interface. Sounds more expensive — until you price your first weekend invoicing outage.
Idempotency: the word that saves your weekend
The most common silent integration bug: the same order gets written twice on a retry. The fix is idempotency — every operation has a unique key, so repeating it breaks nothing. It sounds like a developer detail, yet it decides whether a network glitch leaves you with order or with double invoices. Ask your vendor how they handle repeated and delayed messages.
How to test an integration before it burns
An integration tested only on “nice” data falls over in its first real month. So we test the ugly scenarios too: an ERP outage, incomplete data, a changed code list, double the load. Far better to find the problem in staging than on a Friday night during close. Resilience is not built after an incident — it is built before one.
Our configurator charges +8 % per integration. Now you know what is inside those percent: not optimism, but twenty years of other people's mistakes.