ERP integration is usually budgeted as connector work and overruns on everything else: whose definition of a customer wins, which system owns a price, and how much of your business logic lives in customisations nobody documented. Published figures put integration at a third to a half of initial development cost, and ERP is the clearest example of why.
- 01The connector is days. Reconciling what the two systems mean is the project.
- 02Master data ownership has to be decided per entity before any interface is written.
- 03Customisations are where business logic hides, and they do not travel.
- 04Batch versus real time is a business decision disguised as a technical one.
- 05If the integration cannot be reversed step by step, the plan is a hope rather than a migration.
What makes ERP integration harder than it looks?
Not the connection. Most ERPs expose reasonable interfaces and moving records between systems is understood work that can be estimated with confidence.
What cannot be estimated from the outside is the mismatch underneath: two systems holding different definitions of a customer, an order, a unit of measure or a completed job, each correct within its own context and incompatible across the boundary.
Published analysis puts integration work at roughly a third to a half of initial development cost and names it the most frequent source of overrun. ERP is where that shows up most clearly, because the ERP is usually the system with the strongest opinions.
What is master data ownership and why decide it first?
It is the answer to "when these two systems disagree about this field, which one is right". It has to be decided per entity, sometimes per field, and it is an organisational decision rather than a technical one.
Deciding it first is cheap. Discovering it during testing means rework, and discovering it after go live means two systems quietly diverging while everyone assumes they are in sync.
The common failure is deciding it implicitly, through whichever interface was written first. That produces a rule nobody chose and nobody can find.
Why are customisations such a problem?
Because that is where the business logic lives. The workflow rule, the discount exception, the approval step someone added years ago and the report finance depends on are usually not in the standard system, and they are not in the documentation either.
They also do not travel. Logic written in a vendor specific language has to be rebuilt rather than moved, and the first step is working out what it currently does, which is slower than rewriting it.
This is the strongest argument for reading the system rather than interviewing about it. What people describe and what the customisations actually do differ more than anyone expects, and the gap is where the overrun comes from.
Real time or batch?
A business question before a technical one. Real time costs more to build and considerably more to operate, because every failure becomes urgent and every dependency becomes a availability requirement.
Batch is cheaper, more forgiving and sufficient for most reporting and reconciliation. It is insufficient wherever someone acts on the data inside the batch window, and that is the only test that matters.
Mixing them deliberately is usually the right answer, and mixing them accidentally, by making everything real time because it sounded better, is a common way to buy operational burden nobody needed.
How should the rollout be staged?
One entity at a time, each stage reversible, with both systems answering until the new path has been right for long enough to trust.
That is slower on paper than a cutover and it is the difference between a bad afternoon and a bad quarter. If a plan cannot describe how to reverse each step, it is not a migration plan.
Run a reconciliation during the parallel period and look at it. A parallel run nobody checks provides the comfort of a safety net without the safety.
What should be in the business case that usually is not?
The running cost. An integration has a permanent relationship with systems that are upgraded, renamed and replaced without consulting it, and annual maintenance commonly runs 15 to 25 percent of the original build cost.
The cost of the parallel period, which is real and is the price of being able to reverse.
And who owns it in year two. Integrations fail quietly rather than loudly, and an integration with no owner degrades until someone downstream notices a number is wrong, by which point the cause is months old.
We build and integrate the production systems a business runs on: internal platforms, customer facing applications, integrations, and the data layer underneath them.