Skip to content

Scoping a first analysis

The most common reason a first process mining analysis disappoints is not the tool or the data. It is scope.

It runs often. Hundreds of cases in the period you can export, ideally thousands. A process that runs eleven times a quarter has no statistical shape to find.

It is mostly in one system. Every extra system multiplies the extraction effort and introduces identifier-matching problems. Order-to-cash in one ERP is a good first project; a customer journey spanning four systems and a call center is a second or third one.

Somebody is unhappy about it. Not because you need a complaint, but because you need someone who will act on the answer. An analysis nobody asked for produces an interesting deck and no change.

Turn “look at the process” into a question

Section titled “Turn “look at the process” into a question”

Vague scope produces vague output. Compare:

Analyze the procurement process.

with:

Why do 15% of purchase orders take more than twenty days from request to approval, when the target is five?

The second names a population, a measure, a threshold and a gap. It tells you which case ID to use, which time window to export, and — critically — what would count as an answer.

Good first questions usually take one of four shapes: why is this slow?, why is this done twice?, why does this skip the control?, or why is this so different across teams/sites/regions?

The case ID defines what “the process” means, so let the question pick it:

  • “How long do customers wait?” → the order.
  • “Where do parcels get stuck?” → the delivery.
  • “Why are invoices paid late?” → the invoice.

If the honest answer is “the interesting part is how orders, deliveries and invoices interact”, note it and carry on with a single case for the first pass. Object-centric analysis handles it properly, and it is a better second analysis than a first one.

Long enough to contain complete cases — several times the typical end-to-end duration — and recent enough that the process has not changed underneath you. If there was a system migration or a reorganization in the window, either exclude it or make comparing before and after the analysis itself.

Before you look at anything, ask the process owner: if we find that X, what would you do? If there is no answer, you are producing a report, not an analysis. Better to find that out now than at the end.

Getting event data out of a system.