What process mining tells you in the first 90 days

Process mining tools demonstrate beautifully and disappoint reliably, for one reason: the demo starts with clean event data and your organisation does not have any. Here is a ninety-day plan that assumes that.

Days 1-15: pick one process and find its trail

Not a department. One process, end to end, with a clear start event and a clear end event. "Purchase requisition to payment" is a process. "Finance" is not.

Then find where the timestamps live. They are rarely in one place: part in the ERP, part in the ticketing system, part in a mailbox. The output of these two weeks is a list of sources and the join key between them, which is usually the document number and is usually inconsistent.

Days 16-35: extract and be disappointed

The first extraction will be wrong. Cases with no end event. Events out of order because two systems disagree about time zones. Duplicate rows from a retry. This is normal and it is also the first real finding: the gaps tell you where the process leaves the systems entirely and runs on email.

Budget for two extraction cycles. Teams that plan for one spend the difference arguing about whether the tool works.

Days 36-55: conformance before discovery

The temptation is to open the discovery view, see a spaghetti diagram and take a screenshot of it. Resist. Start with the process as documented, and measure conformance against it: how many cases follow the official path, and where exactly do the others leave it?

That question has an owner and an answer. "Here is a complicated picture" does not.

Days 56-75: quantify the deviations

Each deviation gets three numbers: how often, how much time it adds, how often it causes rework. Most will be irrelevant. Two or three will account for the bulk of the cost, and at least one will surprise the people who run the process daily - usually a wait, not a task.

Days 76-90: produce a ranked backlog and a decision

The deliverable at day ninety is not a dashboard. It is a short list of candidates ranked by measured cost, each with a recommendation: automate, fix the policy, or leave alone. Dashboards get admired and closed; a ranked list with recommendations gets argued with, which is what you want.

What derails it

Scope creep into a second process before the first is finished. Waiting for perfect data. And buying a platform in month one - most of this can be done with exported event logs and SQL, and knowing the questions before you buy makes the eventual platform choice far cheaper.