Process mining overview

What it does

Process mining reads event logs out of the systems a business actually runs on and reconstructs how a process really executes — every path, every loop, every wait. You get a process map, variant analysis and timing metrics based on what happened rather than on what the process documentation claims happens.

Key facts

The input is an event log: at minimum a case ID, an activity name and a timestamp per event. Everything else the tool produces is derived from those three columns.
Optional attribute types extend the analysis — resource, event-level attributes, case-level attributes interpreted from the first or last event, and financial values per event or per case.
Process mining is included with the Power Automate Premium user licence, which carries 50 MB of process mining data storage per licence, up to 100 GB per tenant. Process and Hosted Process capacity licences don't include it.
The full analytics set is split. Several KPIs, filters and views are marked Premium and need your own Power BI workspace, or a Fabric workspace, connected to the capability. Fabric workspace support is in preview.
A process map shows activities as nodes and transitions as edges, with frequency, performance and rework metrics layered over the top.
The capability sits under Process mining in the maker portal and covers both process mining from event logs and task mining from desktop recordings.

When to use / skip

Worth doing when the client genuinely doesn't know how a process runs — high volume, multiple systems, a lot of manual variation, and a business case that hangs on where the time goes. Skip it when the process is low-volume, when a fortnight of workshops would tell you the same thing more cheaply, or when the event data doesn't exist in a usable form. The tool is good; the data preparation is the project, and clients consistently underestimate it.

Configuration decisions

Which process to start with, and whether the event data for it actually exists before you commit to a scope.
What the case ID is — the choice determines what "one instance of the process" means, and it's the single most consequential decision in the setup.
Whether you run on the default analytics or connect a Power BI or Fabric workspace to get the premium views.
Which optional attributes to bring in, particularly financial ones, since they're what turns "this step is slow" into a number the client cares about.
Storage sizing — 50 MB per Premium licence goes quickly on a real event log.

Gotchas

Getting the event log out of the source system is the hard part. Many applications store current state, not history, and the historical record lives in a specific log table that needs joining to others to get readable IDs and names.
Pick the wrong case ID and the process map is meaningless. Multiple candidate ID columns are common and it's worth trying more than one.
A lot of the analysis people expect from a demo is Premium-tier and needs a Power BI or Fabric workspace wired up first. Demos on the default view underwhelm.
The event log is only as good as the events that got logged. Steps done in email, on paper or in someone's head don't appear, and the map looks cleaner than reality.

Consultant notes

Sell the data extraction as its own phase with its own effort. Every process mining engagement that goes badly went badly at the data stage.
Bring the client's application owners in early — they know where the log table is and you don't. Contacting the IT team that manages the source application is a step in Microsoft's own guidance for a reason.
Set the expectation that the first process map will be an ugly spaghetti diagram and that's a finding, not a failure. Clients who expect their documented flowchart get defensive.

Revisit when the Fabric workspace support leaves preview, or if the storage entitlements change