Turn event logs into an evidence-based view of how work really flows.

Choose what one case represents

A case might be an order, batch or another defined unit of work. Choose the key from the question you need to answer. Splits, merges and rework need explicit handling; a reused order number can join unrelated journeys.

Check the event record before drawing paths

For each event, check the case key, activity, timestamp, source and relevant status. Separate the time an activity occurred from the time someone entered or received the record.

Look for missing events, duplicates, clock differences and overwritten states. Keep uncertain intervals visible rather than filling them with a plausible sequence.

Choose event logs or desktop recordings for the question

An analyst uses process mining to follow cases across recorded business events. Task mining examines recorded desktop actions within a person's work. Microsoft distinguishes these two inputs in its Power Automate documentation. Neither input captures every conversation, physical check or reason behind a decision.

Use the route-approval example to see the limit: a system can stamp approval at 11:00 even when someone reviewed the request earlier. A desktop recording may show screen activity, but it cannot establish the unrecorded review time. Ask the people doing the work before treating either record as elapsed task effort.

Before anyone records a desktop, the responsible team must approve the purpose, people, applications, access and retention. Check how the chosen product pauses recording and removes sensitive material. Keep credentials and personal information out of the record. A recording feature does not grant permission to monitor staff.

Discovery, conformance and enhancement use the same record differently

The IEEE Process Mining Manifesto distinguishes three uses. Discovery builds a process model from recorded events. Conformance checking compares a log with an existing model or rule. Enhancement uses the log to improve or extend a model, for example by adding waiting-time information.

Compare routes and waiting

Start with the common paths and the variants associated with a real delay or failure. Compare the same product and case boundary. An unusual path can be a necessary recovery step rather than waste.

Validate a finding at the work

Walk through a few complete cases with operators and planners. Ask what each timestamp represents and which actions are absent. Use the diagram to guide that discussion, not to overrule it.

1. Set the boundary

   Define the case and the start and finish events.

2. Audit the log

   Check identities, ordering, missing events and clock meaning.

3. Inspect the paths

   Find variants and waiting linked to the chosen problem.

4. Confirm the explanation

   Compare the digital path with people and physical records.

5. Test a change

   Measure a specific workflow change against a comparable starting period.

Keep capture quality with the result

Record which systems and manual steps the analysis includes. If event coverage changes, the apparent process can change even when the work does not. Keep that limitation beside any before-and-after comparison.

When process mining is useful, and what its three types mean

Process mining remains useful when a consequential question can be answered from representative event records. It can expose repeated rework or approval waits across many orders. It adds little when manual steps are missing, case identifiers mix unrelated jobs, or a direct observation would answer the question more reliably.

The IEEE Process Mining Manifesto identifies discovery, conformance checking and enhancement as three types of analysis. They are sometimes called pillars, but they are not three required implementation stages. Discover a path from events, compare observed behavior with a stated model, or improve a model with performance information. Data coverage and a useful operating decision matter for all three.

Five process-mining products to evaluate, without a made-up ranking

A practical shortlist can include Celonis, SAP Signavio Process Intelligence, IBM Process Mining, UiPath Process Mining and Microsoft Power Automate Process Mining. Each has current official product documentation. This is a set of examples, not a tested top-five ranking, a free-software list or an endorsement.

Ask each supplier to demonstrate the same anonymized sample: an order with rework, a split batch, a missing event and a corrected timestamp. Check whether the model preserves those relationships, whether a user can trace a result to its records, and whether the output can be exported. Compare data preparation, hosting, access control and the full license cost. A polished sample dashboard cannot answer those questions for your own data.