What does your process data really tell you?

A record tells only part of the story

Process mining reconstructs process flows from events in systems. Task mining examines actions performed while carrying out tasks, for example through recordings. These approaches therefore start with different evidence. First establish which questions existing records can answer before recording screen activity.

A recorded status change is not necessarily a complete description of the work. Someone may call a customer without the conversation appearing as a separate process step. A gap between two records may have several explanations. Treat the map as an evidence-based hypothesis and discuss unusual patterns with the people who perform the work.

Check meaning, relationships and completeness

For conventional process mining, Microsoft requires a case identifier, activity and time information. The person or technical resource performing the activity is optional. Case attributes may use the first or last recorded value. These choices are explained in its process data preparation documentation.

Start with a small sample you can follow from the original record to the report. Record what each check establishes and what remains uncertain. This prevents a polished visualisation from suggesting more certainty than the underlying evidence supports.

On a narrow screen, scroll the table sideways. With a keyboard, focus the table and use the arrow keys.

Keep a record of transformations between source and analysis. Researchers discuss tracing data quality problems across transformation stages in the official EDOC research abstract. Their evaluation concerns manufacturing sensor data; it does not establish a general financial return.

Fictional example: a service request stays open

A fictional service organisation finds that requests sometimes remain open for days. Its first idea is to automate closure. Checking individual records reveals that staff often help customers on the same day but close requests only after receiving confirmation from a supplier.

The measured duration therefore combines customer service, waiting for the supplier and administrative closure. Automatic closure could improve the chart without improving the service. A useful next step is to distinguish these events and investigate which waiting time can actually be reduced. This example shows why the meaning of a measurement needs to be established before acting on it.

Limit the information collected about employees

Explain the purpose, who can access the data and when recordings will be deleted. Use process or team information where possible. Do not collect names or screen images simply because a product supports them.

Microsoft describes pausing recordings, editing input information and removing screenshots in its privacy guidance for process recordings. Check captured images before sharing them. These controls do not automatically make a particular use lawful.

In the Netherlands, the Dutch Data Protection Authority explains that a system may count as an employee monitoring system if it can track employees, even when that capability is not being used. The works council has approval rights over the relevant arrangement. Assess how this applies before introducing the system. Broader responsibilities are covered in automation governance.

Recent development: easier setup still needs checking

Practical do’s and don’ts

• Do: maintain a glossary and check unusual cases against original records.
• Do: make missing information and uncertainty visible in the analysis.
• Do not: treat a successful import as proof that the content is accurate.
• Do not: quietly turn process research into individual performance comparisons.

Frequently asked questions

Do I need a full mining platform immediately?

Not necessarily. A limited export and an examination of sample cases can establish whether the required data exists. Choose tools that fit the question and scale afterwards.

Are screen recordings essential?

No. First determine which questions system records can answer. Consider recordings for specific missing information and assess the privacy implications.

What should I do with a reliably identified bottleneck?

Decide whether it calls for a process change, better data or automation. The guidance on which processes to automate can help you choose the next step.