Task mining

What it does

Task mining captures what people do on their desktops — clicks, keystrokes, applications, screenshots — using the Power Automate desktop recorder, then turns those recordings into a process map. Where process mining reads system event logs, task mining watches the human.

Key facts

Recordings are made with the Power Automate recorder, launched from the process details page in the maker portal. Power Automate for desktop has to be installed on the machine.
The recorder supports pause, resume, deleting individual actions and resetting the whole recording before you save.
Raw actions get grouped into named activities before analysis. The capability auto-groups similar actions, and you can edit, add and delete activity groupings by hand.
You need at least two activities for a process map to mean anything.
Process owners can define recommended activity names for a process so recordings from different people group consistently. Names outside that list appear as custom names.
Screenshots and text captured during recording can be deleted individually before analysis — that's the mechanism for stripping sensitive data.
Output includes a process map, variant analysis, and an application analytics report showing where time is spent by application.
Automation recommendations surface where recorded actions map to available Power Automate connectors, and can generate a draft flow.

When to use / skip

Task mining earns its keep when the work happens in the user interface and leaves no system trace — copy-paste between systems, spreadsheet manipulation, legacy green-screen apps, anything a person does that the database never sees. Skip it when the process is already visible in event logs, because process mining gives you thousands of cases where task mining gives you however many recordings you can persuade people to make. The two answer different questions and the honest answer is usually that you want both.

Configuration decisions

Who records, how many of them, and how many recordings each — the sample is the study, and three recordings from one enthusiastic person isn't a sample.
Whether recommended activity names are defined up front, which they should be if more than one person is recording.
What the consent and communication approach is before anyone starts recording their screen.
What gets scrubbed from recordings and who checks — screenshot and text removal is manual, per step.
Whether the output feeds a desktop flow, a cloud flow, or a redesign of the process rather than an automation.

Gotchas

The recorder captures screenshots of whatever is on screen, including the customer record, the payroll figure and the browser tab nobody meant to have open. Removing them is a manual step someone has to actually do.
Recording quality is entirely down to the recorder's discipline. Stray clicks, mis-selections and interruptions all land in the data unless deleted.
Actions can take a while to appear after a long recording finishes. People assume it failed and record again.
Auto-grouping is a starting point, not an answer. Left unedited, the activity names it produces make the process map hard to read.
Sample sizes are small by nature, so variant analysis from task mining tells you about the people who recorded, not about the population.

Consultant notes

The consent conversation comes before the tooling conversation. Screen recording of employees touches works councils, unions and data protection in a way process mining doesn't — in some jurisdictions it's a formal consultation, not a courtesy email.
Be explicit with the client that this is observational research with a small n. Presenting task mining variants as if they were statistically representative is where credibility gets lost.
Recruit recorders who do the job properly and one who does it badly. The gap between them is usually the most valuable finding in the whole exercise.

Recheck the privacy and data-removal guidance before any engagement involving employee recording — the rules move faster than the product does