Aggregating the Captures

The deliverable of a task mining study is a description of how a task is performed. Individual traces are working material.

What the output should be

A task model: the steps, their order, the applications involved, the variants, the durations.

Frequencies across participants, not per participant.

The manual steps identified, with the click sequence where automation is the purpose.

The variation between participants, described as approaches rather than attributed to people.

Nothing in the final output should identify anyone.

Getting there

Sessionise the raw events into task instances.

Label the steps, which is manual work and where the analyst's judgement enters.

Merge equivalent steps performed differently by different people.

Count across participants.

Then discard the raw captures, which is the step that is always deferred and should not be.

What to delete and when

Screenshots: as soon as the step labelling is complete.

Keystroke and clipboard detail: immediately, if captured at all.

Window titles containing case or customer data: after sessionisation, retaining only the application.

Participant identifiers: once the cross-participant counts are built.

Set the deletion dates when the study starts, and confirm in writing when they pass.

Handling the variation

Different people will do the same task differently, and that is the most useful finding.

Describe the approaches, not the individuals: "three approaches were observed; the fastest avoids the reconciliation step by exporting once at the end".

Do not name who was fastest, which converts a process finding into a performance comparison.

Ask the participants which approach they would recommend, which frequently produces the standard the process never had.

Showing participants first

Before the output goes anywhere else.

Their own trace, if they want it, which corrects errors no analyst would find.

The aggregate model, with a chance to say what it got wrong.

This is both an obligation in many jurisdictions and the step that makes the model accurate, since participants will identify the steps the labelling missed.

What the aggregate model is for

Filling the gap in the process map, which was the original question.

Automation scoping, where the click sequence is the specification.

Standardising a task, using the approach the participants themselves rate best.

Estimating manual effort, properly, since task mining measures effort where the event log only measures duration.

Not: staffing decisions about the participants, or any comparison between them.

The deletion schedule

Set when the study starts, confirmed when each date passes.

Screenshots: on completion of step labelling.

Window titles containing case or customer references: after sessionisation.

Keystroke and clipboard detail: immediately, if captured at all.

Participant identifiers: once cross-participant counts are built.

Raw event stream: on completion of the task model.

Confirmed in writing to participants, which is both an obligation in several jurisdictions and the thing that makes the next study possible.

Describing approaches, not people

The framing that keeps a variation finding usable.

"Three approaches were observed" rather than three named participants.

Describe what distinguishes each: the order, the tool used, the step skipped.

Report which was fastest as a property of the approach, not of the person.

Ask participants which they would recommend, which frequently produces the standard the process never had and gives them ownership of it.

Never name who was fastest, which converts a process finding into a performance comparison and ends the cooperation.

Aggregating at capture

The design decision that determines how sensitive the deployment is.

Derive the metrics on the endpoint and transmit those, not the raw activity stream.

Application and duration, not keystrokes, wherever the question allows.

Group by role or team before storage.

Set a minimum group size below which no figure is reported.

Discard the raw stream once the derived data exists.

Test the result: attempt to reconstruct an individual's day from what was stored. If you can, the aggregation is nominal.