Scoping a Task Mining Study

Task mining works as a study and fails as a standing capability. The difference is in the scoping.

Start from the gap

Run the process mining first.

Name the specific gap: the four-hour unexplained interval between two events, the step nobody records, the manual reconciliation everyone mentions.

State the question in one sentence. "What happens between order confirmation and picking release, and how much of it is manual?"

If the question cannot be stated, the study should not run. A capability deployed to see what it finds is the definition of disproportionate.

Choose the participants

Volunteers, genuinely, with a real option to decline and no consequence for declining.

Enough for the task to be characterised, which is usually a handful rather than a department.

Representative, including people who do it differently.

Told exactly what is captured and shown their own data.

Coercion defeats the purpose as well as the obligations: a reluctant participant works differently while observed.

Bound the capture

Applications: only those involved in the task, by allow list rather than by exclusion list.

Hours: the working period, not around the clock.

Duration: days or weeks, with a stated end date.

Content: metadata and event sequences where possible, screenshots only where the sequence cannot be reconstructed otherwise.

Exclusions technical: personal applications, email content, browser history outside the allow list.

Every one of these narrows the exposure without narrowing the answer, which is the test of a good scope.

The consent question

Consent from an employee to their employer is frequently not a valid basis in itself, because of the power imbalance.

Which does not make asking pointless. A study with willing participants produces better data and a defensible process.

Rely on a proper basis — a documented legitimate interest with the proportionality assessment, or whatever your jurisdiction requires — and treat participation as voluntary regardless.

Ending it

Stop on the stated date.

Build the task model, then discard the raw captures rather than retaining them "in case".

Show participants the output before it goes anywhere else.

Report what was found and what changed, to them first.

Confirm deletion, in writing.

A study that quietly becomes permanent is the failure mode, and the end date is the only thing that prevents it.

What to refuse

Extension without a new question.

Widening to more people because the tool is already licensed.

Individual comparison of participants.

Retention beyond the study.

Reuse of the captures for a different purpose, which is the drift that turns a study into surveillance.

The allow list

The configuration decision that does most of the work on exposure.

List the applications in scope for the task, and capture only those.

An allow list, not a block list, because a block list misses whatever nobody thought of.

Everything outside it not captured at all, rather than captured and filtered later.

Published to participants, which is what makes the assurance credible.

Reviewed if the task turns out to involve an application nobody listed, by adding it deliberately rather than by widening the capture.

Discarding the first days

Observed work is different work, and the effect fades.

Expect the first days to be unrepresentative: faster, more compliant, fewer workarounds.

Discard them from the analysis and say so.

Extend the window to allow for it, rather than analysing a short study that is entirely novelty effect.

Ask participants when they stopped noticing, which is a better guide than any rule of thumb.

Where the whole study is short enough that the effect dominates, the result describes the documented process, which the organisation already had.

Bounded study, not standing capture

The arrangement that makes task mining defensible.

A defined study period, weeks rather than indefinitely.

A volunteer sample, not a whole department.

A named process, not everything the person does.

Capture limited to the applications in scope, with everything else excluded technically.

Data deleted at the end of the study, with the findings retained rather than the recordings.

A standing deployment across a department produces a surveillance system; a bounded study produces an answer. They use the same software.