The Skills a Programme Needs
Tool training is the easiest part and the most often recruited for. What actually determines whether an analysis produces a change.
Reference
Process mining capability is usually built by sending people on a tool course. That is the smallest of the five things needed.
SQL and data handling
The largest gap in practice.
Building the log is a join and a group-by, and doing it correctly requires understanding the source schema.
Checking it requires profiling: distributions, outliers, counts.
Without this, the analyst is dependent on someone else for every question, which is what makes exploration stop.
Process understanding
What distinguishes a finding from an artefact.
Knowing that a step is batched nightly, or that one team handles exceptions, or that a code is legacy, prevents most wrong conclusions.
Acquired by talking to the operation, not from documentation.
An analyst who has never watched the process work will produce technically correct findings that get rejected.
Statistical caution
Understanding ordinary variation, so noise is not reported as change.
Knowing that a segmented comparison is confounded until the confounders are listed.
Knowing that testing twenty splits guarantees one looks interesting.
Reporting spread alongside every average.
This is the skill whose absence produces the most damaging errors, because they are confident and plausible.
The relationship with the operation
Validating findings before reporting them.
Asking why, and being able to hear an answer that contradicts the analysis.
Being trusted enough that people say what actually happens.
A relationship, not a technique, and it takes months.
Tool skills
Last, and easiest.
A competent analyst learns a mining platform in days.
A tool expert without the other four produces maps.
How to build it
Start with someone who knows the process and teach them SQL, rather than the reverse. The process knowledge takes far longer to acquire.
Pair them with a data engineer for the extraction, permanently rather than for the project.
Do the first analysis manually, which teaches more than any course.
Have findings reviewed by someone who will challenge them, which is where the statistical caution develops.
What to look for when recruiting
Ask for an example of a finding they reported that turned out to be wrong, and what they did.
Ask how they would check whether a slow step is slow to do or slow to reach.
Ask what they would do about same-second timestamps.
Ask how they would validate a case identifier.
These four questions distinguish someone who has done this from someone who has been trained on it, and the difference is most of the value.
Building the capability internally
The sequence that works, and it is the opposite of the usual one.
Start with someone who knows the process and teach them SQL. Process knowledge takes far longer to acquire than query skills.
Pair them with a data engineer permanently, not for the project.
Do the first analysis manually, which teaches every decision explicitly.
Have findings challenged by someone who will push back, which is where statistical caution develops.
Send them on the tool course last, when they know what they need from it.
The four interview questions
They distinguish someone who has done this from someone trained on it.
How would you check whether a slow step is slow to do or slow to reach?
What would you do about events sharing a timestamp?
How would you validate a case identifier?
Tell me about a finding you reported that turned out to be wrong.
The last one matters most. Anyone who has done this work at length has one, and the answer reveals whether they check their findings or defend them.
What the role actually requires
The job is less analytical and more about data plumbing and persuasion than most descriptions suggest.
SQL, competently. Most of the work is extraction and joins.
Understanding of the source systems, which is domain knowledge rather than technique.
Enough statistics to distinguish signal from ordinary variation, and no more.
The ability to present a finding to people it implicates without them becoming defensive.
Willingness to report a null result.
Hire for the last two. The tooling is learnable in weeks; the credibility to get a finding acted on is not.