Segmenting the Analysis

A single map of a whole process averages case types that behave differently. The split is usually more informative than the whole.

The splits worth trying first

Case type or product. Frequently the largest source of variation, and the differences are usually legitimate rather than deviations.

Value band. High-value cases attract extra approvals, and the resulting delay is rarely quantified.

Channel. Cases arriving by different routes often follow entirely different paths.

Region or site, which reveals local practice divergence.

Customer segment, where service levels differ by design or by accident.

Period, which shows whether the process is drifting.

Fast against slow

The most productive single comparison.

Take the fastest quartile and the slowest quartile by cycle time.

Discover each separately and compare.

The differences are the finding: an extra step, a handover, a rework loop, a particular resource, a particular case attribute.

Frequently one attribute explains most of the difference, and it is usually not the one anyone guessed.

Clean against reworked

Cases with no loop against cases with at least one.

Compare cycle time, which gives the cost of rework directly.

Compare the early steps, which is where the cause of the later loop sits.

Compare case attributes, which frequently identifies a predictable subset that will rework.

A predictable rework subset is the strongest finding available, because it can be routed differently before the failure.

Where segmentation misleads

Small segments. A split producing groups of forty cases is measuring noise.

Post hoc segmentation, where you try twenty splits and report the one that looks interesting. Something always does.

Segments defined by the outcome, which guarantees a difference and explains nothing.

Decide the splits you will test before looking, and report how many you tried.

Confounded splits

Region also differs in case mix, staffing and system version.

Channel also differs in customer type.

So a difference between segments is rarely attributable to the segment itself.

List the alternatives before concluding, and test the ones you can by splitting on both at once.

Reporting a segmented finding

Name the segment and its size.

Give both figures and the difference.

State what else differs between the segments, honestly.

Say what it would take to confirm the cause, which is usually asking the people who handle those cases.

A segmented finding presented without its confounders will be dismissed by the first person who knows the operation, and rightly.

Declaring the splits in advance

The discipline that prevents the most common false finding.

Write down which splits you will test, before looking.

Report how many you tried, which is what makes a positive result interpretable.

Something always looks interesting across twenty splits, and reporting only that one is selection rather than analysis.

Where an unplanned split looks striking, treat it as a hypothesis and test it on a different period.

This costs nothing and it is the difference between a finding and a coincidence.

The fast-slow quartile comparison

The single most productive analysis available, and it needs no hypothesis.

Take the fastest and slowest quartiles by cycle time.

Discover each separately.

Compare: extra steps, extra handovers, rework presence, resource pattern, case attributes.

Usually one attribute explains most of the gap, and it is usually not the one anyone guessed.

Then check it is not simply case difficulty, by segmenting on the attribute and re-comparing within segments.

Comparing segments without misleading

Segmentation produces most of the useful findings and most of the unfair comparisons.

Segment by case attribute: customer type, product, value, region, channel.

Check the mix before comparing. Two regions with different product mixes are not comparable on duration.

Normalise or state what differs.

Show the spread, not just the average, because a segment with the same mean and wider spread has a different problem.

Never segment by individual resource for a performance comparison, which is the point at which an operational analysis becomes a personnel matter and stops being cooperated with.