Every improvement claim rests on a baseline. Most baselines are too short, taken at the wrong time, and measured differently from the follow-up.
How long
Long enough to contain the normal variation, which is several weeks at minimum in most operations.
Covering a full demand cycle: weekly pattern, and monthly if one exists.
Not covering an unusual period — a peak, a system migration, a major absence — unless that is what you are measuring.
Not one good week, and not one bad one.
If the baseline is a single week, any subsequent comparison is measuring ordinary variation and will show whatever the analyst hopes.
What to capture
The measure itself, obviously.
The spread, not just the average.
The work content: volume, mix, order profile.
Staffing, by shift and by experience.
Conditions: equipment availability, temperature, anything unusual.
Exactly how it was computed, in enough detail to repeat it identically.
The last point is the one most often lost, and it makes the follow-up incomparable.
The measurement effect
Performance changes when measurement starts, for reasons unrelated to any intervention.
Which contaminates a baseline taken immediately after deployment.
Allow a settling period before the baseline period begins.
Expect an initial improvement that decays, and do not attribute it to anything.
Documenting it
A short written record: period, measure definition, filters and exclusions, conditions, and the figures with their spread.
Kept somewhere findable, because the comparison will be made a year later by someone else.
Including the exclusion rules, which if changed between baseline and follow-up invalidate the comparison silently.
The comparison
Same measure, same definition, same filters, same period length.
Same season where possible, or normalise for it.
Report the spread alongside the average.
State whether the difference exceeds ordinary variation, which most claimed improvements do not.
Show the whole series, not two numbers, so the reader can see whether the change coincided with the intervention or preceded it.
Where before-and-after is not available
Sometimes the intervention has already happened.
Use a comparison area — another zone, another shift, another site running unchanged — as the control.
Use the historical series to establish what would have been expected without the change.
State the weakness honestly. An uncontrolled before-and-after comparison is weak evidence, and presenting it as strong damages the credibility of the next one.
The honest claim
"Waiting time at packing fell from a median of thirty-eight minutes to twenty-two, over comparable eight-week periods with similar volume and mix, which is outside the normal range for this measure."
Specific, checkable, and it names the conditions.
Not: "productivity improved fifteen percent", which nobody can verify and which will be quoted without the caveats until someone checks it.
Writing the baseline record
A short document that makes the comparison possible a year later.
Period covered, with dates.
Measure definition, precisely, including the counting point.
Filters and exclusion rules, verbatim.
Conditions: volume, mix, staffing, equipment, anything unusual.
The figures, with the spread, not just the average.
Who produced it.
Stored somewhere findable, because the person making the comparison will not be you.
Using a control area
The strongest available design when a randomised comparison is impossible.
Pick a comparable zone, shift or site running unchanged.
Measure both, before and after.
The difference in the differences is the effect, and it removes seasonality, demand changes and the novelty effect at once.
Choose the control before the intervention, not afterwards, which is what prevents selecting a flattering comparison.
Report both series, so a reader can see whether the control moved too.