Front-to-Back Process Review

1,434 cases · 8,577 events · 27 activities · 10 work groups

A process-mining review of a real case-handling workflow — variants, rework, handoffs, queue time, service levels and automation screening. The headline is that almost none of it matters as much as it looks like it does.

59.1d
mean cycle time
9%
of it inside the logged workflow
29.3%
of cases miss the committed date
3.79
handoffs per case
50%
on the single commonest path

Where the 59 days actually go

Intake 38% Worked 9% Closure 52%

Everything a conventional process review produces — variant maps, handoff diagrams, queue times, rework counts — describes the middle band. Remove every queue, handoff delay and rework loop in the logged workflow and mean cycle time falls from 59.1 to 53.6 days. The other 91% of the clock is not instrumented at all.

ScopeCasesCycle (d)IntakeWorkedClosure% worked
all channels1,32759.122.65.530.99.4%
Internet1,14657.9225.330.59.2%
Desk10768.825.75.937.18.6%
Post5259.426.37.725.412.9%
e-mail2176.232.210.833.214.1%
Intern122.613.418.34.3%

The pattern holds in every channel. It is structural, not a quirk of one route.

How this nearly went wrong

The bottleneck analysis came out clean: one transition accounts for 2,422 days of waiting across 791 occurrences — about 31% of all queue time. The obvious thing to fix. Then the arithmetic refused to close: the gaps between logged steps total roughly 5.4 days per case against a 59-day cycle. The bottleneck is real, and it is 31% of a window worth 9% of the total — about 1.8 days of the 59. A programme that opened with it would have reported a real improvement and moved the number almost not at all.

FromToCountMean (d)Total days
T05 Print and send confirmation of receiptT06 Determine necessity of stop advice7913.12,422
Confirmation of receiptT02 Check confirmation of receipt1,0790.8901
T04 Determine confirmation of receiptT05 Print and send confirmation of receipt1,1770.6758
T10 Determine necessity to stop indicationT02 Check confirmation of receipt1553.1475
T06 Determine necessity of stop adviceT10 Determine necessity to stop indication1,1650.4440
T02 Check confirmation of receiptT04 Determine confirmation of receipt1,1190.3339
Confirmation of receiptT06 Determine necessity of stop advice2391.1258
T10 Determine necessity to stop indicationT05 Print and send confirmation of receipt1022.5255

Service levels

ChannelCasesMean cycle (d)Breach rateMean days late
Internet1,14657.928.2%30.3
Desk10768.838.3%35.5
Post5259.432.7%15.6
e-mail2176.238.1%70.5
Intern122.60.0%0

The highest-volume channel performs best. e-mail is worst on both counts — small volume, but when it breaches it breaches by more than two months.

Steps in caseCasesMean cycle (d)Breach rate
a. 1-4 steps11936.215.1%
b. 5-6 steps1,0686029.4%
c. 7-9 steps4892.156.2%
d. 10+ steps9260.532.6%

Cases with 7–9 steps breach at 56.3% against 15.1% for 1–4. Usable, because step count is visible on day three while cycle time is only knowable at the end. Not monotone — 10+ step cases breach less than 7–9 — so it is a flag worth testing, not a scoring model.

Rework and handoffs

ActivityCasesWith repeat% repeatingWorstWasted
T06 Determine necessity of stop advice1,309594.5%10107
T02 Check confirmation of receipt1,316352.7%652
T03 Adjust confirmation of receipt371129.7%518
T07-2 Draft intern advice aspect 215853.3%617
T07-1 Draft intern advice aspect 129827.6%916
T07-5 Draft intern advice aspect 516531.2%611
T17 Check report Y to stop indication20210.0%56
T18 Adjust report Y to stop indicition22100.0%44

Highlighted rows are the process's own explicit correction steps. One of them repeats within a case up to five times — a correction that itself needs correcting.

FromToHandoffsMean wait (h)Max wait (h)
Group 3Group 299211.6314.5
Group 4Group 38571138.9
Group 2Group 180073.76,455.3
EMPTYGroup 47120.7166.6
EMPTYGroup 142919.12,397.4
Group 1Group 433611.9917.7
Group 1EMPTY321112.62,663.4
EMPTYGroup 317119.1172.5

The slowest handoff averages 73.7 hours, with a worst case of 6,455 hours — 269 days on a single handoff. "EMPTY" is the log's own value where no group was recorded; kept rather than dropped.

Automation screening

ActivityExecutionsPeopleMean hoursCoVScore
T05 Print and send confirmation of receipt1,3003319.12.7353
T04 Determine confirmation of receipt1,3073710.24.9222
T02 Check confirmation of receipt1,3684027.95.3217
T06 Determine necessity of stop advice1,4163451.85.5217
T10 Determine necessity to stop indication1,2833411.36.2179
T03 Adjust confirmation of receipt5517129.72.516

Frequent, consistent and quick steps score highest — a ranking, not a decision. The top candidate is executed 1,300 times by 33 people, and is also the source of the biggest queue in the process.

What was excluded

CheckCasesTreatment
Still open at extract105no end date, so no cycle time
End date before start date2impossible; excluded, not clamped to zero
Committed date differs from deadline37two fields disagree; SLA uses the committed date
Usable for cycle time1,327the denominator behind every timing figure

Real-life event log, publicly released for research. A public-sector case-handling process, not a banking one — the domain differs, the method does not. No cost or headcount data, so benefits are stated in days rather than money. · SQL, workbook and memo on GitHub