Rework, waiting time and conformance
Most first analyses produce value through one of four findings. Look for all four; report the ones that hold up.
1. Rework
Section titled “1. Rework”What it is. A case passes through the same activity more than once — an approval bounced back, a form corrected, a decision revisited.
How to find it. Backward arrows in the model. Then quantify: what share of cases loop, and what does looping cost them in elapsed time?
Why it pays. Rework is unambiguous waste and usually has a mundane cause — a required field missing at submission, an unclear threshold, a handover with no checklist. Fixes tend to be cheap.
Sharpen it. Split by whatever segments the population: team, site, order value, customer type. “18% of cases rework” is interesting; “38% in one region against 6% everywhere else” tells you where to look.
2. Waiting time
Section titled “2. Waiting time”What it is. Elapsed time between one activity finishing and the next starting. In most processes this dwarfs actual working time.
How to find it. Switch the model to performance and read the arrows, using median duration — a few pathological cases will otherwise pull the mean somewhere unhelpful. Then look at the mean too: a large gap between the two means the step is usually fine and occasionally terrible, which is a different problem with a different fix.
Why it pays. It reframes the conversation from “our people are slow” to “work sits in a queue for three days”, which is both truer and easier to fix.
3. Batching
Section titled “3. Batching”What it is. Work released in clumps rather than flowing — a nightly job, a weekly committee, a report run each Friday.
How to find it. The Dotted Chart. Vertical stripes are batches; empty bands are nights and weekends.
Why it pays. The fix is scheduling, not redesign. Moving an approval run from weekly to daily can take days out of the median with no change to how anyone works.
4. Conformance
Section titled “4. Conformance”What it is. The difference between what the process is supposed to do and what it does.
How to find it. You need a reference model. Either build one — discover from a clean subset, then edit it to reflect the intended process — or use a model the organization already has. Then run alignment-based conformance with the log and the model selected together.
The result tells you two distinguishable things:
- The log did something the model forbids — the process is deviating from the rule.
- The model expects something the log lacks — the step was skipped, or it is not being logged at all.
Do not report the second as non-compliance before checking which it is. A cluster of “missing” steps on one transition is very often a logging gap, and presenting it as a control failure is how an analysis loses its credibility.
Check every finding twice
Section titled “Check every finding twice”Before it goes in front of anyone:
- Is it a data artifact? Truncated cases, coarse timestamps, a system migration inside your window.
- Does someone who does the work recognize it? Show the finding — not the tool — to a person who performs the process. They will either confirm it in one sentence or explain the boring reason it looks that way. Both save you.