# Process Variant Analysis: Separate Useful Exceptions From Rework

<a id="main-content"></a>

Operations &amp; Systems

October 28, 2026  ·  9 min read  ·  Victor Valentine Romo

On this page

1. [Fix the Case Boundary and Selection Rule](<#fix-the-case-boundary-and-selection-rule>)
2. [Group the Log Into Three Explicit Sequences](<#group-the-log-into-three-explicit-sequences>)
3. [Keep Case Frequency Separate From Event Count](<#keep-case-frequency-separate-from-event-count>)
4. [Check the Business Reason for an Exception](<#check-the-business-reason-for-an-exception>)
5. [Treat a Repeated Review as a Question](<#treat-a-repeated-review-as-a-question>)
6. [Separate Elapsed Duration From Active Work](<#separate-elapsed-duration-from-active-work>)
7. [Test the Record Before Labeling the Process Defective](<#test-the-record-before-labeling-the-process-defective>)
8. [Use a Review Worksheet That Preserves Uncertainty](<#use-a-review-worksheet-that-preserves-uncertainty>)
9. [Close With an Evidence-Bounded Finding](<#close-with-an-evidence-bounded-finding>)
10. [Sources](<#sources>)

One service request passes through review once. Another needs specialist approval. A third returns for correction and passes through review again.

All three reach completion. Their paths deserve different questions.

Process variant analysis groups recorded paths so you can inspect those differences. Microsoft defines a variant by activity sequence: cases with the same activities in the same order share a variant. [Microsoft: process variants](<https://learn.microsoft.com/en-us/power-automate/minit/variants>).

The exercise below uses twelve fictional service requests. It assumes a readiness review has established their case identities and event order. The worksheet, [business rules](</articles/business-decision-table-gaps-overlaps/>), and timing examples are proposed teaching material. They describe no real company or employee.

<a id="fix-the-case-boundary-and-selection-rule"></a>

## Fix the Case Boundary and Selection Rule

For this exercise, one case means one service request [from receipt through completion](</articles/service-blueprint-small-b2b-team/>). A correction before completion remains in that case. A later, separately submitted request begins another case.

Use a synthetic sample containing twelve completed requests and no other cases. Each has its full recorded history. Each appears once in the case list. These assumptions make the arithmetic below reproducible; they do not describe a typical business.

State the same boundaries in your own review. If your sample [excludes open requests](</articles/work-item-aging-small-service-teams/>), do not present its results as a description of every request received. If a filter removes activities, label the result as a filtered sequence.

Keep the definition stable while comparing paths. Changing [from request IDs to company IDs](</articles/join-cardinality-business-reports/>) would change what each path represents. A long path could then combine unrelated requests.

Use an [operations automation audit](</articles/operations-automation-audit/>) to choose a process worth examining. Here, hold the selection decision fixed. The task is to interpret recorded sequences before proposing any change to how work runs.

Write the review question above the case list: which additional steps serve a documented need, and which repeat work after a correction?

<a id="group-the-log-into-three-explicit-sequences"></a>

## Group the Log Into Three Explicit Sequences

Read each case's events in their established order. Retain repeated activities. Give each distinct sequence a local label for the worksheet.

The fictional log contains exactly these sequences:

- **Path A:** Received → Reviewed → Completed.
- **Path B:** Received → Reviewed → Specialist approval → Completed.
- **Path C:** Received → Reviewed → Correction requested → Revised → Reviewed → Completed.

Path C includes Reviewed twice. Both occurrences remain visible. Removing the second occurrence would change the path being studied.

The following table specifies every case in the synthetic log. Each identifier in a row has the exact sequence above, with no extra events.

| Path | Case identifiers | Cases | Events per case | Total events |
| --- | --- | --- | --- | --- |
| A | SR-201, SR-202, SR-203, SR-204, SR-205, SR-206 | 6 | 3 | 18 |
| B | SR-207, SR-208, SR-209 | 3 | 4 | 12 |
| C | SR-210, SR-211, SR-212 | 3 | 6 | 18 |
| Total | Twelve distinct request IDs | 12 | Varies | 48 |

This compact specification is enough to reproduce the grouping. Expand each identifier into its stated sequence, then count the rows. It does not provide timestamps for every case, so it cannot support [an average duration](</articles/mean-median-business-reports/>) for each path.

Power Automate can group variants using a selected mining attribute, and its variant selection is independent of the process map's attribute selection. Confirm that you are comparing activity sequences when using that view. [Microsoft: variant mining attribute](<https://learn.microsoft.com/en-us/power-automate/minit/variants#variant-mining-attribute>).

<a id="keep-case-frequency-separate-from-event-count"></a>

## Keep Case Frequency Separate From Event Count

Path A appears in six of twelve cases: fifty percent. Paths B and C each appear in three of twelve cases: twenty-five percent each.

Now count events. Path A contributes eighteen of forty-eight events, or thirty-seven and a half percent. Path B contributes twelve, or twenty-five percent. Path C contributes eighteen, or thirty-seven and a half percent.

The arithmetic describes this synthetic log only. It estimates no broader population and predicts no business outcome.

Paths A and C have the same event total. Path C occurs in half as many cases. A chart sorted by event volume can therefore emphasize a different path from a chart sorted by case frequency.

Microsoft's variant view offers separate sorting by case count and event count. Read the selected measure before interpreting its order. [Microsoft: variant attribute panel](<https://learn.microsoft.com/en-us/power-automate/minit/variants#attribute-panel>).

Write both [denominators](</articles/two-way-tables-business-reporting/>) beside your findings. “Three cases follow Path C” identifies the affected requests. “Those cases contain eighteen events” describes their recorded steps. Neither statement establishes how much labor those events require.

Keep the case list available for inspection. If a filter changes the sample to nine cases, recompute the shares using nine. Do not leave percentages from the original twelve-case review attached to the smaller view.

<a id="check-the-business-reason-for-an-exception"></a>

## Check the Business Reason for an Exception

For the fictional service, a documented rule requires specialist approval for requests involving a custom specification. SR-207, SR-208, and SR-209 meet that condition. Their records contain the corresponding approval evidence.

Under those assumptions, Path B is a legitimate exception to the shorter path. Its extra step serves an identified requirement. Calling it waste because it contains another activity would ignore that requirement.

Record the rule, its applicable period, and the evidence connecting each case to it. A general statement that some requests need approval is weaker than a case-specific explanation.

Now change the example for the remaining review. SR-209 retains its approval event, but its record no longer establishes the custom-specification condition. Hold that case unresolved. The other two cases do not prove that SR-209 needs the same step. The worksheet below uses this revised evidence state.

Use [SOP documentation](</articles/sop-documentation-b2b/>) to locate the intended rule. Compare its effective version with the case date. A later-written procedure needs supporting history to establish which requirement applied when the request was handled.

Keep the business decision with the person who owns the rule. The analyst can identify a mismatch between the rule and the record. The analyst should not invent an exception policy from the shape of a diagram.

<a id="treat-a-repeated-review-as-a-question"></a>

## Treat a Repeated Review as a Question

Path C contains a return through review. In the fictional case notes for SR-210 and SR-211, the first review identifies an omitted specification field. A revision supplies it, and the second review accepts it.

That supporting evidence justifies describing those two cases as correction-related rework. The repeated Reviewed label alone would not establish the reason.

For SR-212, leave the correction reason missing. Its sequence matches Path C, but its explanation remains unknown. A shared sequence does not make every underlying business circumstance identical.

Ask the process owner what evidence distinguishes avoidable correction from a required iterative review. Ask whether the original request could contain the needed information at submission. Keep the answer attached to the specific cases examined.

[Do not assign cause](</articles/incident-postmortem-small-service-team/>) from a role label. A review event associated with one department does not establish that department created the omission. The missing field may depend on another party, an unclear request, or a later scope change.

Record plausible explanations as questions. Mark an explanation confirmed only when the supporting record answers it. Avoid turning a diagram into an employee ranking or an account of someone's effort.

<a id="separate-elapsed-duration-from-active-work"></a>

## Separate Elapsed Duration From Active Work

A single event timestamp does not supply an activity interval. Power Automate's data guidance distinguishes [event logs](</articles/process-mining-event-log-readiness/>) from activity logs with start and end timestamps. [Microsoft: timestamp fields](<https://learn.microsoft.com/en-us/power-automate/process-mining-processes-and-data#data-requirements>).

Give SR-210 one additional synthetic timing record. All times below occur on the same fictional day in UTC. Received is a milestone at 09:00. Completed is a milestone at 12:00. The intervals explicitly represent uninterrupted work in this example.

| Recorded interval | Start | End | Duration |
| --- | --- | --- | --- |
| First review | 09:10 | 09:30 | 20 minutes |
| Revision | 10:30 | 11:00 | 30 minutes |
| Second review | 11:30 | 11:45 | 15 minutes |

The case spans one hundred eighty minutes from receipt to completion. The three nonoverlapping work intervals total sixty-five minutes. Subtracting them leaves one hundred fifteen minutes outside those intervals.

Call that remainder unclassified until evidence explains it. Suppose the fictional case history separately confirms queue waiting from 09:30 to 10:30 and 11:00 to 11:30. Those two intervals account for ninety minutes of waiting. The other twenty-five minutes remain unclassified.

Do not label the entire remainder idle time. Do not assign it to a person. If intervals overlap in another case, adding their durations would also double-count part of the elapsed timeline.

Keep this one-case calculation separate from the path totals. One timed case cannot establish Path C's average work or waiting duration. If your source has only milestone times, report elapsed intervals and leave active work unknown.

<a id="test-the-record-before-labeling-the-process-defective"></a>

## Test the Record Before Labeling the Process Defective

Suppose a later export produces Received → Completed for SR-204. Before announcing that review was skipped, inspect its source history and the export settings.

Check whether Reviewed changed names during the period. Check whether an activity filter removed it. Check whether two systems use different case identifiers. Check whether the case began before the export window.

These are investigation steps, not explanations established by the suspicious path. Microsoft notes that application records may omit historical events of interest. Missing evidence remains a coverage limit. [Microsoft: log history](<https://learn.microsoft.com/en-us/power-automate/process-mining-processes-and-data#where-to-get-log-data-from-your-application>).

If the source contains the review event, correct the sample through [a documented transformation](</articles/data-lineage-small-business-reports/>) and repeat the grouping. If the source cannot resolve the question, mark the case uncertain. Do not insert a review event merely because the intended procedure requires one.

Also inspect tied [timestamps](</articles/event-time-processing-time-business-reports/>). If their precision cannot establish order, the exact sequence needs qualification. A path created by an arbitrary tie-breaking choice is insufficient evidence of a process defect.

Keep the original observation, the evidence checked, and the revised interpretation together. A later reviewer should see why your conclusion changes.

<a id="use-a-review-worksheet-that-preserves-uncertainty"></a>

## Use a Review Worksheet That Preserves Uncertainty

Use one row per finding that needs a different explanation. The proposed worksheet below separates observation from interpretation and a question from its answer.

| Observed sequence | Business context | Evidence gap | Owner question |
| --- | --- | --- | --- |
| Path A, six cases | Ordinary request path in this sample | No timing records for these six cases | Which cases need deeper timing review? |
| Path B, SR-207 and SR-208 | Custom specification requires approval | None in the fictional rule check | Does the recorded approval satisfy the applicable rule? |
| Path B, SR-209 | Same extra activity as the other B cases | Qualifying condition absent | What record establishes why approval is required? |
| Path C, SR-210 and SR-211 | Missing field corrected after first review | Upstream reason for omission | Is the field available and clearly requested at intake? |
| Path C, SR-212 | Repeated review follows revision | Correction reason absent | What change makes the second review necessary? |

The worksheet now distinguishes sequence membership from confirmed business context. Path B still contains three cases, even though one case's justification is unresolved. Path C still contains three cases, even though only two have a documented correction reason.

Keep those counts intact. Excluding uncertain cases silently would make the review appear [more conclusive than its evidence permits](</principles/>).

For questions about assignment history, the [dispatch system article](</articles/dispatch-system-client-management/>) provides adjacent workflow context. This analysis stops at identifying the question and its evidence. It does not redesign routing or introduce another queue.

<a id="close-with-an-evidence-bounded-finding"></a>

## Close With an Evidence-Bounded Finding

Write the finding in terms another person can check: twelve synthetic cases follow three sequences; two cases have confirmed correction-related rework; one matching case lacks its reason.

Then state what remains unknown. The example has one timed case, so it supports no average duration by variant. Missing business context prevents a complete judgment about every additional step.

Choose a next review action that resolves one of those gaps. Inspect the missing condition for SR-209. Ask the source owner about SR-212's correction reason. Keep proposed [process changes](</articles/workflow-change-impact-assessment/>) separate until the necessary evidence exists.

A frequent path can still contain a defect. An uncommon path can serve a valid requirement. Frequency tells you where to inspect; supporting records explain what you find.

Keep the exception when its purpose holds. Investigate the repetition when its reason is missing.

<a id="sources"></a>

## Sources

- [Microsoft Learn: Analyze processes with variants](<https://learn.microsoft.com/en-us/power-automate/minit/variants>). Variant definitions, mining attributes, and separate count measures. Checked 2026.10.28.
- [Microsoft Learn: Prepare processes and data](<https://learn.microsoft.com/en-us/power-automate/process-mining-processes-and-data>). Timestamp types and limits of source history. Checked 2026.10.28.
