Key stats
- 3 rework definitions to separate: reopened change, corrected defect, and discarded work
- 2 denominators to report: completed items and reviewed changes
- 4 cause groups: unclear brief, missed context, review gap, and changed requirement
- 1 rule: rework is reviewed by cause and class, never by individual score
Key takeaways
- Define rework precisely before counting it, because the label can cover very different events.
- Report rework against completed items and reviewed changes so the rate has a stable base.
- Attribute the cause across the brief, the build, the review, and the requirement change.
- Expect some rework in discovery-heavy work and do not treat zero as the goal.
- Review the pattern by work class and use it to fix one cause at a time.
Question and scope
This report examines rework as a delivery signal in outsourced software work, where a buyer and a provider share the path from brief to accepted change. The unit is one work item that was completed and then had to be changed, corrected, or redone before or after acceptance. The scope covers defects, reverted changes, and work that was discarded because the brief or the context was wrong. The analysis draws on published guidance about software quality and estimation rather than on a proprietary dataset. It offers a measurement approach, not a benchmark for any particular vendor.
The question is whether rework can tell a buyer something useful about the health of the delivery system without becoming a blunt vendor score. Rework is expensive, visible in the tracker, and easy to count badly. The analysis treats it as a diagnostic that needs a definition, a denominator, and a cause review before it can support a decision. The report therefore separates the count from the judgment and shows where a buyer should ask for context. A number without that context is a conversation starter, not a conclusion.
What counts as rework
Three events are often mixed under one label. A reopened change is one that was reviewed or accepted and then modified for the same requirement. A corrected defect is a failure found after the change was believed complete. Discarded work is effort that was thrown away because the approach or the requirement changed. Each has a different cause and a different remedy.
Write the definition with the same care as a quality rule. State whether a defect found in the same sprint counts, whether a design change requested by the buyer counts, and whether a change that was never merged can be rework. If the definition is vague, two teams will report different rates for the same behavior and the comparison will be meaningless. Publish the definition where both the buyer and the provider can see it, and version it when it changes. A definition that lives only in a slide deck will drift within a quarter.
Denominator and window
The denominator matters as much as the numerator. A rate against completed items shows how often finished work comes back. A rate against reviewed changes shows how often review sends work back before merge. Both are useful, but they answer different questions. Report them separately rather than blending them into one number.
Choose a window that matches the work. A two-week window catches immediate corrections, while a quarterly window catches defects that surface after release. State whether an item can be counted more than once and how a split item is treated. The rule should be simple enough that a delivery lead can apply it without a spreadsheet argument.
Attribution across the handoff
Not all rework belongs to the pod that did the work. A brief may omit an acceptance criterion, a buyer may change a requirement mid-sprint, or a reviewer may approve a change without the context needed to spot a conflict. The cause should name the layer where the gap entered, not the person who happened to touch the ticket last. The review should ask what information was missing and who could have supplied it. That question is more useful than assigning blame after the fact.
A small cause set is enough: unclear brief, missed context, review gap, changed requirement, and environment difference. Each points to a different owner. A cluster of unclear briefs points to intake. A cluster of missed context points to onboarding or documentation. A cluster of changed requirements points to the buyer's decision process, which may be outside the pod's control.
Why blended rates mislead
A single rework rate across a portfolio mixes exploratory work with routine maintenance. Exploratory work has more uncertainty and may legitimately produce more rework. Routine work should produce very little. Blending them lets a healthy exploratory stream hide a weak maintenance stream, or the reverse.
The blended number also hides size effects. A large migration may generate rework out of proportion to its item count, while many tiny fixes generate almost none. Segment by work class and by size band before comparing periods. The goal is to see where rework concentrates, not to produce one number for a dashboard.
Using rework with review evidence
Rework is more useful when read beside review evidence. A low rework rate with very few review comments may mean the review is shallow, not that the work is clean. A higher rate with detailed review findings may mean the team is catching issues earlier. Neither number is good or bad without the other.
Pair the rate with defect escapes, reopened tickets, and the time between completion and correction. A change that comes back within a day is likely a review miss, while one that comes back after a month may be a requirement or environment issue. The timing helps attribute the cause more accurately than the count alone. Also check whether the same cause appears in more than one period. A repeated cause is a process signal, while a one-off may be noise.
Gaming and goodhart risk
Any rework metric tied to consequences will be managed. A team may avoid reopening by creating a new ticket, relabeling a correction as a new requirement, or pressing the buyer to accept a partial fix. These responses are rational when the number becomes a score. The control is an audit of a sample of completed work and a culture that treats a correction as information.
A very low rate deserves the same scrutiny as a high one. It may mean the definition is narrow, the review is weak, or the buyer is not reporting problems. Check a sample of closed items against the timeline. The metric should be plausible in both directions, not only in the direction leadership prefers.
A practical protocol
Write the definition, pick two denominators, and add a cause field at completion. Report the rate by work class and by size band. Review the week's corrections with the pod and the buyer, and sample the text of a few items to check the cause assignment. Keep the fields short enough to fill in seconds.
Choose one cause and fix it. If unclear briefs dominate, improve the intake template. If missed context dominates, add a context note or a pairing block. If changed requirements dominate, separate discovery from delivery and make the change visible. Measure the same fields for a month before judging the fix.
A worked example
A pod completes twenty items in a month and six come back. The blended rate is thirty percent, which looks alarming. The breakdown shows four of the six are small corrections to a migration whose requirements changed twice, and two are review misses on a new area. The rest of the work has no rework.
The team treats the two groups differently. The migration gets a discovery step and a written change rule. The new area gets a named reviewer and a short context note. The next month's rate is lower, but the more important change is that the causes now have owners. The number followed the fix rather than leading it.
Data to collect first
Record the item, the completion date, the correction date, and the cause category. Add the work class, the size band, and whether the correction happened before or after acceptance. Keep the fields short and use a controlled list for the cause so the data can be aggregated. A free-text field is useful for context, but it should not be the only record.
Collect for at least a month before comparing periods. Check a sample of items against the tracker and the review to confirm the cause assignment is consistent. If two reviewers disagree, write a short rule and apply it to the sample again. The quality of the rate depends on the honesty of the classification, not on the size of the dataset.
Counterevidence and limits
Some rework is healthy. A team that never revises a design may be ignoring feedback or shipping a poor fit. Early-stage discovery work should expect changes, and a zero-rework target would punish learning. The analysis recommends understanding the cause, not eliminating every correction.
Rework is also hard to separate from changing requirements in fast-moving products. A buyer may legitimately change direction, and the pod may be blamed for a decision it did not make. Use the cause field and the timeline to keep that distinction visible. Without it, the metric becomes a source of friction rather than a guide to improvement.
Evidence table
| Signal | What to inspect | Owner |
|---|---|---|
| Outcome | Acceptance evidence for the bounded task | Task reviewer |
| Control | Access, test, and approval boundary | Internal owner |
| Handoff | Open risks and next decision | Next owner |
Good distributed work is observable at the handoff: the result, evidence, limitations, and next owner are explicit.
Frequently asked questions
Is a lower rework rate always better?
No. A very low rate can mean the definition is narrow or the review is shallow. Read it beside defect escapes and review evidence.
Should buyer-requested changes count as rework?
Count them, but give them their own cause category so they are not attributed to the pod. The point is to see where the change originated.
How often should rework be reviewed?
Weekly for the correction pattern and monthly for the trend by work class. A quarterly view is useful for contract discussions but too slow for day-to-day fixes.
