A release slips, and the client notices. Ah, of course… The retrospective produces a familiar conclusion: “Communication broke down.”
An action is assigned to “improve communication,” and work continues “as planned”.
Months later, the same failure returns, perhaps under a new action item.
Maybe you are familiar with this situation… Well, there is something curious about this ritual.
Organizations can document problems without learning to prevent them (and the documentation can look impressive). Trouble begins when a description of events becomes an explanation of causes.
“Poor communication” might be part of the story, of course.
But which information was missing, why was it missing, and what made its absence capable of delaying a release?
Those questions are less comfortable and more useful.
Today I want to talk about a disciplined root cause analysis that can be as easy as four connected moves.
“The release was a mess” may capture the mood of a retrospective, but it offers little direction for an investigation.
First establish what happened and, perhaps more interestingly, what did not. Charles Kepner and Benjamin Tregoe developed an approach often called Is / Is Not analysis. It compares a problem across what, where, when, and extent, looking for differences between situations where the failure appears and situations where it does not.
Consider a payment service that fails during Friday production deployments but works in staging and survives releases earlier in the week. Those distinctions narrow the investigation.
Friday, production, or the payment service deserves attention. None is necessarily the cause, but the team now has a question it can investigate.
Which raises another question: how often do retrospectives begin with proposed solutions before anyone has agreed on the exact problem?
The Five Whys, associated with Taiichi Ohno and Toyota, is attractive because it is simple. Ask why repeatedly until the underlying conditions become clearer. The trouble is that simplicity also makes the method easy to rush.
Why did the service fail? A configuration was incorrect.
Why? An engineer entered the wrong value.
Why? Lack of attention.
The investigation ends, and the corrective action practically writes itself: remind the engineer to be careful. But this explanation depends on an assumption about the person, and it leaves the process untouched.
A different investigation might discover that Friday releases require a manual configuration step, that automation has remained unfinished in the backlog, and that nobody has clear ownership of the deployment pipeline.
That sounds more promising, but it is still a hypothesis (a convincing story is not the same thing as evidence). Deployment logs, process comparisons, and the people who do the work need to confirm it. An Ishikawa fishbone diagram can organize competing explanations, and Pareto analysis can prioritize recurring problems.
Once the team has a candidate explanation, ask what would have happened if that condition had not existed. Would the release still have failed?
This removal test is a useful challenge to the reasoning, rather than a formal proof.
Many failures have several contributing conditions, and removing one may reduce the risk without making the problem impossible.
James Reason’s Swiss cheese model helps here, explaining that a serious failure can pass through several imperfect defenses when their weaknesses line up.
So identifying an incorrect configuration does not explain why the validation check missed it, why the process depended on manual work, or why that dependency remained unresolved.
There is an uncomfortable management question inside all this. When “human error” ends an investigation, has the organization found the cause, or simply found the person closest to the failure?
The distinction matters because someone can be responsible for a mistake without being responsible for every condition that made it possible.
Finding a plausible cause is only part of the job. The organization must decide what to change, who will change it, and how anyone will know the change worked.
Safety engineering’s hierarchy of controls offers a useful principle: removing a risky condition or building protection into the process is generally stronger than asking people to remember another instruction. For the deployment example, that might mean eliminating the manual step, automating validation, and assigning ownership of the pipeline. Training still has a place, but it should not carry the entire burden.
And then comes the part that tends to disappear from the retrospective agenda: checking the result.
Was the action completed?
Did the automated check catch errors?
Did Friday releases actually become more reliable?
The next time a retrospective ends with “improve communication,” “be more careful,” or “follow the process,” resist the temptation to close the discussion too quickly.
Ask what, specifically, will change in the way the team works, who can make that change happen, and what evidence would show that the problem has actually been addressed.
And perhaps there is another question worth asking…
How many of the recurring problems in your organization have already been investigated, documented, discussed, and assigned to someone, sometimes more than once?
A completed action item is useful, of course, but it tells us surprisingly little about whether the underlying problem has disappeared (spreadsheets are quite good at showing progress, even when reality disagrees).
Before your next retrospective, revisit the lessons learned from the previous one. Choose one recurring issue and examine what actually changed since it was recorded. If the same conditions are still present, there is probably more work to do.
Because six months from now, when the team faces another difficult release, the quality of that retrospective will be measured by what happens differently, not by how well everyone explained what went wrong.
An action is assigned to “improve communication,” and work continues “as planned”.
Months later, the same failure returns, perhaps under a new action item.
Maybe you are familiar with this situation… Well, there is something curious about this ritual.
Organizations can document problems without learning to prevent them (and the documentation can look impressive). Trouble begins when a description of events becomes an explanation of causes.
“Poor communication” might be part of the story, of course.
But which information was missing, why was it missing, and what made its absence capable of delaying a release?
Those questions are less comfortable and more useful.
Today I want to talk about a disciplined root cause analysis that can be as easy as four connected moves.
1. Define what actually failed
“The release was a mess” may capture the mood of a retrospective, but it offers little direction for an investigation.
First establish what happened and, perhaps more interestingly, what did not. Charles Kepner and Benjamin Tregoe developed an approach often called Is / Is Not analysis. It compares a problem across what, where, when, and extent, looking for differences between situations where the failure appears and situations where it does not.
Consider a payment service that fails during Friday production deployments but works in staging and survives releases earlier in the week. Those distinctions narrow the investigation.
Friday, production, or the payment service deserves attention. None is necessarily the cause, but the team now has a question it can investigate.
Which raises another question: how often do retrospectives begin with proposed solutions before anyone has agreed on the exact problem?
2. Follow the Five Whys beyond the obvious answer
The Five Whys, associated with Taiichi Ohno and Toyota, is attractive because it is simple. Ask why repeatedly until the underlying conditions become clearer. The trouble is that simplicity also makes the method easy to rush.
Why did the service fail? A configuration was incorrect.
Why? An engineer entered the wrong value.
Why? Lack of attention.
The investigation ends, and the corrective action practically writes itself: remind the engineer to be careful. But this explanation depends on an assumption about the person, and it leaves the process untouched.
A different investigation might discover that Friday releases require a manual configuration step, that automation has remained unfinished in the backlog, and that nobody has clear ownership of the deployment pipeline.
That sounds more promising, but it is still a hypothesis (a convincing story is not the same thing as evidence). Deployment logs, process comparisons, and the people who do the work need to confirm it. An Ishikawa fishbone diagram can organize competing explanations, and Pareto analysis can prioritize recurring problems.
3. Question the cause you think you found
Once the team has a candidate explanation, ask what would have happened if that condition had not existed. Would the release still have failed?
This removal test is a useful challenge to the reasoning, rather than a formal proof.
Many failures have several contributing conditions, and removing one may reduce the risk without making the problem impossible.
James Reason’s Swiss cheese model helps here, explaining that a serious failure can pass through several imperfect defenses when their weaknesses line up.
So identifying an incorrect configuration does not explain why the validation check missed it, why the process depended on manual work, or why that dependency remained unresolved.
There is an uncomfortable management question inside all this. When “human error” ends an investigation, has the organization found the cause, or simply found the person closest to the failure?
The distinction matters because someone can be responsible for a mistake without being responsible for every condition that made it possible.
4. Change the process, then check whether it worked
Finding a plausible cause is only part of the job. The organization must decide what to change, who will change it, and how anyone will know the change worked.
Safety engineering’s hierarchy of controls offers a useful principle: removing a risky condition or building protection into the process is generally stronger than asking people to remember another instruction. For the deployment example, that might mean eliminating the manual step, automating validation, and assigning ownership of the pipeline. Training still has a place, but it should not carry the entire burden.
And then comes the part that tends to disappear from the retrospective agenda: checking the result.
Was the action completed?
Did the automated check catch errors?
Did Friday releases actually become more reliable?
The next time a retrospective ends with “improve communication,” “be more careful,” or “follow the process,” resist the temptation to close the discussion too quickly.
Ask what, specifically, will change in the way the team works, who can make that change happen, and what evidence would show that the problem has actually been addressed.
And perhaps there is another question worth asking…
How many of the recurring problems in your organization have already been investigated, documented, discussed, and assigned to someone, sometimes more than once?
A completed action item is useful, of course, but it tells us surprisingly little about whether the underlying problem has disappeared (spreadsheets are quite good at showing progress, even when reality disagrees).
A lesson learned is not necessarily a lesson applied. The difference becomes visible only when the next release, the next decision, or the next difficult week puts the supposed improvement to the test.
Before your next retrospective, revisit the lessons learned from the previous one. Choose one recurring issue and examine what actually changed since it was recorded. If the same conditions are still present, there is probably more work to do.
Because six months from now, when the team faces another difficult release, the quality of that retrospective will be measured by what happens differently, not by how well everyone explained what went wrong.
Posted on: November 18, 2026 09:00 AM |
Permalink



