Maya writes a careful postmortem after an outage and churn continues anyway, and finds the accounts that left were counting occurrences rather than reading apologies
The Third Apology — Volume I, Understanding Humans Before Products · Trust · Failure & recovery · Practitioner. Maya writes a careful postmortem after an outage and churn continues anyway, and finds the accounts that left were counting occurrences rather than reading apologies
01 · Opening
Maya writes a careful postmortem after an outage and churn continues anyway, and finds the accounts that left were counting occurrences rather than reading apologies
02 · The setup
Four hours down on a Tuesday. Maya writes the postmortem herself and it is the best one the company has published.
Maya thinks: Timeline, root cause, what we are changing. No hedging anywhere in it.
03 · The setup
Priya: Three customers replied to thank us for the detail. That never happens.
04 · The evidence
Six weeks later the churn number has not moved back.
Churn, enterprise accounts — Baseline monthly: 1.1%; Month of outage: 2.9%; Month after: 2.7%; Postmortem read rate: 68%
05 · The evidence
Maya: Two thirds of them read it and they left anyway.
Sam: They read it. I am not sure they believed it.
06 · The evidence
So Maya stops reading the dashboard and calls four accounts that cancelled.
Maya thinks: I want to know which sentence lost them. I do not think there is one.
07 · The evidence
Priya: It was a good note. So was the one in March. So was the one in November.
08 · The evidence
Maya goes back through the archive and finds the shape of it in about ten minutes.
Incidents and follow-through — November — postmortem published: yes; November — actions completed: 1 of 4; March — postmortem published: yes; March — actions completed: 0 of 3; This outage — same root cause as March: yes
09 · The evidence
Dev: We wrote down the same fix in March and it never got scheduled.
Maya: So the note was not a promise. It was a description of one.
10 · The evidence
Count the occurrence, not the incident — The customer is not scoring this outage. They are scoring how many times this has happened and what changed after the last one. Find out which number they are on.
Name the mechanism, not the intention — 'We will improve monitoring' is an intention. 'This alert now pages a human at two minutes and here is the alert' is a mechanism. Only one of them is checkable next time.
Report the previous actions first — Open the note with what you said last time and whether it happened. If that paragraph is uncomfortable to write, the apology was never the missing piece.
11 · The evidence
Sam: So we write a better one next time.
Maya: A better one is worse. It is the same failure with more craft in it.
12 · You make the call
The story does not tell you first.
A. The four hours of downtime
B. That this was the third occurrence and nothing had visibly changed
C. The postmortem was too technical for a business audience
13 · What happened
The March actions get scheduled. The next incident note opens with a paragraph about them, and it is short because there is finally something to say.
Maya: Three of four done, one dropped and here is why. That is the whole opening.
14 · What happened
Maya thinks: I spent two days on the writing and none on the thing the writing was about.
15 · Complete
An apology is a claim about a change, and the customer is scoring the change. Next: why the same users look like nothing when you sort them by company size and like a pattern when you sort them by what they were trying to finish.