What to do when something surfaces late

What makes it different

Nobody is looking for it. An early failure arrives during a period when everybody expects things to go wrong and the transition is still an open file. A late one arrives into normal operations, where the first explanation reached for will not be a change made fourteen months ago.

The connection has to be made. On this record, two of the three late failures were initially diagnosed as something else. The resignations were a retention problem. The repeated analysis was somebody being inefficient. The connection to a removal was made afterwards, by reading the list.

It arrives with a cost already paid. An early failure is a gap. A late one is an invoice, an audit, three departures. The conversation starts from the damage rather than from the absence.

It restores at full weight. This is the expensive part, and it is documented elsewhere on this site: a removal reversed by an incident comes back as the thing that was removed, not as the thing that was needed. Negotiating the size of a precaution immediately after it has failed is a conversation nobody is willing to have.

The sequence that works

Four steps, and the order matters because the second is the one everybody skips.

1. Fix the immediate thing. Obviously. The audit needs figures, the contract needs renegotiating, the new starters need somebody to ask.

2. Find out whether it was a removal, before deciding what to put back. This is the step the list exists for. Read it. The question is not "what would prevent this?" but "did we stop doing something that used to prevent this, and what exactly was it?"

On this record the answer was yes three times out of three, and in two cases nobody would have found it without the list.

3. Put back what was needed, not what was removed. With the entry in front of you, the original procedure's dimensions are visible — frequency, scope, who did it — and so is the specific failure. The reconciliation failed on assembly across quarters; it went back as an annual assembly with a named owner, at forty per cent. The shadowing failed on new starters having nobody to ask; it went back as a named person for a month, at fifty per cent.

Both of those are better than the originals. Neither would have been arrived at without the entry to look at.

4. Check the neighbours. A late failure is evidence about a category, not just about itself. Of the five problems here, four were knowledge-carrying removals. The first one to surface should have prompted a look at the other eight entries of that kind — and the list makes that a five-minute exercise.

What not to do

Do not reverse the transition. Three late failures across thirty-eight removals, against fourteen clean savings and twenty-four resizings, is not evidence that the subtraction was wrong. It is evidence that three particular removals were.

Do not add a process to prevent recurrence. The natural response to a late failure is a new checking procedure, which is a new entry on a list nobody will read in two years. The failure was not caused by insufficient process. It was caused by nobody being positioned to notice, and the remedy for that is an appointment with a document.

Do not stop reading the list. The temptation after a late failure is to treat the matter as closed. On this record the sixty-four-week failure arrived after the forty-seven-week one, into an organisation that had already had its lesson.

The thing to write down

One line against the entry: what failed, when, how it was found, and what went back at what weight.

It is the only part of a late failure that is worth anything to anybody afterwards, and it takes a minute. Everything useful on this site came from thirty-eight such lines.

The diagnosis that takes two minutes

Step two above — find out whether it was a removal — sounds like an investigation and is not. With the list in front of you it is a keyword search and a glance.

What makes it hard is that nobody thinks to do it, because by the time a late failure arrives the transition is not among the candidate explanations. On this record the resignations were investigated as a retention problem for several weeks before anybody opened the spreadsheet.

The cheap fix is a line in whatever process handles incidents, for as long as the list is live:

Check the removal list before concluding.

It costs two minutes per incident and it will produce nothing in the great majority of cases. On this record it would have produced something three times out of three, and in two of those the connection was otherwise made weeks later or by accident.

Eighteen months is a reasonable life for that line. After that the list has stopped being the first place to look, and the procedures that quietly returned have made it unreliable anyway.