On method: the failure rate

Every volume of Principles of Practice ends by naming the number its field repeats most often and explaining why the book does not use it. For automation it is the failure rate: seventy percent of digital transformations fail, or eighty five percent of AI projects, or some neighbouring figure.

The numbers circulate widely and they share a defect that is not really about sourcing. Even where a real survey exists, the word failure is never defined. Failure to complete, to deploy, to meet its business case, to be adopted, or to still be running two years later are five different measurements with wildly different rates, and the headline figure does not say which.

A statistic whose central term is undefined cannot be checked, cannot be compared across studies, and cannot inform a decision. It can only be quoted.

It is also worth noticing what the number is usually for. It appears most often in material selling the thing that will prevent the failure, which is the same profile as every other claim this series declines to use.

This volume argues its case from mechanism instead. Automation fails silently because output continues. The check requires retained skill. The world drifts and the rule does not. Each of those is something you can test on your own systems this quarter, which is worth more than a percentage about somebody else's.

The four questions from Volume I apply unchanged. Who produced it and what were they selling. What was the sample and of what population. What exactly was measured, in units. Has anyone independent reproduced it.