The exceptions are the work, not the edge case

Most automation handles the comfortable ninety-five per cent and leaves the week exactly as long as it was. The other five per cent is where the week goes.

Nukes AI

AI automation team

The cases an automation handles are the ones that were easy to describe, and the ones that were easy to describe were rarely the expensive ones. A team can automate the bulk of its volume and find that the week is exactly the same length, because the week was never made of the ordinary cases in the first place.

Volume is not effort

A process where ninety-five cases in a hundred are routine and five are not will spend most of its hours on the five. The routine ones take a minute each. The others take a conversation, a lookup in a second system, and a decision somebody has to be senior enough to make.

Automating the ninety-five removes the minutes. It leaves the hours exactly where they were, and it was the hours that people were complaining about.

Why the easy cases get built

Because they are the ones that can be specified. An exception is by definition the case nobody wrote down, so it is absent from the requirements, absent from the sample data and absent from the demo.

The first version is therefore fitted to the cases that were visible during the fitting, and the ones that were not visible come back with a note saying they were out of scope.

Look at the calendar

The work that costs a team is not the work it does most often. It is the work it cannot do without stopping.

The way to find the expensive cases is to ask where the day was actually interrupted, not what the volume report says. Those two answers are almost never the same one.

Exceptions have a shape

They are not random. Sorted, most operations find the same handful repeating:

data that arrived incomplete
two systems that disagree
a case outside the agreed policy
a customer who has asked twice

Each wants a different response, and only the first is a candidate for automation in the ordinary sense. The other three want a decision made by a person with the right information already in front of them.

Route, do not resolve

A system that tries to resolve every exception will resolve some of them wrongly and quietly. A system that recognises an exception, classifies it and puts it in front of the right person has removed the part that was actually slow: the finding, the sorting and the waiting.

Give the handler the context

Most of the time an exception takes is spent assembling the picture — opening three tabs, finding the earlier ticket, working out what the account is entitled to. That assembly is automatable even where the judgement is not, and it is usually the larger half.

Exceptions are a backlog

Logged with their reasons, they become a list sorted by frequency, and the top of that list is the specification for the next thing worth building. An exception that recurs a hundred times a month has stopped being an exception.

The measure that matters

Not the share of cases handled without a person. The change in how long the difficult ones take, and whether anybody still has to go looking for them.