Automation that handles the exceptions.

High volume, stable, and painful enough that somebody can describe them precisely. If nobody can write down the rules, that is the finding — the process needs deciding before it can be automated. Frequency matters more than complexity, because a fiddly task done daily beats a simple one done twice a year.

They go to people, and designing that route is most of the work. Automation handles the ordinary path; the value is lost when a rare case silently takes the wrong branch. We measure what proportion needs human handling and make that visible, because a rising exception rate is the early warning that the process has changed.

Usually it removes the parts of a job nobody wanted. The realistic outcome is the same team handling more volume and spending their time on the exceptions and the customers. We would rather say that plainly than have the project discovered as a headcount exercise halfway through.

Automation follows rules you decide; AI makes a judgement you have to check. Most useful systems are both — deterministic steps for everything predictable, with a model at the points needing interpretation. Starting with rules is cheaper and easier to debug, so we do that first.

It breaks, which is why we integrate through APIs wherever one exists and treat screen-level automation as a last resort. Where it is unavoidable, the automation is monitored so a break is detected in minutes. Fragility comes from the integration method rather than from automation itself.

5 mo on the work we have delivered. The honest calculation counts the maintenance as well as the build, because an automation nobody owns degrades and quietly stops saving anything.