A manual onboarding process handles about 400 cases a month. Automation covers the standard path, and after six months roughly 20% of cases still go to a person because each has some exception. The team proposes automating the remaining exceptions. How should this be sequenced?

The sequence

1. Classify the 80 exceptional cases a month by cause, for one month, before writing any code. The distribution is almost always uneven: two or three causes account for most of the tail, and a long list of singletons accounts for the rest. Without this data, the team automates the exception it remembers most vividly rather than the one that occurs most.
2. Automate the two or three largest causes and measure the fall in manual volume. Expect this to take the manual share from 20% to perhaps 8%, at a fraction of the cost of the whole tail.
3. Make the remaining exceptions first-class rather than hidden. A case that cannot be automated should be routed, tracked and timed as a legitimate path, with its own SLA, not treated as a failure of the happy path.
4. Stop, and re-measure quarterly. The tail changes: a partner's data improves, a regulation arrives, a product change introduces a new case. Automating a permanently moving target at 100% is a commitment nobody should make.

Where data can diverge, and how you would know

Automating an exception means encoding a judgement somebody was making by hand, and the two will disagree. The safe mechanism is to run the automation in shadow for a period: it produces a decision, the human still decides, and the differences are reviewed. A disagreement rate above a few percent means the rule is wrong, not the humans.

The point of no return

When the manual path is decommissioned and the people who knew the exceptions have moved on. That is worth delaying deliberately. Keep the manual route usable, and keep at least one person who can exercise it, for at least a full business cycle after the automation is live.

What it really costs

The standard path was 80% of cases for maybe 20% of the effort. The tail inverts that ratio: each remaining cause is rarer, more complex, and has fewer examples to test against. A reasonable planning assumption is that automating from 80% to 95% costs about as much as getting to 80% did, and the last 5% costs more than both.

When not to automate the rest

When the cost of automating a cause exceeds the cost of handling it manually for two years. At 400 cases a month with a 20% tail and, say, 20 minutes of work each, the manual effort is about 27 hours a month. That is a real number, and it is smaller than most of the automation proposals it attracts. The right answer is frequently a better tool for the human, which halves the 20 minutes for a fraction of the cost of removing the person entirely.