The work moves up, it does not disappear
When routine cases automate, what remains in the queue is systematically harder. The team's average case difficulty rises even as volume falls, which is why 'we automated 80% of the work' rarely means 'we need 80% fewer people'.
Plan for this. The people who used to process cases become the people who handle exceptions, tune thresholds and improve the definitions — a genuinely different and more valuable job.
Measure the queue, not the automation rate
Automation rate is a flattering metric that hides the important question: is the remaining work getting done well and on time?
- Queue depth and age, not just throughput
- Time to resolution at the 95th percentile
- Correction rate — how often review disagrees with automation
- Cost per completed case, including model spend
Give the exception path the same care
A review queue with poor context is where good automation goes to die. Reviewers need the original input, the model's output, the rule that fired and the ability to correct in one action — otherwise they open four other tabs and you have recreated the problem you started with.