The support-triage models nobody talks about are the ones that actually keep working six months after launch.
Most AI pilots die quietly around month four, when the model's blind spots start costing more than the time it saves.
The models that survive share one trait: a human stays the final approver on anything customer-facing, and the model is scored on how often its draft needed no edits — not on a generic benchmark.
Applied ML isn't a model — it's a feedback loop. Build the loop first.