Summary
Zendesk's Intelligent Triage assigns an intent to incoming tickets based on subject line and first public comment. When the classification is wrong, agents can manually correct it. Those corrections have no effect on future classifications. The model does not learn from them.
For accounts where intent accuracy matters -- routing logic, prioritisation, Explore reporting -- this creates a ceiling on how good the system can get. The only path to improvement is ongoing manual curation of intent definitions, which doesn't scale.
The gap
Intelligent Triage is positioned as a system that classifies tickets automatically, reducing manual handling. In practice, the accuracy floor is set at initial configuration and stays there. Agent corrections are a rich, continuous source of ground-truth signal and currently none of it feeds back into the model.
We've used the quality recommendations feature, worked through intent definitions, and reduced overlap. After all of that, we're still seeing a meaningful volume of misclassified tickets. The tooling for improving intent definitions is reasonable. The problem is that corrections made at the ticket level -- the most direct signal available -- are simply discarded.
This was raised in the Intelligent Triage article comments in 2022. The response at the time was that the system didn't learn from agent feedback "at the moment." Three years on, confirmed with Zendesk support, the answer is the same.
What we're asking for
Agent corrections to the intent field should be used as a training signal to improve future classifications. This doesn't need to be real-time or fully automated. A supervised process where accumulated corrections periodically inform the model would be sufficient. The data already exists every time an agent overrides a classification.
Who this affects
Any account using Intelligent Triage at scale where intent accuracy is load-bearing -- routing automations, SLA triggers, or reporting built on intent as a dimension. The current model works reasonably well at launch and degrades in usefulness relative to the effort required to maintain it.