Use AI to classify topics and intent
Add semantic decisions when exact text matching is too brittle.
Before you begin
Open Workflows in the workspace navigation. Decide which event should start the flow and which agents should receive it. Start with an unpublished draft.
Step by step
- 1
Define the category
Describe the intended subject and exclusions. “Installation support” is clearer when you explain whether product sizing or order tracking belongs elsewhere.
- 2
Provide representative context
Use the message/conversation fields the condition supports. Avoid sending unnecessary personal data to a decision provider.
- 3
Choose the supported decision settings
Use the offered AI/provider options. Jev-backed decisions require the workspace’s relevant privacy approval and server-side configuration; availability is not implied by selecting a label.
- 4
Test ambiguity
Try paraphrases, mixed topics, short messages and an unrelated topic. Inspect the trace and fallback behavior rather than trusting one positive example.
Good to know
AI classification is probabilistic. Keep important eligibility and identity checks deterministic. Available fallback LLM decisions can consume model usage.
A topic match should route help, not silently authorise refunds, account access or policy exceptions.
Check the result
Test matching, non-matching and missing-data cases. Review the publication, publish the steps and save the intended agent assignments.
Questions & troubleshooting
A category is selected too often
Narrow its definition, add exclusions and test nearby categories. Use deterministic conditions around the AI decision when possible.
Rate this guide on this device. Ratings are not sent to Corthex.