Workflow node: AI Classifier (route by intent)

What it does

The AI Classifier (route by intent) node asks a language model to read the customer's last message and place it into ONE of the categories you defined, then continues down the matching branch. It's not a chat — it's a single classify-and-route call meant for triage.

When to use it

Great for splitting inbound messages by intent before deciding what to do next: for example, separating Sales, Support, and Renewals at the start of a WhatsApp conversation, or detecting a complaint before escalating.

Configuration

• Categories (at least 2, required): each with a label and, optionally, a hint ("pricing questions, demos…") to help the model decide.
• Reference AI agent (optional): reuses an existing agent's configured model instead of picking one manually.
• Model (optional): fallback model when no reference agent is set.
• Minimum confidence (0 to 1, default 0.6): below this threshold, the message goes to the unmatched branch.
• Unmatched label: the name of the branch used when no category clearly fits (a plain greeting, an empty message, insults, or ambiguity).
• Save to variable (optional): exposes the result under whatever name you choose, in addition to the standard $ai_classify.category variable.
• Include history: when enabled, the model sees the last 5 inbound messages instead of just the latest one.

Example

In a WhatsApp reception workflow: the contact's first message hits an AI Classifier node with the categories Sales, Support, and Renewals. If it detects "Sales" with high confidence, the workflow follows that branch to an Assign AI agent node pointing at the Sales agent; if nothing matches, it falls to the "Unmatched" branch and a generic welcome message.

Tips

The node consumes the organization's AI credit, just like other Omnifox AI features — every classification counts against the plan's cap.

Troubleshooting

If everything keeps falling into "unmatched", check that your category labels roughly match the language you used in the hints, and that the minimum confidence isn't set too high for your use case.