Route conversations by intent with the AI Classifier in Omnifox

Hand-sorting every incoming chat is slow and error-prone: someone has to read it, understand it, and drag the conversation to the right team. Routing conversations by intent with AI removes that bottleneck. The AI Classifier node in Omnifox uses a language model to read the customer's last message, decide the category, and open one branch per intent. In seconds the chat lands with the right people.

What the AI Classifier is

The AI Classifier is a special condition node. Instead of rigid rules ("if the text contains the word invoice"), you describe the categories that matter to you and the LLM decides which one the message fits. Each category becomes an output of the node, plus you get a fallback branch for when the AI isn't confident. This is far more robust than keyword matching because it understands synonyms, typos, and indirect phrasing.

The full flow, node by node

A typical intent router looks like this:

1. Trigger Message received. Turn on the first only filter so classification runs when the customer opens the conversation, not on every line. Optionally filter by channel if you only want to route WhatsApp.
2. AI Classifier condition. Wire in the trigger's output. Define three categories, for example: sales, support, and billing. Give each a short description ("sales: wants to buy, pricing, plans"; "support: something's broken, technical help"; "billing: payments, invoices, charges").
3. Sales branch -> Transfer to an agent: send the chat to a human rep or a sales team.
4. Support branch -> Assign AI agent: let a text AI agent handle first-level questions before escalating.
5. Billing branch -> Transfer to an agent on the collections team, and maybe an Add tag "billing" for reporting.
6. Fallback branch -> Send message asking for clarification ("To help you faster, is this a purchase, support, or billing question?") or straight to a general Transfer to an agent.

The short version: Message received (first only) -> AI Classifier -> (sales) Transfer to an agent / (support) Assign AI agent / (billing) Transfer to an agent / (fallback) Send message.

How to write good categories

Routing quality depends on how you describe the categories:

• Few and well-separated: 3 to 5 categories work better than 10 that overlap. If sales and billing get confused, sharpen the descriptions.
• Examples in the description: add 2 or 3 sample phrases per category. The AI uses them as reference.
• Always define the fallback: never leave the customer without an exit when the model can't classify confidently. The fallback branch is your safety net.

Combine with other conditions

The AI Classifier plays well with other nodes. A powerful pattern is placing it after a Business Hours condition: within hours the support branch goes to a human; after hours the same branch goes to Assign AI agent so AI covers the overnight window. You can also put Detect language in front to classify and respond in the customer's language.

Common mistakes

• Overlapping categories: if "support" and "billing" share words ("my payment failed"), the AI hesitates. Clarify the boundaries in the descriptions.
• Forgetting the fallback: without that branch, ambiguous messages have no route and the customer is left hanging.
• Classifying every message: without first only, the classifier runs on each line and can reassign the chat mid-conversation. Classify at the start and let the flow continue.
• Expecting it to pull data that isn't there: the node classifies intent, it doesn't extract order numbers. For that use Ask a question or a more specific downstream AI Classifier.

Advanced variant

You can nest classifiers: a first AI Classifier splits sales/support/billing, and inside support a second AI Classifier distinguishes "app bug" vs "how-to question," routing each to a different team with Transfer to an agent. That builds a routing tree as fine-grained as you need without writing a single text rule.

Tag and measure to improve routing

Here's a detail almost nobody does at first but that pays off over time: on each branch, alongside the main action, chain an Add tag with the category name ("intent-sales", "intent-support", "intent-billing"). After a week you can filter your conversations by tag and see the real volume of each intent. If you discover "billing" is 40% of traffic, it may deserve its own team or a dedicated AI agent. And if a category barely shows up, its description might be poorly written and the AI is sending it to the fallback. Tags turn the AI Classifier into a source of business data, not just a router.

When NOT to use the AI Classifier

If your routing depends on an exact, structured value, an AI Classifier is overkill. Examples: routing by the inbound channel (use the trigger's channel filter or an If condition), by whether the contact has a specific tag (Contact has tag), or by time of day (Business Hours). The AI Classifier shines when the signal lives in the customer's natural language, which is exactly where rigid rules fail. Reserve AI for that and use deterministic conditions for the structured stuff; the flow ends up faster, cheaper, and more predictable.

Routing by intent is one of the biggest time-savers from day one. Try it in Omnifox and let AI put every conversation in the right hands.