Short answer
Keyword routing sends a message to a queue or workflow when it contains certain words, while intent classification routes it by what the customer wants, however they phrase it. Keyword rules are simple and predictable; intent routing handles rewording and messages with several requests.- Keyword routing
- Rules that match words or patterns in the message
- Intent-based routing
- A model labels what the customer wants; rules act on the label
- Keyword rules break on
- Synonyms, negation, typos and other languages
- Intent routing breaks on
- Requests that fit no defined intent
- Often combined
- Patterns for fixed formats, intents for meaning
Both approaches answer the same question: where should this message go? Keyword routing answers it by looking for words. Intent classification answers it by reading the message for what the customer wants. The difference only shows on the messages that matter most, the ones customers word in their own way.
The difference at a glance
Keyword routing is a list of rules; intent-based routing is a label first, then rules that act on the label.
| Keyword routing | Intent-based routing | |
|---|---|---|
| How it decides | "If the message contains refund, send to billing" | "This customer wants a refund", then route |
| Rewording | Misses phrasings nobody wrote a rule for | Reads different wordings as one request |
| Negation | "I don't want a refund" still matches refund | Understands the customer does not want one |
| Several requests in one message | First matching rule wins | Can label more than one request |
| Upkeep | A new rule for every exception | Update the intents and their examples |
| Transparency | Easy to see why a rule fired | Needs labels visible to the team |
Where keyword routing goes wrong
Keyword rules fail on ordinary messages from a Shopify store's inbox.
| Message | Keyword rule result | What the customer wants |
|---|---|---|
| "I sent it back two weeks ago, where's my money?" | No match for refund | Refund status |
| "Can I still cancel? It hasn't shipped" | Cancellation queue | Cancellation, correctly matched |
| "Cancel the gift note, the order is fine" | Cancellation queue | A change to the order |
| "Où est ma commande ?" | No match | Order status, in French |
| "Love it, but do you have it in a medium?" | No match | Exchange or product question |
Keyword classification, labelling text by the words it contains, is the simplest form of text classification. It works until customers stop using your words.
How intent-based routing works
Intent classification, the step that labels a message with what the customer wants, comes first; routing rules then act on that label.
- The message is read as a whole, with the subject line and earlier messages in the thread.
- It is labelled with one or more of the team's intents, or none when nothing fits.
- Rules send each label where it belongs: a team, a person, an automated reply, or AI drafting for a person.
- Messages with no label go to a general queue, where someone can spot a new kind of request.
This is what an intent router does in a chatbot or helpdesk. In many tools the team's labels are called topics rather than intents; the idea is the same.
When keyword rules still make sense
Patterns beat models wherever the format is fixed and the meaning is not in question.
- Order numbers, tracking numbers and email addresses, matched with a regular expression.
- Specific senders, such as a marketplace or a payment provider.
- Automated mail: delivery notifications, out-of-office replies, no-reply addresses.
- Words that must always reach a person, whatever the context, such as a mention of legal action.
Most teams end up with both: patterns for structure, intents for meaning. The practice this page rejects is the endless keyword rule set, where every misroute adds a rule and nobody can say any longer why a message went where it did.
In Aide, sender rules handle that layer before any topic is read: automated mail is filed, contact-form messages are opened for the customer they carry, and every filed message says why, with a one-click correction.
Moving from keyword rules to intent routing
A switch goes best in small steps.
- Export the current rules and list what each one is trying to catch.
- Turn those into a short list of intents with a one-line description each, plus the request types the rules never caught.
- Run intent labels next to the old rules for a few weeks and compare where they disagree.
- Move routing to the labels one intent at a time, and keep the pattern rules that still earn their place.
With Aide, the list in step 2 does not have to be finished up front: when no topic fits a message, Aide suggests one right on the message, with a name, a group and a one-line description, and the real messages a team saves to a topic as examples sharpen it from there.
Where Aide fits
Aide classifies every customer message against the team's own topics and picks up to three, best first, or none. Topics are written back to Zendesk, Gorgias or Front as tags or a ticket field, where the helpdesk's routing rules can use them, and each topic's AI level decides whether Aide answers, drafts or leaves the conversation to the team. A procedure can combine a topic with a message or subject pattern, so meaning and fixed formats work together. See AI ticket tagging.