Businesses often jump straight to conversational AI when a list of keywords would resolve most of their inbound volume. Keyword routing is unglamorous, transparent, and debuggable at three in the morning — which is more than can be said for the alternative.
What it does
Match words in an incoming message, then act: send a reply, tag the contact, assign to a team, or start a flow. That is the whole idea. Its virtue is that when it misbehaves you can see exactly why.
Start from your actual messages
Do not invent categories. Export the last few hundred inbound messages and read them. You will usually find that five or six intents account for most of the volume, and that customers describe them in a small vocabulary you can enumerate in an afternoon.
Typical outcome for a retail business: order status, delivery timing, returns, product availability, pricing, and "is anyone there".
Match generously
- Case-insensitive, always.
- Include misspellings you actually see — "recieve", "adress", "cancle".
- Include the languages your customers write in, including transliterated Hindi.
- Match on substrings so "tracking" catches "tracking?" and "my tracking no".
- Include the short forms people really type: "wru", "eta", "cod".
Order your rules deliberately. "cancel" appears in both "cancel my order" and "how do I cancel a return". First-match-wins with a badly ordered list sends people to the wrong place. Put the most specific patterns first and the general ones last.
What to do on a match
Resist answering everything automatically. The strongest pattern is usually to route rather than to reply: tag the conversation, assign it to the right team, and let a human answer with context already attached. The customer gets a person; the person gets a head start.
Reserve automatic answers for questions with one correct, stable answer — opening hours, address, return window length. Anything account-specific should go to a human or a flow that can look it up.
Handle no match well
Most messages will not match, and that is fine. The fallback should be a straightforward handoff to a person, not an apology loop. Never make a customer guess the magic word.
Keep a review loop
- Weekly, sample unmatched messages.
- Add patterns for anything that recurs.
- Check the mis-routes — messages that matched the wrong rule — and reorder.
- Delete rules that never fire. Dead rules make the list harder to reason about.
A keyword list maintained this way for a couple of months will handle a large share of routing accurately, and you will know precisely what it does. If you later want a model-based classifier, you will also have a labelled dataset and a baseline to beat — which is a far better position than starting from nothing.