· · 5 min read

How Usetta actually reads intent: inside the classification engine

"Keyword matching" is why most chatbots feel dumb. Here's what Usetta does instead, and why it holds up on the kind of messy, real conversations that break simpler systems.

A person smiling while looking at organized message threads on a laptop screen

Most automated messaging tools work by scanning for keywords: if the message contains "price," trigger the pricing reply. It's cheap to build and it breaks constantly, because real people don't write in keywords. "Is this worth it for a team our size?" is a pricing question with no price-shaped words in it at all. "Not sure this is for us right now" is an objection that a naive keyword scanner might miss entirely, or worse, misfire on if the word "sure" happens to trip a positive-intent rule.

Usetta doesn't scan for keywords. Every inbound message is read by Claude and classified based on what the message actually means, in context, before any reply gets generated.

Six intent classes, one classification per message

Every message gets sorted into exactly one of six categories:

Example messageClassified as
"This looks interesting, tell me more"interested
"What's the cost for a team of 15?"question
"Seems a bit expensive for what it does"objecting
"Can we do a call Tuesday at 4?"booking
"Please stop messaging me"unsubscribe
"hey do you sell shoes"not_relevant

The classification is what determines everything that happens next: which reply gets generated, whether the lead's state moves forward, and whether a calendar link goes out. A message classified as booking triggers auto-scheduling. A message classified as objecting gets a reply shaped by your objection playbook, not a generic acknowledgment. Get the classification wrong and everything downstream is wrong too, which is exactly why this step doesn't rely on surface-level pattern matching.

Where keyword rules actually fail

Three patterns break simple keyword systems constantly:

Reading a message the way a person would, rather than scanning it for trigger words, is the difference between a reply that actually addresses what was asked and one that technically responds to the message while missing the point of it.

Classification feeds the lead-state funnel automatically

Intent classification isn't just about picking a reply. It's also what moves a lead through the funnel: cold to engaged to hot to booked to closed. A string of question and interested classifications on the same conversation signals real engagement and moves the lead forward. A booking classification is what actually triggers the calendar link to go out, at the exact moment the intent was expressed, not on a delay.

Underneath: Usetta uses Claude to read and classify messages, then generates the reply itself using your knowledge base and your do-not-say rules as hard constraints. Classification and reply generation are separate steps on purpose, so the reply is always shaped by both what the model understood and the rules you actually set.

Why this matters more than it sounds like it should

A wrong classification doesn't just produce an awkward reply. It can send a hot lead a generic response, or worse, fail to trigger a booking link for someone who was ready to schedule right now. Getting intent right on the first read is the difference between automation that feels genuinely useful and automation that quietly loses deals while looking like it's working.

See it read your own inbound

3-day free trial · card required · live in under 10 minutes

Start for free →