Use an LLM to label each inbound reply with one intent from a fixed list (interested, question, reschedule, wrong number, opt-out, other), then route the label to a person or a rule. Let it sort, never send. A model that classifies is useful and low-risk. A model that texts people unsupervised is neither.

Reply volume grows faster than a team can read it. Most replies are simple: "yes," "not now," "who is this." Sorting them first puts the interested ones at the top and keeps the opt-outs out of the sales queue. For a broader view of model guardrails, read LLM texting agent guardrails.

The pattern

  1. Receive the reply on your webhook and store it. Answer fast.
  2. Pre-filter with rules. STOP words and clear opt-out phrases are handled by Ready already: STOP, UNSUBSCRIBE and phrases like "please stop texting me" or "take me off" opt the contact out. Don't send those to a model.
  3. Classify the rest with a constrained output: one label from your list and a confidence value.
  4. Route. High-confidence "interested" goes to the owner with an alert. "Wrong number" updates the contact. Low confidence goes to a human queue.
  5. Log the text, the label and the model version, so you can audit and improve.

Prompt shape

Give the model the reply, the last outbound message for context, and a closed list of labels with a one-line definition for each. Ask for JSON with label and confidence, and reject anything else.

LabelDefinitionRoute
interestedWants to continue or asks for next stepOwner, high priority
questionAsks about price, time, detailsOwner
rescheduleWants another timeCalendar flow
wrong_numberNot the personMark contact, stop texts
not_nowDeclines for nowPause sequence
otherAnything unclearHuman queue

Guardrails

  • Opt-out wins. Odd wording no list catches still needs a person to read replies. If the model thinks it's an opt-out, route to a human and mark the contact do-not-disturb meanwhile.
  • No auto-send. Draft suggestions are fine for staff to approve. Ready's own AI features help write and check messages, and they don't send replies by themselves.
  • Treat the reply as untrusted input. A reply can contain instructions aimed at your model. Never let its text change the label list, call tools or reach other systems.
  • Keep PHI out. Ready isn't HIPAA compliant and doesn't sign BAAs, so don't route health details through texts or model calls you can't cover.
  • Version and test. Keep 200 labeled replies as a regression set, and rerun it when you change the model or prompt.

Cost

A received reply is billed at $0.0075 per segment plus the $0.003 received carrier fee, whether or not a model reads it. Model calls are a separate bill from your model provider. Classifying a short reply is cheap, but price it at your volume.

Frequently asked questions

Can the model decide who to stop texting? Let it flag, and let rules or a person confirm. A wrongly kept opt-out is a compliance risk.

What about non-English replies? Most models handle them. Test your top languages in your regression set.

Can I use an MCP server? Ready runs one, which lets agents send texts and read conversations. See the SMS MCP server guide. Classification doesn't need it.

Next step

Label 200 recent replies by hand, run the model on them, and look at the disagreements before you route anything live.

Accounts opened before November 12, 2026 keep their signup prices, shown under Billing.