How AI Can Classify SMS Replies

Send 500 renewal texts and you'll get a wall of replies in two languages. Classification is how that wall becomes a worklist.

The reply problem

Outbound SMS works in insurance — clients actually read texts. The cost is the reply flood: "SI", "who is this?", "STOP", "my new address is…", "yes but call after 5", photos of documents. Unsorted, this is an afternoon of reading. Sorted, it's a short worklist.

What a classifier decides

• Intent: interested / question / not interested / opt-out / wrong number.
• Language: reply in the client's language, store the preference.
• Urgency and routing: a billing complaint outranks a thumbs-up; route accordingly.
• Confidence: the most important output — low-confidence messages go to a human review queue, never to an auto-reply.

The compliance floor (non-negotiable)

Opt-out handling must not depend on AI being right. Regex-level STOP detection should fire regardless of classification, mark the record do-not-text, and confirm the opt-out. A polite "no gracias" is different from STOP — a good system distinguishes a soft decline (mark disinterested, stop the campaign) from a legal opt-out (block the channel).

What lands in the CRM

Each classified reply becomes structured data on the household: a disposition, a task if interested, a queued draft if a reply is needed, and a clean audit trail of what the AI decided and who approved what. That's the difference between "we have AI" and AI you can defend to a carrier or regulator.