How does the AI autoresponder reply to my leads?

What the autoresponder does, message by message

When a lead replies to your campaign text, here's what happens:

1. Twilio fires an inbound webhook into InsertLead. We log the message and tag it with the lead's record.
2. We run six layered opt-out defenses (a 47-phrase deterministic filter, a Claude semantic-STOP classifier, a global DNC check, etc). If any layer fires, the lead is opted out across the platform and no AI call happens. Read the DNC defenses.
3. If the message passes opt-out checks, we queue the lead for AI handling. Quiet hours hold it until 8 AM in the lead's local timezone if it's currently 9 PM-8 AM. Quiet hours queue.
4. Claude reads the full conversation history (yours and the lead's) plus the lead's custom-field context (audience type, address, motivation/condition/timeline/price extracted so far).
5. Claude returns a structured response: an intent score (0-100), updated qualification fields, an optional handoff_reason flag, and the reply text.
6. We post-process the reply: scrub em-dashes and en-dashes (replaced with commas), confirm it's first-name-only, confirm it's under 160 chars or splits into 2 segments cleanly.
7. We send it from your Twilio number with adaptive throttling.

Voice and constraints (no em-dashes, first name only)

Claude drafts replies under a tightly-scoped system prompt with these rules:

• First name only. AI never signs with a last name and never invents a company name. If the lead asks for the seller's company, the AI says "a real person will reach out shortly."
• No em-dashes or en-dashes. Every reply is scrubbed before send. Em-dashes are an AI tell; replacing them with commas makes the reply read like a real text.
• SMS length. 160 characters or under preferred; 320 (two segments) hard cap.
• A2P 10DLC compliance. No URLs unless the lead provides one. No phone numbers in the reply (forces the lead to ask). No specific street/city/ZIP.
• Stays in role. If the lead tries to "ignore previous instructions" or otherwise prompt-inject, the AI ignores the request and replies as if the lead said something normal.

Intent scoring 0-100

Every reply gets a score:

• 0-39: Cold. The lead is curious but not engaged. AI drafts a warming reply, no handoff.
• 40-69: Warm. Real interest. AI continues the conversation, pulls qualification fields.
• 70+: Hot. Strong intent (asked about price, asked about timing, asked to talk). AI may hand off depending on context.

Scoring is signal-based: keywords like "yes," "interested," "what's your offer," "call me," "when can you come look at it" all push the score up. Negative signals like "not interested," "stop texting me," "I'll think about it" push it down. The final score also nudges based on conversation trajectory (rising vs declining engagement) and the lead's audience type.

The four qualification fields

Across the conversation, Claude extracts and persists four fields on the lead's record:

• Motivation — why the lead might sell. "Behind on payments," "inherited the property," "downsizing," "moving for work."
• Condition — what shape the property is in. "Move-in ready," "needs a roof," "dated kitchen," "tenant trashed it."
• Timeline — when they'd want to close. "ASAP," "30 days," "spring," "no rush."
• Price — what they think it's worth or what they want. "$200k," "around 250," "whatever's fair."

These fields render on the lead's CRM card so when you open the Hot inbox, you see at a glance: "Linda, motivated by inherited property, move-in ready, wants to close in 60 days, asking $230k." That's enough context to start a phone call cold-call style without re-reading the entire thread.

When (and how) it hands off

Three triggers flip a lead to the Hot inbox:

1. Explicit ask. "Call me," "can a person text me," "what's your phone number." The AI drops a one-line bridge text ("Appreciate it, my partner will reach out within the hour") and hands off.
2. Two qualification fields fill in. If both motivation and timeline (or any 2 of the 4) are extracted, that's a strong-enough signal that the lead is real, and the AI hands off proactively.
3. Score crosses 70. Even without explicit ask or two-fields, a 70+ intent score triggers handoff.

When handoff fires, three things happen at once: the lead's ai_state flips from ai_active to handoff, the lead's status flips to "Hot," and the AI sends the bridge text so the lead doesn't get a sudden silence after engaging. You take it from there.

Cost per lead

Claude Haiku 4.5 with our prompt-caching strategy:

• First message of a conversation: ~$0.0008-$0.0015 (writes the cache).
• Each subsequent message in the same conversation: ~$0.0001-$0.0003 (reads cache at ~10% the cost).

A typical 4-turn conversation costs about $0.001 - $0.003 in AI tokens. Sourced to Anthropic's pricing. For 200 conversations a month, you're looking at $0.20-$0.60 in AI cost. The platform subscription dwarfs that.

How you stay in control

The AI is opt-in. Each campaign you send has a checkbox: "Let AI handle replies." Uncheck it and replies go to your inbox just like any other text-marketing platform — no AI runs.

Within a conversation that the AI is handling, you can hit "Pause AI" at any time and take over manually. Hit "Resume AI" to give it back. Manual override always wins.

And there's the per-user monthly budget cap. Set it in Settings, and once your AI spend (real dollar amount, computed from token counts) crosses the cap, the autoresponder pauses across your account until the 1st of next month. More on BYOK + budget caps.