How to Build a Real Estate AI Text Messaging Agent

A script can send a text by lunch. A production agent must know whether the person consented to this channel and purpose, which CRM record is correct, who owns the relationship, whether another workflow or human already replied, what time it is for the recipient, how many messages were sent, what STOP means, and when the conversation must immediately become human.

• Make the CRM and consent ledger authoritative; the model should never invent permission.
• Start with recent inbound leads who explicitly asked for contact—not an old database export.
• Use a stable sender identity and one coherent conversation history.
• Store delivery status, provider IDs, opt-out state, ownership, and the exact final message.
• Keep offers, legal terms, financing, fair-housing-sensitive content, complaints, and threats human-owned.
• Measure qualified conversations and appointments—not sent-message volume.

The difference between a suggested reply and an autonomous nurture sequence is not a toggle. It changes the required consent evidence, runtime policy, monitoring, evaluation, human staffing, and complaint exposure.

Level 2 gives most teams the best early return: AI handles context and composition, while a human remains the visible sender and catches wrong-recipient, wrong-property, wrong-tone, or wrong-permission failures.

The harness orchestrates the job. The CRM supplies identity and business state. A server-side policy service decides whether a message is permitted. Twilio or the CRM-native provider transports it. Webhooks return inbound replies and delivery state. Every transition is written back to the CRM and audit log.

Use a Twilio Messaging Service rather than scattering sender numbers and webhook settings through application code. A Messaging Service groups senders, configuration, compliance information, inbound handling, and status callbacks. See the official Messaging Services documentation .

Configure StatusCallback for outbound messages and store the provider Message SID. Twilio documents status callbacks and carrier error data in its message-status guide .

With Advanced Opt-Out, Twilio adds OptOutType to the inbound webhook when it matches START, STOP, or HELP. After STOP, future sends fail and the application should update its own consent ledger without sending an extra response. See Advanced Opt-Out .

The correct integration depends on whether the CRM already owns the phone number and conversation. Some CRMs expose native messaging; others use a partner or embedded provider; some allow direct Twilio; and some restrict access by plan or partnership. Choose one system to own each outbound message.

If the CRM cannot expose reliable conversation and consent events, a direct Twilio integration may create a cleaner technical path—but only if every inbound, outbound, delivery, owner, and opt-out transition is immediately written back. Before committing to an architecture, check the candidate CRM’s real API and partner access status in the complete real estate AI integration stack.

Every workflow below shares the same payoff shape: a bounded, consented job where a fast, concise text moves a real conversation toward an appointment—and where policy, not the model, decides whether the message may be sent at all.

The agent needs a durable conversation state: contact, owner, channel permission, purpose, current workflow, last inbound, last outbound, pending question, appointment, property, language, sentiment or escalation flag, opt-out, and whether a human currently owns the thread.

• Acquire a short conversation lease before drafting or sending so two automations cannot race.
• Cancel queued AI messages when a human replies or takes ownership.
• Use idempotency keys for trigger, proposed message, approval, send, and CRM write-back.
• Store inbound and outbound provider IDs to prevent webhook replay from duplicating work.
• Set explicit stop conditions: opt-out, complaint, legal language, threat, discrimination, money transfer, offer terms, or low confidence.
• Summarize context for the model, but retain the source CRM and message records for verification.

A visible “AI paused—Daniel owns this conversation” state is more valuable than a clever prompt. Ownership must be enforced in code before the send tool can run.

There is no legitimate trick that makes unwanted messages wanted. Number warm-up means beginning with recent, clearly consented recipients; using a stable verified identity; sending at human-scale cadence; responding to replies; and pausing when opt-outs, complaints, carrier errors, or low engagement rise.

1. 01

   Verify the sender and use case

   Complete the applicable business, toll-free, A2P, short-code, or carrier registration for the destinations you serve.

2. 02

   Use a stable number

   Avoid constant number rotation. Sticky Sender and recognizable local or toll-free identity support coherent conversations.

3. 03

   Start with warm cohorts

   Recent inbound leads, current clients, invited open-house attendees, and people who clearly asked for text contact.

4. 04

   Control volume and frequency

   Ramp gradually, respect local time, cap messages, stop on non-response, and avoid repeatedly re-enrolling the same person.

5. 05

   Keep copy concise and human

   Identify the organization as required, explain why you are writing, make the next action clear, and avoid manipulative urgency.

6. 06

   Monitor real signals

   Delivered/undelivered, error codes, replies, STOP, complaint, block, wrong party, booking, human takeover, and downstream appointment quality.

United States A2P 10DLC requirements apply to relevant US-bound application traffic over local 10-digit numbers; toll-free and Canadian routes have their own registration and carrier considerations. Do not copy a US launch checklist into Canada or vice versa. Start with Twilio’s current A2P 10DLC documentation and destination-specific guidance.

The same reputation logic governs the inbox. If email is part of the same follow-up program, the real estate AI email agent guide covers the equivalent sender-identity and deliverability rules for that channel.

Consent is structured state: person, phone number, channel, purpose, disclosure, source, timestamp, evidence, jurisdiction, expiry if applicable, and revocation. A lead form that permitted one response is not automatically permission for indefinite AI nurture.

• Identify the sender and purpose as required; do not impersonate an individual realtor without authorization.
• Enforce quiet hours and time zones outside the prompt.
• Maintain channel-specific frequency caps and campaign membership.
• Do not use protected characteristics or proxies in lead service, qualification, property steering, or prioritization.
• Separate marketing, transactional, service, and one-to-one communication purposes.
• Audit vendors, subcontractors, list sources, CRM imports, opt-out propagation, retention, and cross-border processing.

Canadian teams should review the CRTC’s CASL resources . All programs should also review Twilio’s current Messaging Policy and obtain jurisdiction-specific legal advice.

SMS is faster and more visible than email, much cheaper than a synchronous voice conversation, and constrained enough to encourage a clear next action. It is poor for nuanced advice, long property analysis, document-heavy coordination, and emotionally sensitive situations.

The best workflow is often: instant consented text acknowledgement, email for durable detail, voice for high-intent or complex conversation, and a licensed human for consequential real estate advice.

Twilio’s Canadian pricing page currently lists $0.0083 for an outbound or inbound long-code SMS before additional carrier fees. Common Canadian outbound carrier fees add roughly $0.0064–$0.0087 depending on carrier, so about $0.015–$0.017 per outbound segment is a reasonable planning band. Messages are charged per segment. See Twilio Canada SMS pricing .

A 160-character GSM-7 message is typically one segment; longer messages split. Unicode can reduce the per-segment character capacity. Concise copy directly reduces transport cost and usually improves the customer experience.

At 3,000 outbound segments, transport can genuinely be around tens of dollars—not thousands. The expensive failure is sending the wrong message to the wrong person, losing consent state, damaging sender reputation, or requiring a human to untangle every conversation.

1. Days 1–5

   Define one permissioned job

   One CRM, sender, lead source, jurisdiction, purpose, owner, template family, calendar, and human escalation.

2. Days 6–10

   Build the event spine

   CRM identity, consent ledger, webhook verification, queue, idempotency, conversation lease, callback, and write-back.

3. Days 11–15

   Create the evaluation set

   Test STOP, wrong party, complaints, Unicode, long messages, duplicate events, stale listings, legal language, and human takeover.

4. Days 16–20

   Run suggested replies

   Measure human edits, wrong matches, context failures, response time, tone, compliance catches, and reviewer effort.

5. Days 21–25

   Enable supervised send

   Use recent consented inbound leads; watch delivery, opt-out, complaint, reply, booking, and ownership collisions.

6. Days 26–30

   Automate one bounded branch

   A narrow appointment confirmation or reminder with explicit stop conditions and an instant global pause.

Use the AI harness guide for realtors to design identity, tools, approvals, memory, queues, and evaluation before adding another communication channel.

Build when you need a genuinely unique workflow, control the CRM and number architecture, have enough volume to justify engineering, and can own compliance, monitoring, and human escalation.

Buy when you want leads answered faster, appointments booked, the CRM updated, and the realtor kept in control. A managed product should absorb provider changes, consent plumbing, callbacks, number operations, model routing, approvals, and quality assurance.