How to Build a Real Estate AI Voice Agent

A demo can call a lead by Friday. A production system must know which lead it is allowed to call, what it can say, which CRM fields are authoritative, how to recover from interruptions, when to transfer to a human, how to book without double-booking, what to retain, and how to prove consent after a complaint.

• Start with one narrow job and a human escape hatch.
• Treat the CRM—not the model—as the system of record.
• Use official telephony, CRM, calendar, and listing interfaces.
• Price failed attempts, supervision, QA, and compliance—not only connected minutes.
• Do not scale outbound traffic until consent and caller reputation are demonstrably healthy.
• Assume every model, voice, carrier, and webhook will fail in a different way.

“AI voice agent for real estate” can mean a calendar bot or something expected to sound like an experienced inside sales agent. Those are different products with different data, risk, latency, evaluation, and cost profiles.

Level 1 is usually the right starting point. Level 3 is not just “a better prompt.” It is a permissioned real estate application with a voice interface.

A reliable system separates real-time conversation from authoritative business actions. The model may plan the next turn. Deterministic services decide whether a call is allowed, read or write the CRM, query inventory, calculate availability, and produce an audit record.

The harness helps engineers design, code, test, and operate the integrations. It is not the live telephony runtime. Claude Code and Codex are repository-first engineering tools; OpenClaw is an open-source agent harness that adds hosting and operational ownership; Manus is useful for delegated prototyping; Perplexity is strongest for current research and source discovery; Homies is the managed real-estate product path.

Monthly figures are the working assumptions in the stack calculator above, not vendor quotes; plan tiers and credit rates change.

The important boundary is simple: use the harness to build and maintain the product; use a low-latency voice runtime to conduct the call; use the CRM and approved property systems to hold the record. For a deeper treatment of what a harness is and why it matters more than the model, read the realtor’s guide to AI harnesses.

There is no universally cheapest provider because the bill can include a platform fee, language-model tokens, speech-to-text, text-to-speech, carrier minutes, concurrency, recordings, transcription, retention, and enterprise compliance.

Carrier cost is separate in many architectures. Twilio currently lists outbound calls to the U.S. and Canada from $0.014/min, inbound local calls from $0.0085/min, local numbers at $1.15/month, recording at $0.0025/min, transcription at $0.05/min, answering-machine detection at $0.0075/call, and branded calling at $0.12/call. See Twilio Canada Voice pricing . Telnyx publishes a lower Voice API layer beginning around $0.002/min before SIP trunk and destination fees; see Telnyx Voice API pricing .

This is where the SEO query and the engineering reality meet. Realtors search for “AI Follow Up Boss,” “AI kvCORE,” “AI Lofty CRM,” or “GoHighLevel voice AI,” but a useful integration is not a generic Zap. It needs identity mapping, event ordering, duplicate prevention, field ownership, signed webhooks, suppression state, and a replay strategy.

The minimum CRM write-back contract

• Store provider call ID, internal trace ID, lead ID, assigned agent, purpose, and consent basis.
• Write structured outcomes—not only a prose transcript: connected, qualified, appointment, transfer, opt-out, wrong number.
• Use idempotency keys so webhook retries cannot create duplicate notes, tasks, appointments, or messages.
• Preserve do-not-call and do-not-text fields as hard policy inputs, never prompt suggestions.
• Keep raw recordings and transcripts in a governed store with explicit retention; write a minimized CRM summary.
• Route low-confidence or consequential changes to a human review queue.

These five platforms are the head of a much longer tail. For the full map of realtor CRM, MLS, IDX, email, voice, and website systems—with official API links and honest access status for each—see the complete real estate AI integration stack.

A human ISA knows the team’s offer, lead sources, neighbourhood language, escalation paths, calendar norms, and CRM habits. A strong realtor knows far more: current inventory, property facts, market context, agency duties, disclosures, financing caveats, professional boundaries, and when not to answer.

A production knowledge base is an ongoing content operation, not a one-time upload. Someone must own source approval, chunking, metadata, field permissions, freshness, conflict resolution, deletion, and test coverage. IDX or VOW data should normally be queried through a scoped tool; it should not be copied into a generic vector database and forgotten.

Start outbound with the warmest, clearest-consent cohort: recent inbound inquiries who asked to be contacted. Do not begin with an old database export and a throughput target.

Raw voice minutes are only one line. A defensible budget includes product and engineering labour, runtime, carrier, number registration, CRM middleware, observability, call recording/transcription, evaluation, legal review, support, and the human who handles exceptions.

The implementation ranges below assume a focused DIY build using a modern coding harness and existing provider primitives. They do not price every integration as a custom agency project. Formal procurement, proprietary middleware, enterprise security, bespoke data agreements, and multi-market legal work belong in a separate budget.

Can ElevenLabs cost more than a human ISA?

On raw connected minutes, a well-designed AI stack is usually cheaper than a staffed hour. But that is not the correct denominator. If answer rates are low, calls are too long, qualification is wrong, appointments no-show, or a human must review every transcript, the cost per useful conversation can exceed a competent human ISA.

Canada’s Job Bank currently reports a median call-centre customer-service wage of roughly CAD $22/hour nationally. Benefits, payroll burden, management, training, occupancy, software, and performance incentives raise the loaded cost. See Government of Canada wage data . The calculator lets you change the loaded human rate and compares it on connected talk time at 45% occupancy.

Voice is expensive because it buys synchronous attention, richer qualification, and faster objection handling. SMS and email are dramatically cheaper and usually better for confirmations, low-intent nurture, simple updates, and asynchronous property links.

The best sequence is often multimodal: immediate consented SMS, a voice call when intent warrants it, email for durable detail, and a human for consequential advice. CASL, TCPA, carrier, consent, and unsubscribe requirements still apply to the cheaper channels. If text should carry most of your funnel, the companion field guide on building a real estate AI text messaging agent covers consent, deliverability, and webhook architecture at the same depth.

Where Slybroadcast voicemail drops fit

Slybroadcast publishes pay-as-you-go pricing from $12 for 100 deliveries and volume tiers that decline with scale. It can be useful as a consented follow-up or service notification. It is not a loophole around telemarketing law; its own terms put legal compliance on the customer. See Slybroadcast pricing and terms .

• Store consent provenance: source, language, purpose, channel, timestamp, evidence, expiry, and revocation.
• Disclose identity and the nature of the automated or AI-assisted call as counsel requires.
• Enforce jurisdiction-aware calling hours, internal DNC, national registry rules, frequency caps, and immediate opt-out.
• Obtain separate advice for call recording, transcription, biometrics/voice, data residency, and cross-border processing.
• Make the calling organization—not the model—responsible for every list, vendor, script, and outcome.
• Run fair-housing/human-rights tests so qualification does not use protected classes or proxies.

“Warm up” should mean earning legitimate reputation, not rotating numbers until spam systems lose track. A new number is not a clean-reputation hack, and aggressive number rotation can make the program look more abusive.

1. 01

   Verify the business

   Complete carrier/customer profiles, use accurate legal identity, and obtain the highest appropriate STIR/SHAKEN attestation.

2. 02

   Use stable numbers

   Assign recognizable local or toll-free numbers to real use cases; configure CNAM or branded calling where supported.

3. 03

   Register legitimate traffic

   Use carrier analytics and number-registration programs such as Hiya where appropriate; monitor labels across networks.

4. 04

   Begin with consented cohorts

   Call recent inbound leads in small batches. Keep cadence conservative and respect time zone, suppression, and frequency.

5. 05

   Watch quality signals

   Answer rate, immediate hang-up, very short call, blocked call, complaint, opt-out, wrong party, transfer success, appointment quality.

6. 06

   Pause before damage compounds

   When labels, complaints, or short calls rise, stop the campaign, inspect audience and script, and remediate the root cause.

Twilio documents STIR/SHAKEN , branded calling , and caller-reputation practices . Hiya explains that number registration can associate legitimate business identity with calling traffic, but does not guarantee removal of spam labels; see Hiya Connect Number Registration .

1. Days 1–5

   Define one job

   Choose one lead source, one consent basis, one script, one calendar, one CRM outcome set, and one human owner.

2. Days 6–10

   Build the event spine

   Implement identity mapping, authorization, signed webhook intake, idempotency, CRM read/write, call trace, and suppression.

3. Days 11–15

   Create the evaluation set

   Test accents, interruptions, silence, objections, voicemail, wrong numbers, hostile content, transfers, and every prohibited claim.

4. Days 16–20

   Run internal and invited calls

   Use staff and explicitly invited participants; measure latency, task completion, false statements, tool failure, and handoff.

5. Days 21–25

   Launch a consented micro-cohort

   Call a small recent-inbound audience during conservative hours with live monitoring and immediate pause controls.

6. Days 26–30

   Scale only the proven path

   Increase volume gradually only if opt-outs, complaints, wrong-party calls, spam labels, appointment quality, and human rework are healthy.

And remember the product-design test hiding in every customer-service phone tree: when a robotic voice agent traps you, what is the first thing you say? “Can I speak to a human?” Your transfer path is not an edge case. It is a core feature.

Build when you have a genuinely unique call workflow, enough volume to justify dedicated engineering, legal and operations owners, strong CRM and property-data access, and a reason to control every runtime layer.

Buy when your advantage is converting leads, serving clients, and closing real estate. The managed product should absorb the telephony, model routing, CRM plumbing, quality assurance, compliance controls, and continual provider changes—while you still verify that it fits your jurisdiction and brokerage.