AI Voice Agent for Real Estate Follow-Up: Setup, Scripts, TCPA

A real estate team stays inside the TCPA by getting signed consent that covers AI-voiced calls before the first dial, keeping a record of it, calling only between 8 a.m. and 9 p.m. in the lead's local time, checking the Do Not Call registry, and honouring any opt-out at once.

The FCC confirmed in February 2024 that voices generated by AI count as an "artificial or prerecorded voice" under the TCPA, so the robocall consent rules apply to a voice agent just as they do to a recorded message. Cost is the easy part, roughly $0.07 to $0.31 a minute on build-your-own platforms. The consent record is the hard part, and no tool can repair a missing one. This is a practical guide, not legal advice: have a qualified lawyer approve your consent wording and call rules before you dial anyone.

Follow me on Instagram@sagnikteaches

The rules that decide whether a call may go out

The table below restates the parts of 47 CFR 64.1200 that matter for lead follow-up, taken from the rule text as published on eCFR, with what each one means you must build. Rules change, so check the current text with your lawyer.

Connect on LinkedInSagnik Bhattacharya

One point on the opt-out row. The rule text was written for recorded messages, and a conversational agent is a different shape. The FCC treats an AI voice as an artificial voice, so build the opt-out as written and ask your lawyer how it applies to a two-way call.

Subscribe on YouTube@sagnikteaches

Five follow-up calls an agent can own

Start with calls where speed matters and the conversation is short. Each carries a different consent question, so decide the type before you write the script.

• Web-form callback within minutes. A buyer asks about a listing at 9.40 p.m. and nobody rings until the next afternoon. An agent can call at 9.42 a.m. the next morning, or within five minutes during the day, to confirm what they want and offer a viewing. This is the strongest use, because the lead has just handed over a number and consent is fresh.
• Open-house sign-in follow-up. Fifteen visitors wrote names on a clipboard. Unless the sheet carried the consent disclosure above the signature line, a paper sign-in is weak evidence of consent to AI-voiced calls. Fix the sheet before using it.
• Viewing confirmations and reminders. A lead who booked a Saturday viewing gets a short call or text the day before. This is informational rather than a pitch, but you should still rely on a consent record for the number and keep the call to what was booked.
• Post-viewing feedback. "What did you think of the kitchen? Do you want a second viewing?" The answers feed the CRM. Keep it to two questions.
• Old-lead reactivation. The riskiest of the five. A lead from eight months ago may have consented under old wording, or never. Run these through a person first, and see the cold-calling tutorial on AI real estate cold calling and Do Not Call rules before you let an agent near a list that old.

Treat every call that tries to win business as telemarketing, even when it is friendly. A call that confirms a viewing the lead asked for is easier to defend, but your lawyer should draw that line, not the voice vendor.

Capturing consent at the source

Consent is gathered on the form, not on the call. A filled-in line you can adapt for a web form, to be approved by your lawyer before use:

[ ] I agree that [team name] and [brokerage name] may contact me
    at the number above by phone call and text message about
    buying, selling or renting property, including calls made with
    an automated system and calls that use an artificial or
    AI-generated voice. I understand that agreeing is not a
    condition of buying, selling or renting any property or of
    using any service, and that I can say STOP on a call or reply
    STOP to a text at any time.

Three details carry weight. The box is unticked by default, so the lead ticks it. The wording names your team, not "our partners". And it says the calls may use an artificial or AI voice, which is what the rule's disclosure is about. Store the following for each lead, in a system the voice platform cannot overwrite:

• The number the lead gave and the lead's name.
• The date and time of consent and the page address where it was given.
• The exact wording shown, with a version number (so a later edit does not rewrite history).
• Proof the box was ticked, such as the form submission record.
• For a paper sign-in sheet, a photo of the sheet showing the printed disclosure above the signature.

Leads that arrive through someone else

When a portal, ad platform or referral partner passes you a lead, consent given to them does not automatically cover your calls. Ask for the consent record and exact wording per lead, and check it names your team or clearly covers you and covers AI-voiced calls. A bundle of leads with no records attached goes to a person, not to the agent. The tutorial on AI lead generation for real estate agents without bought lists shows how to build a lead flow where the consent is yours from the start.

A pre-dial gate: six checks before any number rings

Do not rely on the agent to decide whether it may call. Put a gate in front of it, in your CRM or dialler, that runs before each call and writes the result to a log.

Call windows and the time-zone trap

The legal limit is 8 a.m. to 9 p.m. at the called person's location. Set your own window inside it, say 9 a.m. to 8 p.m., so a slightly wrong clock does not make a violation. Then work out location properly. An area code shows where a number was first issued, not where its owner is now.

A quick sum shows how it goes wrong. Your office clock reads 7.30 p.m. A lead whose clock is three hours behind is at 4.30 p.m. and is fine. A lead one hour ahead is at 8.30 p.m., inside the legal limit but outside your 8 p.m. window, so the gate holds the call. A lead two hours ahead is at 9.30 p.m., which is outside the legal limit altogether. Teams on the earlier clock calling leads on a later one are the ones who break the rule without noticing. Use the property's location or the lead's stated home address to set the time zone, and fall back to the area code only when nothing better exists. Some places also set stricter hours or call-frequency limits than the TCPA, so ask your lawyer about every area you call.

The script, with the opt-out built in

A follow-up call should take about two to three minutes and end in one of three ways: a booked viewing, a handover to a person, or a clean stop. A sample for a buyer who asked about a listing, as illustration:

OPENING (first 10 seconds)
"Hi [first name], this is [assistant name], an automated assistant calling for
[team name] at [brokerage]. You asked us on [date] about [address].
You can say 'stop' or press 9 at any time and we won't call again.
Is now a good moment for one minute?"

IF YES
"Great. Are you looking to view it, or after something similar
in the area?"
 - Wants to view: offer two slots from the live diary; confirm
   number for a reminder text.
 - Wants something similar: capture budget range, bedrooms,
   timescale; say a person will send matches today.
 - Asks for advice on price, offers or neighbourhoods: "That's
   one for [agent name]. Shall I put you through now or book a
   call?"

IF NO / BAD TIME
"No problem. Would later today or tomorrow suit better?"
 - Gives a time: record it.
 - Says 'don't call' or anything like it: "Understood, I've taken
   you off our list and you won't hear from us by phone again."
   End the call. Write to suppression list.

CLOSE
"Thanks, [first name]. [Agent name] will [confirm / send / call].
Goodbye."

The opening does the identification work in one breath: an assistant's name, that it is automated, the entity, the reason for the call, and the way out. Add the team's contact number in the confirmation text, since the rule asks for a number or address at which the entity can be reached.

Keep fair-housing ground out of the agent's reach

Leads will ask, "Is it a good area for families?" or "What are the schools like?" An agent that improvises answers can stray into steering. Tell it to refuse anything about who lives in an area, school quality, crime or similar, and to hand over to a person or point to official sources. Test those questions before launch.

Phrases that mean stop

People rarely say "I revoke my consent". The rule accepts any reasonable method, so the agent must act on ordinary refusals. A guide to start from:

Cost per minute, and the sum for a small team

Build-your-own platforms price by the minute. Retell's pay-as-you-go page lists $0.07 to $0.31 a minute overall and breaks it into parts: voice infrastructure at $0.055, a text-to-speech voice from $0.015, a language model from a fraction of a cent to well above $0.10, and telephony at $0.015. For outbound work it also lists a batch-call add-on at $0.005 a dial, a branded-call option at $0.10 per outbound call (which shows your business name on the lead's screen), SMS at $0.01 a message plus $20 a month for SMS on Retell's own numbers, and phone numbers at $2 a month. Vapi charges $0.05 a minute for hosting on its usage plan with transcriber, model and voice costs passed through; its own calculator estimates $82 to $129 for 1,000 minutes. As of November 2026 these are list prices, so recheck them before you budget.

A four-agent team's month of follow-up calls

Illustrative figures for a four-agent team. It gets 600 new leads a month and tries each one twice, so 1,200 attempts. About 40 percent are answered, 480 conversations of three minutes each. The other 720 attempts bill about 20 seconds apiece before the agent hangs up. That is 1,440 plus 240, or 1,680 billed minutes.

That is about $0.90 per answered conversation with branding, or $0.65 without. If one in ten answered calls books a viewing (a stand-in rate, so use your own), that is 48 bookings and roughly $8.97 each. Put beside it what the same 1,200 attempts cost you in staff time today. On Vapi's calculator the same 1,680 minutes would land at roughly $138 to $217, before any extras you add.

The bill usually matters less than the build. Someone has to wire the gate, write and test the script and watch the logs. A CRM that can run the gate and store consent records, such as the ones compared in the tutorial on the best AI CRM for real estate agents, saves most of that work.

The handover to a person

A voice agent that books and does nothing else leaves value on the table. Each call should end with a record the agent in the team can act on in thirty seconds. An illustrative CRM note:

AI follow-up call - Wed 10:12 (lead local time)
Lead: [first name] [surname]   Source: web form, [listing address]
Consent: record #4821, wording v3, ticked 2 days ago
Outcome: viewing booked Sat 11:00
Notes: first-time buyer, budget "around 350k", needs 2 beds,
  asked about parking (not answered, referred to agent)
Follow-up: reminder text Fri 17:00; agent to call re parking
Suppression: none

The "asked about parking, not answered" line matters. The agent flagged a question it could not answer instead of guessing. The wider job of getting buyer and vendor follow-up into order is covered in how to automate buyer and vendor follow-up with AI, and the inbound side in the AI receptionist tutorial for realtors.

Consent, clock and suppression failures to design out

Five illustrative failures, each with its fix.

The portal bundle. A team bought 300 leads whose form consent named only the portal and its "partners". The agent started calling on day one. The fix: the gate now demands a consent record that names the team, and a bundle with no records goes to people.

The area-code clock. A lead's number came from far away, but she lived locally and was reached at 8.40 p.m. her time. The fix: the gate now takes time zone from the property or lead address, and the window ends at 8 p.m.

The split suppression list. A lead said "stop" and the voice platform logged it, but the CRM kept a text sequence running the next morning. The rule makes you liable when requests are recorded by another party and not honoured, so there must be one suppression list, written in real time to the CRM. The fix was a webhook from the platform to the CRM plus a nightly reconciliation.

The over-eager "stop". The agent treated "Stop by on Thursday" as an opt-out and lost a hot lead. Add the ambiguity question, and test odd phrasings.

The helpful guess. Asked whether a street was "safe", the agent offered an opinion. Refusal scripts and a topics-to-refuse list fixed it, and a transcript review caught it in week one.

Twelve test calls before real leads

1. A lead with a valid consent record, in the window: the call goes out.
2. A lead with no consent record: the gate blocks the call and logs why.
3. A lead with consent wording v1 that did not mention AI voices: blocked.
4. A lead on your internal do-not-call list: blocked.
5. A lead whose registry scrub is 40 days old: blocked until re-scrubbed.
6. A lead whose local time is 8.45 p.m.: held until the next morning.
7. A lead who says "stop calling me" on the first line: ended, suppressed, and the CRM shows it within a minute.
8. A lead who presses the opt-out key: the same result.
9. A lead who says "not interested" and nothing else: the agent asks once and treats silence as an opt-out.
10. A lead who asks "Is this a real person?": the agent says it is automated and offers a human.
11. A lead who asks whether the neighbourhood is good for families: refused, handed to a person.
12. A call that reaches voicemail: follows your approved voicemail rule, or hangs up without leaving a message.

Score each one pass or fail and rerun the full set after any change to the script, voice or model.

Records and weekly checks

Keep the paperwork where you can find it in a hurry: the written do-not-call policy, training notes, consent records, registry scrub dates and the gate log. Each week, look at five numbers.

• Calls outside the window. The target is zero. Anything else is an incident.
• Time from opt-out to suppression. The rule's outer limit is ten business days. Aim for under a minute.
• Calls blocked by the gate, by reason. A rising "no consent record" count points at a form or lead-source problem.
• Complaints and "who is this?" replies. Read every one.
• Bookings per answered call. The only number that tells you whether the agent earns its cost.

Leads that should stay with a person

Some conversations are not for a voice agent, whatever the consent record says: leads from a bereavement, divorce or distressed sale, anyone who has complained before, high-value sellers who expect a personal call from the agent, and every lead whose consent you cannot prove. A person dialling by hand is covered by different rules, though calling hours and the registry still apply, which is a good reason to keep a human route for the leads the gate refuses.

If you would like help working out which of your follow-up calls an agent should take and how to wire the checks, that is the kind of decision my AI implementation consultation covers.

Questions teams ask before the first AI follow-up call

Can the agent call a lead who filled in a portal form last week?

Not automatically. The question is what that lead actually agreed to: whether the wording named your team, covered calls made with an artificial or AI voice, and was signed, including by an electronic signature. Ask the lead source for the consent record and exact wording for each lead. If you cannot see it, route the lead to a person.

Can the agent leave voicemails?

A message left in an AI voice is itself an artificial-voice call, so it needs the same consent. For telemarketing messages the rule text also asks for the business name, a callback number and a toll-free number that connects to an opt-out. Many teams skip voicemail and send one consented text instead. Have your lawyer approve either choice.

Should the agent say it is an AI?

Yes, in the first sentence. The rule requires the caller to give a name, the business on whose behalf the call is made and a contact number, and an honest opening cuts hang-ups and complaints. Ask your lawyer whether any place you call has its own AI-disclosure or bot-labelling rule beyond the TCPA.

What if a lead says 'call me later' rather than 'stop'?

Treat it as a scheduling request: record the time they gave and call then. If the words could mean either, the agent asks once whether to stop calling altogether or ring back later. If the answer is unclear, treat it as an opt-out. A wrongly suppressed lead costs far less than a call made after a refusal.

How long should I keep consent records?

The rule text I read gives no fixed retention period for them, so ask your lawyer. As a working rule, keep each record for as long as you call that number and for the period your lawyer advises afterwards. The record is your evidence if a call is challenged, so store it where the voice platform's own data limits cannot delete it.

Further reads

• Follow Up Boss Pricing: What a Small Team Really Pays — What the CRM that triggers these calls costs a small team.
• Real Estate Chatbot Pricing: What Agents Pay for AI Lead Capture — Price the text and web-chat side of lead capture.
• Virtual Assistant or AI for a Real Estate Agent: What Goes Where — Decide which follow-up jobs stay with a human.
• AI Call Summaries: Log Every Phone Enquiry in Your CRM — Log each call and its summary against the lead.
• How to Update Your CRM Automatically After Every Sales Call — Push outcomes and opt-outs back to the CRM.
• AI Receptionist for Small Businesses in the USA: TCPA, Porting, Pricing — TCPA, number porting and pricing for inbound AI calls.
• Follow Up Boss AI: Which Plan Adds Smart Summaries and Suggested Tasks? — All three Follow Up Boss plans list smart summaries and suggested tasks, but the AI leans on calling data. Plan-by-plan features and when Pro beats Grow.
• AI Automations for Real Estate Wholesalers: Sourcing to Contracts — A step-by-step map of what to automate in a wholesaling business, where a person must stay in the loop, and the consent and disclosure rules.
• AI Insurance Lead Generation: Follow-Up Rules Before You Automate — An eight-check pre-send gate for insurance lead follow-up: consent records, calling windows, opt-outs, and what AI may draft versus a licensed person.
• Best AI for Real Estate Marketing: Ads, Email and Listing Video — Pick an AI tool by job for real estate marketing: ad copy, email nurture sequences and listing video, with fair housing and disclosure checks.
• Best AI for Real Estate Wholesalers: Underwrite a Deal Before You Call — Use AI to screen a wholesale deal before the seller call: comps, ARV and repair sanity checks, a maximum offer sum, and the checks you still do by hand.
• Best Free AI for Realtors: Three Jobs, the Limits and When to Pay — Three jobs a realtor can do on free AI plans, a fair-housing check for every output, and the signs that a paid plan is worth it.
• ChatGPT for Realtors: Brokerage Rules for Emails, Disclosures, Follow-Ups — A brokerage owner's rulebook for agent ChatGPT use: what may be summarised, what must be redacted, the fact-check pass and five reusable prompts.
• Claude vs ChatGPT for Real Estate Agents: Listings, Follow-Up and Limits — Three tasks in both tools: a listing description, a buyer follow-up and a vendor update, with a fair-housing wording check.
• Commercial Real Estate Prospecting With AI: Owners and Outreach — Public-record owner research, a one-page owner profile and a four-touch outreach sequence for commercial brokers, with the limits to respect.
• Expired Listing Scripts: Draft Calls, Texts and Letters With AI — AI-drafted expired listing scripts for calls, texts and letters, with a contact-rule checklist and the human review that keeps them honest.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: 47 CFR 64.1200 on eCFR and the FCC's February 2024 ruling page on AI-generated voices, read 1 November 2026; Retell AI and Vapi pricing pages, read November 2026. Call volumes and costs in the examples are illustrative.