Top 10 AI Agents forReal Estate Brokers in 2026

Every ranking of AI agents for real estate makes the same promise: buy this and you will convert more of your leads. The promise cannot be checked, because there is no independent benchmark of any product in this market. We reviewed fourteen products on 23 August 2026 and found no academic study, no third-party test, no audited conversion data and no trade-body evaluation of any commercial AI tool sold to brokerages. Not one. Every accuracy figure, every response rate, every ROI multiple published in this sector is published by the seller about its own software.

That is worth stating flatly because it is not the situation everywhere. In accounting there are model-level benchmarks. In legal there is a peer-reviewed hallucination study. In restaurants there is a published drive-thru accuracy comparison. Residential real estate has nothing at any level — not on the products, not on the underlying models as applied to this work, not even a partisan study with a disclosed method. When a category has zero independent measurement and universal vendor measurement, the honest response is not to average the vendor numbers. It is to stop scoring performance altogether and score the things that leave a trace.

The demand side of this is real, which is why the marketing works. The National Association of REALTORS® 2025 Profile of Home Buyers and Sellers, published 4 November 2025, reports that eighty-eight percent of home buyers purchased their homes through a real estate agent or broker, and that ninety-one percent of sellers sold with the assistance of a real estate agent, with only five percent FSBO — an all-time low. The same report puts first-time buyers at a historic low of twenty-one percent of all buyers, with a median first-time buyer age of forty. Read the methodology with the number: NAR mailed 173,250 surveys in July 2025 and received 6,103 responses, an adjusted response rate of 3.5 percent, with a confidence interval of ±1.25 percent. It surveys people who completed a purchase, so it measures buyers who bought, not the whole market.

So agents still intermediate almost every transaction, the buyer pool is shrinking and ageing, and the cost of chasing each lead has gone up. That is a genuine operational problem and software can genuinely help with parts of it. What follows is an attempt to describe which parts, using only facts a brokerage owner can verify without trusting us either.

You have been told that a lead is “391 times more likely to convert if you respond within five minutes.” That sentence is wrong in four separate ways, and we can show you the original document. If you have been sold this figure for a decade, this section is the one that will tell you whether the rest of this page is worth your time.

The study everyone is gesturing at is real. It is the Lead Response Management Study, presented by Dave Elkington of InsideSales.com and James Oldroyd, then a professor at MIT's Sloan School of Management, at MarketingSherpa's Business-to-Business Demand Generation Summit on 16 October 2007. We read it. Here is what it says about its own method, verbatim:

We examined 3 years of data across six companies that generate and response to web leads, from over fifteen thousand leads and over one hundred thousand call attempts.

— InsideSales.com / MIT Lead Response Management Study, 16 October 2007

Six companies. Roughly fifteen thousand leads. An observational analysis of one vendor's own CRM data, not a controlled experiment, and not real estate. That is a legitimate piece of work and a narrow one. Now the finding itself, also verbatim:

The odds of contacting a lead if called in 5 minutes versus 30 minutes drop 100 times. The odds of qualifying a lead if called in 5 minutes versus 30 minutes drop 21 times.

— InsideSales.com / MIT Lead Response Management Study, 16 October 2007

Twenty-one, not 391. It describes a drop in the odds of qualifyingwhen you wait from five minutes to thirty — a decay curve, not a lift you can bank. And the study states, in its own words, the sentence that is fatal to every vendor who cites it as a conversion statistic:

This study did not address close ratios.

— InsideSales.com / MIT Lead Response Management Study, 16 October 2007

It gets softer still on inspection. The outcome variable is not standardised across the six companies: footnote 2 states that each company involved in the study had their own way to indicate a qualified lead, and that the analysis took this into account. Even “contact” is not standardised — footnote 1 states that the time defining a contact ranged from two to six minutes depending on the company. So the 21x is a comparison across six differently-defined outcome variables, honestly disclosed by the authors and almost never disclosed by the people quoting them.

Where 391 actually comes from

The number 391 does not appear in that study at all. It traces to Velocify's sales material of roughly 2013, and it is 391 percent — about 4.9 times, not 391 times — and it refers to responding within one minute, not five. It is also routinely attributed to Harvard Business Review, which did not produce it.

So the sentence you have been sold performs four substitutions at once. It converts a percentage into a multiple, 391% into 391x. It moves the time window, one minute into five. It swaps the metric, qualification into conversion. And it swaps the source, a lead-management vendor into MIT or Harvard. Four errors, one sentence, a decade of circulation.

For completeness: there is a real Harvard Business Reviewarticle in this lineage — “The Short Life of Online Sales Leads” by James B. Oldroyd, Kristina McElheran and David Elkington, March 2011. We could not access its body text, so we do not quote its figures and neither should anyone else who has not read it. The 2007 study we did read, and it is the one whose text is above.

What survives the correction

Speed still matters. A 2007 observational study across six companies suggests that the odds of reaching and qualifying an inbound web lead fall sharply between five and thirty minutes, which is consistent with everything anyone who has run a call floor already knows. That is a defensible reason to automate first response. What is not defensible is presenting a decayed qualification odds ratio as a 391-fold lift in closed transactions, and then pricing software against it.

The vendor numbers we will not print as fact

The same discipline applies to everything the vendors publish about themselves. Each of the following appears on a vendor's own site, none carries a disclosed methodology, and none is independently checkable. We name them so you recognise them in a pitch, attributed to the seller and nothing more.

• Structurely: 3-21x average ROI: A seven-fold range described as an average, with no denominator, no time period and no method. This is the clearest single example in the category of a number that is not a statistic.
• Zuma: Rent Collected 100%: A one-hundred-percent claim with no denominator attached. A percentage without a base is a slogan.
• Ylopo: a 48% response rate across 25 million text conversations, a 1% opt-out rate, up to 10% of new leads instantly connected by live transfer, and a case study turning $20,000 into $112,310 GCI in eight months: All vendor-published, none audited. Ylopo also states that it takes seven calls on average to connect with a lead, presented with no source at all.
• Roof AI: a 7.5% lead-to-close rate and 4x more qualified leads: Vendor case-study figures with no method and no comparison group.
• BoldTrail: engage 5-10X more leads and contacts: A range of ranges, unattributed, on a product whose documented mechanics are trigger-based messaging.
• MoxiWorks: 3.5x more sales and 10M+ hours saved annually: Also vendor data — but MoxiWorks footnotes them as its own platform data with individual results varying, which is more disclosure than anyone else offers. Credit where it is due; it is still self-reported.
• Follow Up Boss: 36 of the 50 highest-volume teams in the US use Follow Up Boss: No attribution given for the ranking or the count.
• Zillow: 80% of US residential real estate transactions involve agents who use at least one Zillow product: Zillow's own estimate, sourced by Zillow to Comscore and to its own consumer report. Attribute it; never repeat it as neutral market fact.

We scored eight attributes, every one of which you can re-check yourself in a browser, and refused to score anything that only the vendor can measure. The ranking order below is not a quality judgement about the software. It is an ordering by how much of the product a buyer can verify before signing— which, in a market with no independent measurement, is the only defensible axis left.

Two rules governed every cell. First, every claim needed a citation or the literal words “not publicly disclosed.” Second, we never estimated a price.Where pricing is gated the cell says so, because a guessed price is worse than no price — it gives a reader false precision to plan against.

We also drew a line between two different absences that people conflate. “No public trust page found” is not the same as “not certified.” A vendor may hold SOC 2 and simply not publish anything; several here probably do. What we can report is what a buyer can see, and for an enterprise platform serving tens of thousands of agents, the absence of any published trust page is itself a finding you should raise in procurement.

Everything was checked on 23 August 2026. Vendor pages change, so re-check before you sign; the point of scoring only checkable attributes is that you can. If you want the broader cross-industry framing for how these products are categorised, our 2026 guide to the top AI agents across categories covers the general landscape; this article is the real-estate-specific application of the same method.

Ten products, six verifiable columns, one refused column.Every cell is either something you can open a page and confirm, or the words “not publicly disclosed.” The final column is the same for all ten, and that uniformity is the most important finding in the table.

All attributes checked 23 August 2026. Ownership entries marked “SEC-confirmed” come from filings linked in the sources list below.

Read down the last column first. Ten products, ten identical answers. Then read the ownership column, because it separates the market into two halves: the products whose corporate parent is a matter of public record, and the products whose ownership you simply cannot establish. Real Geeks appears at 100% under the FNF RE Technology Holdings branch of an affiliate ownership schedule filed on Form 485BPOS on 30 April 2026 — and the same schedule lists HBN Media, Inc. d/b/a Commissions Inc (CINC) at 99.1495% and SkySlope, Inc. at 96.022%. Fidelity National Financial quietly owns two competing real-estate CRMs and a transaction-management platform. Follow Up Boss is Enchant, LLC d/b/a Follow Up Boss, acquired by Zillow Group in December 2023 and named as such in Zillow's FY2025 Form 10-K. Sierra Interactive, Ylopo, Inside Real Estate, Rechat, Roof AI, Zuma and HouseWhisper disclose nothing, and none of them files publicly.

That matters more in this vertical than in most, because the parent decides what happens to your database. A CRM owned by a portal has a structural interest in the lead flow that runs through it; a CRM owned by a title and escrow group has a structural interest in the transaction that follows it. Neither is disqualifying. Both are worth knowing before you migrate twelve thousand contacts.

Each entry below states what the product does, what is verifiable, what is not disclosed, who it fits, who it does not, and who owns it. Where a vendor makes a performance claim we attribute it to the vendor and move on. Where a product is not really an agent, we say so.

Five more products are worth knowing about, and none of them met the bar for a ranked row. A ranking is only as good as its exclusions, so here they are with the reason.

• EliseAI: Elise A.I. Technologies Corp., multifamily leasing and resident communication, with LeasingAI, ResidentAI, VoiceAI and EliseCRM. Its voice product confirms both inbound and outbound calling, putting it in the same autonomy and consent class as Ylopo and Structurely — in a sector that also sits squarely inside HUD's tenant-screening and advertising guidance. It publishes a Vanta-hosted trust centre whose contents we could not read by automated fetch, so we assert neither the presence nor the absence of any certification. Its marketing site names no specific property-management-system integrations. A $250M Series E was announced 20 August 2025 and is reported rather than verified here.
• Zuma (Kelsey): Leasing, rent collections, move-outs and an AI receptionist, with a human backstop the vendor states plainly: Kelsey does it, and a team of human experts backs it up. That human layer is a genuine differentiator and worth reporting accurately. No named integrations, no trust page, no published pricing — and headline claims including Rent Collected 100% that carry no denominator.
• HouseWhisper: Positions around lead nurturing and routing, re-evaluating every lead daily on activity and engagement and reviving cold database leads. We could not determine from the public site whether messages send without human approval, and we will not classify a product's autonomy by inference. Founding details attributed to an aggregator profile rather than the company.
• Roof AI: Roof AI Inc., a conversational lead-capture bot for brokerage websites and social channels — autonomous within a chat session, scoped to inbound. Claims MLS and CRM integration with no documentation URL and no named partners; no dedicated security page; pricing gated. Named brokerage testimonials exist; the performance figures attached to them do not carry a method.
• CINC and Entrata ELI+: CINC is HBN Media, Inc. d/b/a Commissions Inc, 99.1495% FNF-owned per the same 2026 SEC affiliate schedule that confirms Real Geeks — a second FNF-owned CRM competing with the first. ELI+ is what Colleen AI became after Entrata acquired it on 20 June 2024; it should never appear in a 2026 ranking under the old name.

Two candidate names circulate in listicles that we could not confirm exist as trading products in 2026 at all, so they are not here and should not be in anyone else's table either. Several adjacent platforms — multifamily leasing systems, transaction-management tools and portal media products — were not researched to the standard this ranking requires, and are named in the limitations section rather than given a row.

An AI agent that dials or texts a real estate lead needs prior express written consent, obtained from the consumer, held by you, and provable. That is the whole constraint in one sentence, and it is the sentence most vendor demos never say. Everything below is why.

An AI voice is an artificial voice

The Federal Communications Commission settled this on 8 February 2024. In Declaratory Ruling 24-17, CG Docket No. 23-362, adopted unanimously on 2 February 2024 and effective on release, the Commission confirmed that the TCPA's restrictions on the use of an artificial or prerecorded voice encompass current AI technologies that generate human voices, so calls using such technologies fall under the TCPA and require the prior express consent of the called party absent an emergency purpose or exemption.

The ruling then closes the defence every vendor reaches for. Verbatim from paragraph 6:

And it also makes clear that the TCPA does not allow for any carve out of technologies that purport to provide the equivalent of a live agent, thus preventing unscrupulous businesses from attempting to exploit any perceived ambiguity in our TCPA rules.

— FCC Declaratory Ruling 24-17, CG Docket No. 23-362, released 8 February 2024, ¶6

Paragraph 9 extends this to any AI technology that initiates any outbound telephone call using an artificial or prerecorded voice to consumers. And footnote 13 supplies the escalation that matters commercially: if these robocalls introduce an advertisement or contain telemarketing, the Commission's rules require the caller to obtain the prior express written consent of the called party, citing 47 C.F.R. § 64.1200(a)(2) and (3). An AI voice agent calling a real estate lead to solicit business is telemarketing. It needs written consent.

What the Eleventh Circuit actually did — and did not do

The holding, from the body of the opinion rather than any front-matter summary, is that the FCC exceeded its statutory authority because the 2023 Order's new consent restrictions impermissibly conflict with the ordinary statutory meaning of prior express consent. The court explained that as long as a consumer clearly and unmistakably states, before receiving the robocall, that he is willing to receive the robocall, he has given prior express consent under the TCPA — and that callers must obtain prior express consent, not prior express consent plus.

Read the status words precisely, because four different states are in play and people collapse them constantly. Vacated applies to Part III.D of the 2023 Order and nothing else. The rest of that order was not before the court. The 2012 Order — the rule that makes prior express consent mean prior express written consent for telemarketing and advertising robocalls and robotexts — was expressly not at issue and survives untouched. And FCC 24-17, which classifies AI voices as artificial voices, is a separate 2024 ruling that the decision did not address at all.

The practical translation for a brokerage: the decision made bundled, lead-generator-style consent lawful again. It did not make consent optional and it did not lower the standard. Combined with FCC 24-17, the 2026 position for an AI agent that calls or texts a lead to solicit business is prior express written consent, held by the caller, and provable on demand.

Revocation: one piece waived, the rest in force

Do not let anyone tell you the revocation rule is delayed. FCC Order DA 26-12, CG Docket No. 02-278, adopted and released 6 January 2026, extended the effective date of 47 C.F.R. § 64.1200(a)(10) to 31 January 2027. That subsection is the revoke-allrequirement — the duty to treat a revocation made in response to one type of message as applying to all future robocalls and robotexts from that caller, including on unrelated subjects. The order is explicit about its own narrowness, stating that the waiver extends only to section 64.1200(a)(10) to the extent discussed therein and does not alter the status quo relating to any other prior Commission rules or rulings addressing revocation of consent.

So the rest applies now: a consumer may revoke consent by any reasonable means, and you must honour it. If your AI texts on one campaign and calls on another, and a consumer says stop to the text, you are not free to treat the call channel as unaffected simply because the cross-channel provision is waived — the underlying revocation obligations, and the state rules below, still bite.

Three state statutes that reach further than the federal floor

Federal consent is the floor, not the ceiling, and the state statutes are where the private-plaintiff litigation risk concentrates. All three below reach text messages and automated dialing, which is exactly what an AI lead agent does.

Florida.Fla. Stat. § 501.059(8)(a) prohibits an unsolicited telephonic sales call involving an automated system for the selection and dialing of telephone numbers, or the playing of a recorded message on connection, without the prior express written consent of the called party. A telephonic sales call expressly includes a text message. Consent must be a written agreement bearing the called party's signature — electronic signatures and affirmative acts such as checking a consent box count — clearly authorising automated calls, texts or voicemails, containing the specific number authorised, and disclosing that automated systems are used and that agreement is not a condition of purchase. The statute carries a private right of action with actual damages or $500, whichever is greater, trebled to as much as $1,500 for a wilful or knowing violation. It was narrowed in 2023 by chapter 2023-150 after a wave of class actions.

Maryland. The Stop the Spam Calls Act of 2023, Chapter 413 (S.B. 90), approved 3 May 2023 and effective 1 January 2024, codified at Md. Code, Commercial Law §§ 14-4501 to 14-4503, prohibits automated-dialing or recorded-message solicitation without prior express written consent — and its consent definition expressly covers a recorded or artificial voice message, language added by amendment. Maryland reached AI voice before the FCC did. It also caps solicitations at three to the same called party in any 24-hour period on the same subject matter regardless of the telephone numbers used, bars calls between 8 p.m. and 8 a.m. in the called party's time zone, and forbids intentionally altering the caller's voice to disguise or conceal identity for improper purposes. And it creates a rebuttable presumption that a solicitation made to any Maryland area code is made to a Maryland resident, which means dialing a Maryland area code presumptively triggers Maryland law wherever the lead actually lives. A violation is an unfair, abusive or deceptive trade practice under the Maryland Consumer Protection Act, which carries a private right of action.

Oklahoma.The Telephone Solicitation Act of 2022, Okla. Stat. tit. 15, § 775C.1, added by Laws 2022, c. 290, took effect 1 November 2022 and is widely described as modelled on Florida's statute, including a private right of action. We verified the act's existence and citation from the statute itself; the damages figures commonly quoted for it come from compliance-vendor guides rather than from the operative section, so treat them as reported and pull the statute before you plan to them.

For scale on how much consumer irritation sits behind these statutes: the FTC's National Do Not Call Registry Data Book for fiscal year 2025 records more than 2.6 million Do Not Call complaints against a registry of over 258 million active registrations. Print the FTC's own caveat with it, because they do: Do Not Call complaint data are based on unverified complaints filed by consumers, and the data are not based on a consumer survey.

The best-sourced sentence available on this subject comes from a state regulator, and it should be printed on the wall of every brokerage running AI: using AI tools to conduct licensed activity may be equivalent to asking an unlicensed assistant to do licensed activity. That is the California Department of Real Estate, in an advisory titled Artificial Intelligence in California Real Estate — Opportunities, Risks, and Compliance Considerations for Licensees — dated 17 March 2026.

The advisory does not stop at the classification. It extends supervision to the software: a broker's supervisory obligation extends to the tools used to conduct licensed or unlicensed activities, including AI-powered software. And it lands liability where operators need to hear it land:

If an AI tool generates inaccurate information, misleading advertising, or improper communications with consumers, responsibility under current law rests with the licensee and their responsible broker, not the technology provider.

— California Department of Real Estate, AI in California Real Estate advisory, 17 March 2026

Read that alongside the vendor compliance findings above. Ten products, and not one publishes a customer-facing compliance programme for the activity the regulator says you are responsible for. The advisory also flags that where training data reflects historical patterns of discrimination or bias, the resulting output may create fair housing concerns, and it cites a long list of California authorities including Business and Professions Code sections 10177(h), 10159.2, 10131, 10176, 10140.6 and 10140.8, DRE Regulations 2725 and 2780, Government Code sections 12955 to 12956.1 and the CCPA. Section 10140.8 governs digitally altered images in advertising, which is directly relevant to AI-enhanced listing photography.

One honest caveat: California is the only state regulator whose position on this we verified.We are not going to write that regulators in several states have warned about AI, because we found exactly one advisory and it is a good one. Other states' positions are undocumented, which is not the same as permissive — the underlying unlicensed-activity and supervision rules exist in every state.

Fair housing: what the agent writes, and who it decides to pursue

HUD's Office of Fair Housing and Equal Opportunity published two guidance documents on 29 April 2024, announced by press release on 2 May 2024. Both matter to a brokerage running AI, and each covers a different half of the exposure.

The screening guidance opens with the sentence that removes the technology defence: the Fair Housing Act applies to housing decisions regardless of what technology is used, and both housing providers and tenant screening companies have a responsibility to avoid using these technologies in a discriminatory manner. It states that it covers screening practices with varying levels of human involvement and automation, including machine learning and other forms of AI; that screening based on imprecise or overbroad criteria may unjustifiably exclude people from housing opportunities in discriminatory ways; and — the line that describes the large-language-model problem exactly — that these technologies can also lead to a less transparent process by obscuring the precise reasons for a denial from the housing provider and applicant.

The advertising guidance covers the other half. It lists the ways ad targeting can discriminate, including denying consumers information about housing opportunities, discouraging or deterring potential consumers, advertising different prices or conditions, and steering home-seekers to particular neighborhoods. It then closes the good-faith defence: importantly, this can happen without the advertiser's direction or knowledge, and can even frustrate an advertiser's intention that an ad be distributed more broadly. Its definition of advertiser expressly reaches residential real estate and related services.

For the multifamily and property-management side of a brokerage, two more status facts belong in the file, stated precisely. Louis v. SafeRent Solutions, LLC, No. 1:22-cv-10800 (D. Mass.), settled, with approval reported in November 2024; reported terms include up to $2,275,000 total and injunctive relief under which SafeRent agreed to stop issuing approve or decline recommendations based on its scoring model for applicants using publicly funded housing vouchers unless the model is validated for fairness, with the court retaining jurisdiction for five years. Nothing was adjudicated. There is no holding that the algorithm violated the Fair Housing Act. Write it as settled on these terms, never as a court found. We did not pull the settlement agreement or approval order, so the dollar figure is reported rather than verified.

On algorithmic pricing, the status words matter just as much. In United States et al. v. RealPage, Inc., No. 1:24-cv-00710-WLO-JGM (M.D.N.C.), the complaint was filed 23 August 2024. RealPage's own proposed final judgment was filed on 24 November 2025 and remained proposed as of the Justice Department's 8 May 2026 response to public comments, in which the United States said it would move the court to enter it. Greystar's final judgment was entered on 2 March 2026.Proposed is not entered, and the two should never be reported as the same thing. The obligations the RealPage proposed judgment would impose are instructive regardless of entry: restricting use of competitors' nonpublic data at runtime, limiting model training to data at least twelve months old, excluding active leases, prohibiting market surveys that gather competitively sensitive information, requiring individual customisation of auto-accept and guardrail features, and a court-appointed monitor for three years.

There is no national MLS rule about AI content

Never write, and never accept from a vendor, that “the MLS allows” or “the MLS forbids” AI-generated content. There is no national answer. RESO, the Real Estate Standards Organization, publishes the Web API — a RESTful and OData standard that replaced the deprecated RETS standard — and certifies MLS compliance with Data Dictionary standards, currently ratified at version 2.0 with 2.1 in draft. RESO states that it does not provide MLS real estate data and that data requests should be made to MLSs. We found no RESO standard, rule or guidance on AI-generated listing content, AI fields, or AI use of media and image data. RESO governs the pipe, not the copy.

Content and display rules are therefore set MLS by MLS, layered over NAR MLS policy. A brokerage operating in five MLSs can face five different answers, and the correct operational posture is a per-MLS matrix maintained by a named person, not a single company policy asserted nationally.

What does exist at national level is ethical and template-based. NAR publishes AI Policy Templates dated March 2026 — one for brokers and one for associations — which is the most current citable NAR artefact on this subject. On AI-enhanced listing photos, NAR ties the question back to Article 2 of the Code of Ethics, which prohibits REALTORS® from exaggerating, concealing or misrepresenting pertinent facts about a property or transaction. NAR also flags as an open issue how listing data, photos and other content may be scraped and used to train AI models. We found no NAR national rule requiring disclosure of AI-generated listing content, and we will not assert one.

Two pieces of context that shape the workflows these tools plug into: the NAR settlement practice changes took effect 17 August 2024, with MLSs given until 16 September 2024 for technical implementation, and final approval was granted on 26 November 2024. The two changes are that offers of compensation may no longer be published on an MLS, and that MLS participants working with buyers must enter into written buyer agreements before touring a home. That second one is now an automation surface — and an agreement a consumer signs is exactly the kind of document a workflow should be handling carefully, not creatively.

Consider a twelve-agent residential brokerage in a single MLS market, running one IDX site, buying leads from two portal sources, and currently paying a part-time inside sales associate to work first response. The owner wants to know what an AI layer costs and whether it pays back. Here is the half of that question that can be answered honestly.

The cost side, from published prices only

Sierra Interactive, Growth tier, annual billing. $599.95 per month covers five users. Users six through ten are billed at $20 each, which is $100; users eleven and twelve at $10 each, which is $20. Add Lead Engage at $199 per month for unlimited leads. One MLS feed is included and the dialer is included at this tier, with no setup fee on annual billing. Total: $918.95 per month, or $11,027.40 per year, of which the AI component is $2,388.

Follow Up Boss, Pro tier, annual billing. $416 per month includes ten users; the remaining two are $17 each, which is $34. AI features are included at no separate charge on all plans. Total: $450 per month, or $5,400 per year, with the AI component at zero incremental cost. Note this is a CRM only — no IDX website is included, so compare it against your existing site rather than against a full platform.

Real Geeks, monthly. $399 platform fee includes two users; users three through ten are $25 each, which is $200; users eleven and twelve are $10 each, which is $20. Geek AI Text is $49 per month, and one IDX board feed is $10. Add the $500 one-time setup. Total: $678 per month, or $8,136 per year, plus $500 in the first year, of which the AI component is $588 — and remember that AI email and AI voice are listed by the vendor as coming soon.

The other seven products cannot be entered into this table at all, because none of them publishes a price. That is not an oversight in our research; it is the finding. A brokerage owner comparing options on public information can construct exactly three columns out of ten, and the first honest number for the rest arrives inside a sales process.

The return side, and why we stop

This is where every competing article produces a payback calculation, and this is where we do not. To compute payback you need an incremental conversion rate attributable to the AI layer. No independent benchmark of any product in this market exists, and every conversion figure available is published by the vendor about itself. If we multiplied one of those figures by twelve agents and a local average commission we would produce a number that looks authoritative and rests on nothing — the precise failure this article spent two sections refusing.

What you can do instead is measure it yourself, which is cheaper than it sounds. Take the last ninety days of leads from one source, split them by an arbitrary rule that has nothing to do with quality — odd and even record IDs work — and run the AI layer on one half only for a full cycle. Measure appointments set and appointments held, not messages sent. Sixty to ninety days on a single lead source will tell you more than any vendor deck, and the resulting number is yours, is auditable, and applies to your market rather than someone else's.

The costs nobody quotes

• Consent remediation: If your existing database cannot produce a written consent artefact per lead, the AI layer cannot legally message most of it. Re-papering consent at the point of capture is cheap; retroactively establishing it for 12,000 legacy contacts usually is not, and the honest answer for much of an old database is that it is not addressable by automated outbound at all.
• Per-state configuration: Time-of-day windows in the called party's time zone, frequency caps that count across numbers rather than per number, and area-code presumptions mean a national dialer is a per-state configuration problem before it is a technology problem.
• Human review capacity: Every draft-for-approval feature moves work rather than removing it. Budget the reviewer's time explicitly, and instrument how long review actually takes — that number is the real efficiency measure, and it is one you own.
• Exit cost: Ask, before signing, how you export contacts, conversation history and consent records, in what format, and how long it takes. A CRM you cannot leave is a pricing position, not a product feature.
• Portal coupling: Where the AI's best signal comes from a portal, part of what you are buying is a deeper relationship with that portal. Price that strategically, not just financially.

In this vertical the first failure is almost never the model saying something silly. It is a consent record you cannot produce. Below are the eight failure modes we would instrument on day one, each with the signal that detects it and the rollback that contains it.

Two of these deserve elaboration. The first is prompt injection, and the honest framing is that it is unsolved. Any agent that reads untrusted input — an inbound lead message, a forwarded email, an uploaded document, a portal enquiry — is exposed, and no vendor detection claim changes that. What you can do is reduce the blast radius: give the agent a separate identity from your staff, scope its credentials to the specific operations the workflow needs, keep the credential that can send messages out of any session that reads untrusted content, and make the kill switch revoke the credential rather than pause the code. Treat any vendor who tells you injection is handled as having told you something useful about the vendor.

The second is autonomy creep. Agents drift toward being helpful, and in a licensed profession helpful shades into advice. Read a weekly sample of what the agent actually said — not the templates, the transcripts — and if it is answering questions about price, terms, condition or negotiation, narrow its permitted intents. Qualification and scheduling are safe ground. Everything past that belongs to a licensee, and the California DRE has already said whose problem it is when it goes wrong.

Three categories: what an agent may do alone, what it may prepare but not commit, and what it must never touch. This table is the artefact to hand your compliance lead and your platform vendor at the same time. If a vendor cannot tell you which row a feature sits in, that is your answer.

The general shape holds across every product on this list: an agent may read, classify, route, summarise, draft and propose; a licensed human commits anything that constitutes advice, any statement that publishes, and any outbound contact whose consent record cannot be produced on demand. That is not caution for its own sake. It is the only architecture that survives both a prompt injection and a regulator asking who authorised a message.

Buying one of the ten platforms above is the right answer for most brokerages, most of the time. A custom build earns its keep in three situations: when your workflow crosses systems no single vendor covers, when you need consent and audit controls no vendor publishes, or when the data you would be handing a vendor is itself the asset. If you are weighing that decision, our build-versus-buy analysis for AI agents walks the decision framework without a thumb on the scale.

Senior-led delivery runs $150–$225 per hour; ongoing retainers run $2,500–$9,500 per month. Every build carries a 30-day post-launch warranty, and we return a fixed-price phased proposal within 5 business days of a discovery call. Full source-code and IP ownership transfers to you at completion — which, in a market where seven of ten vendors will not publish a price and none will publish an exit path, is worth more than it sounds. Frenchy Digital is a senior-led Black-owned agency in Los Angeles; the same team that scopes the work delivers it, and if you want the local context for this kind of build, our AI agent development practice in Los Angeles sets out how the engagements run.

These are our bands, not an industry benchmark. They reflect senior-led delivery on compliance-heavy work in 2026 and should be read as one firm's pricing — exactly how you should read every other number a vendor gives you. If you want to walk your own stack through this, book a discovery call at calendly.com/frenchydigital/discovery-call or call +1 (424) 272-5601, and bring your lead sources, your consent capture forms and your current vendor contracts.

Every item below is something we encountered on a live vendor page while researching this article. None is disqualifying on its own. Two or three together should change what you sign.

• An ROI figure presented as a range and called an average: A seven-fold range described as an average ROI, with no denominator, no period and no method, is a marketing claim wearing a lab coat. Ask for the sample size, the time period and who calculated it, in writing.
• A percentage with no base: Rent collected 100%. Out of what, over what period, for how many properties? A percentage without a denominator is not a statistic.
• A SOC 2 badge that links to an explainer about SOC 2: One vendor's compliance link resolves to a page describing what SOC 2 is, with no type, no auditor, no report date and no request process. A real certification has a Type, an auditor, a date and a way to request the report under NDA.
• A certification list with no type stated: ISO 27001 and SOC 2 named without Type I or Type II, without an auditor and without a date tells you the marketing team wrote the page. Ask for the report.
• No trust page at all on an enterprise platform: For a platform serving tens of thousands of agents, no published security page, and a privacy policy path returning 404, is a procurement question rather than a footnote. It does not mean uncertified; it means unverifiable.
• A dead product name in the pitch deck: If a rep says kvCORE, Chime, Colleen AI or Raiya in 2026, the deck is old. Those names are BoldTrail, Lofty, Entrata ELI+ and Ylopo AI Text, AI Voice and AI Squared respectively.
• Compliance-friendly as a feature: Pre-built compliance-friendly scripts is a phrase, not a programme. Ask instead: where is the consent artefact stored, what fields does it capture, how do I retrieve it for one lead, and how do you enforce time-of-day and frequency rules per state?
• A TCPA resource that is coming soon: One vendor selling autonomous outbound calling lists its customer-facing TCPA best-practices sheet as coming soon. If a vendor sells the highest-exposure product in the category and has not published guidance for it, price that gap into your own compliance work.
• An agent that is a trigger: Property viewed sends a text; property favourited sends an email. That is drip automation. It may be exactly what you need — just do not buy it believing you bought reasoning.
• Ownership that cannot be established: Several vendors here disclose no owner and file nothing publicly. Ask directly who the ultimate parent is and get it in the contract's notices clause, because the parent decides what happens to your database at renewal.
• A performance claim you are told is from Harvard or MIT: Ask for the PDF. In this category, the most-quoted statistic in the industry turns out to be a vendor's percentage misattributed to a business school, and the actual study says it did not address close ratios.

The honest list. Everything below is a gap in the public record or in our own research, and each one is a place to verify before you commit money.

• No independent benchmark exists for any product in this category: Not one. No academic study, no third-party benchmark, no audited conversion data, no trade-body test. This is why the ranking scores integrations, compliance disclosure, pricing transparency, ownership and lock-in, and refuses to score performance. If someone shows you a benchmark, read who paid for it.
• Recent trade press was not swept: This research was built from direct fetches of primary documents — court opinions, FCC orders, enrolled state statutes, SEC filings, HUD guidance and NAR's own research PDF — rather than from search summaries. That makes the legal and ownership findings unusually strong and leaves one blind spot: roughly the last quarter of trade coverage. A mid-2026 funding round or acquisition at a privately held vendor could have been missed. The vendors most exposed to that gap are the ones that file nothing publicly.
• Lofty's 2026 ownership is unknown: Its parent deregistered from the SEC in November 2025 and the current boilerplate names no parent. We report the last filed position with its date rather than guessing at the current one.
• Seven vendors disclose no ownership at all: Sierra Interactive, Ylopo, Inside Real Estate, Rechat, Roof AI, Zuma and HouseWhisper file nothing publicly. Where trade sources report an owner, we hedge it and date it.
• EliseAI's certifications are neither confirmed nor denied: It publishes a Vanta-hosted trust centre whose contents are client-rendered and could not be read by automated fetch. We report that a trust centre exists and that specific certifications are unverified — not that it is or is not certified.
• No public trust page is not the same as not certified: BoldTrail, Sierra Interactive, Real Geeks, MoxiWorks, Roof AI, Rechat, Zuma and HouseWhisper publish nothing we could find. Several may hold certifications they do not advertise. Ask, in procurement, under NDA.
• No TCPA case against a real-estate AI vendor or brokerage was verified: We found none and will not imply a docket we did not read. The risk in the consent section is stated structurally from the statutes and the FCC ruling.
• RealPage's own final judgment status after May 2026 is not confirmed: It was filed 24 November 2025 and remained proposed as of the Justice Department's 8 May 2026 response. Whether it has since been entered is not established here, so we state it with the date anchor. Greystar's was entered 2 March 2026.
• The SafeRent settlement terms are reported, not pulled: The approval date and dollar figure come from clearinghouse and press summaries rather than from the settlement agreement or approval order. And it is a settlement: nothing was adjudicated.
• NAR's Clear Cooperation position in 2025–26 was not verified: We could not confirm the content or effective date of the current policy, so we do not describe it. The Code of Ethics changes reported as effective 1 January 2026 rest on trade sources only and are not relied on here.
• Two quoted regulator sentences were truncated on fetch and are paraphrased: The California DRE advisory's sentences on advertising accuracy and on human review were cut mid-sentence by our fetch, so they appear here as paraphrase rather than quotation. The three quotations we do print — unlicensed activity, supervision, and where liability rests — were read in full.
• Products we deliberately did not rank: Two names that circulate in competing listicles could not be confirmed to exist as trading products in 2026 and are not named here. HouseCanary, Zillow Showcase and the multifamily and transaction-management group — including property-management leasing AI, transaction platforms and portal media products — were not researched to the standard this ranking requires and were excluded rather than guessed at.
• The Real Geeks acquisition price is not verified: Its ownership is SEC-confirmed at 100% within the FNF group; the acquisition price that circulates on aggregator sites is not, so we do not print it.
• Our cost bands are ours: The Frenchy Digital figures are our 2026 scoping bands for senior-led delivery, not an industry benchmark. Read them the way you should read every other vendor number on this page.

None of this argues against deploying AI in a brokerage. It argues for deploying it in the order of consequence — classification and routing first, drafting with a human send second, two-way text against a proven consent record third, AI voice last or not at all — and for pricing the decision on cost, verifiability and exit rather than on a return nobody can substantiate. The firms that get value from this work are the ones that found out what they were permitted to do before they found out what the model could do.

And the boundary holds throughout. An agent reads, classifies, routes, drafts and proposes. A licensee commits anything that constitutes advice, anything that publishes, and anything that contacts a consumer without a consent record you can produce on demand. Responsibility rests with the licensee and their responsible broker, not the technology provider — the regulator has already said so in writing.