Before buying a real estate AI chatbot, check eight things: lead capture fields, qualifying questions, viewing booking against agents' calendars, CRM sync, handover to a named agent, accurate listing data, compliance wording, and reporting you own. Any "must" that fails in a live demo is a reason to walk away, whatever the price.
Most feature lists on vendor sites read as brochures. This one assumes you are a small agency with a handful of agents and a website, and gives each item a pass mark and a way to test it in under two minutes. For what the products cost, see the tutorial on what agents pay for AI lead capture; the checklist below is about what you should get for the money.
Must, Should and the six-agent agency used throughout
Items are marked Must or Should. A Must is a feature without which the chatbot loses you leads or creates risk. A Should is an improvement you can live without for the first quarter. Run the checks yourself with a real listing, a real calendar and a phone, rather than watching a recorded demo.
The worked example throughout is a six-agent independent agency with about 400 website enquiries a month, 35% of which arrive outside office hours. Those figures are made up, so swap in yours.
Group 1: Capturing the lead
Capture fields you can change (Must)
A bot that asks only for a name and email wastes the conversation. A useful one asks for the enquiry type (buy, sell, rent, value), property or area, budget or price expectation, timescale, and best contact method. Make sure you can add, remove and reorder fields without calling the vendor.
Test it: start a chat as a seller and as a buyer. The questions should differ. If both flows ask the same five things, the bot cannot qualify.
A completed record from a good flow looks like this (illustrative):
Type: Seller (valuation)
Property: 3-bed semi, 12 Maple Close [illustrative]
Timescale: Listing within 8 weeks
Price expectation: $380,000 to $400,000
Contact: Phone, weekday evenings
Consent to be contacted: Yes (ticked, 14:02)
Source: Website chat, valuation page
Transcript: AttachedConsent captured as a field (Must)
A lead you cannot lawfully contact is not a lead. The chatbot should record a consent tick or the exact wording the visitor agreed to, with a time stamp, and pass it to your CRM. Check the consent line in the demo; it should not be pre-ticked.
Qualifying without interrogating (Should)
Three or four questions beat ten. Look for the ability to skip a question when the visitor has already answered it, and to let a visitor leave a number and get on with their day. See the tutorial on five lead flows to copy for working question sets.
Group 2: Booking viewings
Live calendar check (Must)
The bot should offer times that are free in the responsible agent's actual calendar, not a list you update by hand. Ask the vendor which calendars it reads and how often it refreshes.
A realistic failure: the bot offers Saturday 10:00 for a property, the booking is made, and the agent is already at another viewing across town. The cause was a calendar that synced hourly and an overlapping booking made by phone. The fix is a calendar sync in minutes, a buffer between viewings, and travel time rules. Test it by booking two viewings for the same agent at the same time from two browsers.
Confirmation, reminders and changes (Should)
The visitor should get a confirmation with the address and a way to reschedule, and a reminder before the viewing. No-shows fall when reminders work; whether they do for your agency is something to measure after launch, not assume.
Rules by property type (Should)
Some properties need accompanied viewings, keys collected from a seller, or two-person visits. The bot should treat them differently from a vacant flat. If it cannot, it should hand over rather than book.
Group 3: Connecting to your CRM
Contact creation with the right fields (Must)
Ask for a field-by-field map, not "we integrate with your CRM". The table shows the minimum worth confirming.
| Chatbot field | CRM field | If it is missing |
|---|---|---|
| Name, phone, email | Contact record | Agents retype and make mistakes |
| Enquiry type and property | Lead source and interest | You cannot report by property |
| Consent and time stamp | Permission field | You cannot prove you may contact them |
| Full transcript | Note or timeline | The next agent asks questions already answered |
| Viewing booked | Task or appointment | Agent finds out from the calendar only |
Duplicate handling (Must)
A returning visitor with the same email should update the existing contact, not create a second one. Test by entering the same email address in two chats and checking the CRM afterwards.
Sync failure alerts (Should)
Integrations break quietly. Ask what happens when the CRM rejects a record. The answer should be a visible error and a retry, not a lost lead. If your CRM is a spreadsheet or inbox, a daily export is acceptable as a stopgap; HubSpot, for instance, has a free CRM tier, which can serve as a simple destination for a small agency, though you should confirm the chatbot's integration with it directly.
Group 4: Handing over to a person
Trigger rules you can edit (Must)
Write the triggers before you demo. Typical ones: a visitor asks for a valuation figure, mentions a legal dispute, expresses frustration, asks for a person, or the bot fails to answer twice. In each case the bot should stop trying and pass the chat.
Visitor: What is my house worth? I need to sell quickly, there's a divorce.
Bot (good): I'm sorry you're dealing with that. I can't give you a value,
but [agent first name] can. Is this the best number to call you on, and
when suits you today?
Bot (bad): Based on similar properties your home is probably worth...The "bad" answer breaks two rules: it invents a valuation and ignores a sensitive situation. Test with exactly this kind of message.
What the agent receives (Must)
Handover is only as good as the packet. The agent should get the full transcript, contact details, property and the reason for the handover, by the channel they actually use: a text, a CRM task or a chat app. Check how fast. An alert arriving 20 minutes later has missed the window in which an enquirer is still on your site.
Out-of-hours behaviour (Should)
With 35% of enquiries arriving out of hours, the weak point is what the bot promises. It should say when someone will respond ("tomorrow morning before 10") only if your rota supports it, and should never offer to "get an agent now" at midnight.
Group 5: Knowing your properties
A named data source (Must)
Ask where the bot learns about your listings: a feed, a CSV, an upload, or a crawl of your site. Then ask when it refreshes. A bot that describes a sold property as available damages trust. The tutorial on AI listing mistakes that mislead buyers shows what to check.
"I don't know" behaviour (Must)
Ask a question the data cannot answer, such as "Is the roof new?" when your listing is silent. The bot should say it does not have that information and offer an agent. If it improvises, it will do so on real buyers.
Subjective questions (Must)
Questions such as "Is it a safe area?" or "Is it good for families?" can lead a bot into steering visitors by who lives where, which equal-treatment rules in many places forbid. The bot should decline to characterise neighbourhoods and offer facts such as distance to a school or station only from a listed source. The leasing chatbot equal-treatment checklist goes into this in depth.
Group 6: Compliance wording
You do not have to be a lawyer, but you need three pieces of wording in place and a qualified adviser's view on whether they are right for your market.
- AI disclosure. If you have customers in the EU, the AI Act's transparency duty for chatbots has applied since 2 August 2026. Put it in the first message: "I'm an AI assistant for [agency]. A member of the team can take over at any time."
- Consent line. "By sending this you agree that [agency] may contact you about your enquiry. See our privacy notice." The visitor should have to tick or type something. Data-protection law such as the GDPR may apply to the transcript, so check retention and deletion.
- No advice disclaimer. The bot should not give legal, tax, mortgage or valuation advice. A one-line reminder and a hard rule in the configuration are better than a long disclaimer nobody reads.
Ask the vendor where transcripts are stored, how long, who can see them, and whether they are used to train models. If they cannot answer in writing, that is a failure on a Must. If you plan to use WhatsApp as a channel, note that Meta's platform terms bar general-purpose AI assistants while allowing bots tied to one business task, so a property-enquiry bot is the permitted kind, but a catch-all assistant is not. WhatsApp's Business Messaging Policy also expects "prompt, clear, and direct escalation paths" when you automate, which makes the Group 4 handover a platform requirement on that channel as well as good practice.
Group 7: Channels, language and look
- Website chat on mobile (Must). Most property browsing is on a phone. Open the chat on a mid-range handset and complete a booking.
- Other channels (Should). Facebook, Instagram or WhatsApp matter only if enquiries already arrive there. Do not pay for channels you cannot staff.
- Languages (Should). If a meaningful share of your enquiries comes in another language, test the bot in it, including the consent line.
- Your brand and tone (Should). A name, a colour and a greeting in your voice. Cosmetic, but a stranger's logo on your chat window looks careless.
Group 8: Reporting and ownership
You should be able to answer four questions from a dashboard without asking the vendor: how many chats started, how many left contact details, how many viewings booked, and how many were handed over. You should also be able to export your transcripts and leads and to cancel with your data intact. Ask what happens to your data after cancellation, and get the answer in the contract.
For the measurements worth taking in the first month, and the questions to put to the vendor live, use the demo test checklist. That one is about challenging a vendor in the room; the checklist here is about what the product should contain.
Six test chats that catch most weak bots
Run these on any shortlisted product, using a real listing and a phone as well as a laptop. Each takes a couple of minutes, and together they test almost every Must above.
- The ordinary buyer. Ask about a listed three-bed: price, bedrooms, parking, and a viewing on Saturday. Pass: correct facts, a live slot, a confirmation.
- The seller in a hurry. "I want to sell in four weeks, what will you get me?" Pass: no promise or valuation, a request for contact details and a handover.
- The unanswerable question. "Has the boiler been serviced this year?" when the listing does not say. Pass: "I don't have that" and an offer of an agent.
- The steering question. "Which street is best for a young family?" Pass: a refusal to characterise areas and an offer of listed facts.
- The impatient visitor. Ask for a person three times. Pass: handover on the first request, not the third.
- The repeat visitor. Use the same email twice. Pass: one CRM contact with two entries, not two contacts.
Record the exact reply to each in a sheet, with a pass or fail. Vendors who have good answers are happy to see you do this; vendors who steer you back to the polished demo are telling you something.
Answers from a vendor that should worry you
Some replies sound like answers and are not. Here are four, with what to ask next.
| What the vendor says | What it may mean | Ask next |
|---|---|---|
| "We integrate with all major CRMs." | A generic connector, maybe via a third-party tool at extra cost | "Send me the field map for my CRM, and say who maintains it." |
| "Our AI is trained on real estate." | A general model with a prompt, which may still guess | "What does it do when it does not know? Show me." |
| "Handover is automatic." | An email alert with no transcript | "Show me what the agent receives, and how fast." |
| "We're fully compliant." | A marketing line, not a document | "Which duties, and where is the written policy?" |
The first week after you switch it on
Launch carefully rather than all at once. This sequence keeps risk low for a small agency.
- Day 1: turn the bot on for one low-stakes page, such as the rentals page, with the disclosure and consent lines in place.
- Days 2 to 3: read every transcript the same evening. Note wrong answers and tidy the knowledge source.
- Day 4: check the CRM: every lead should have consent, source and transcript attached.
- Day 5: check handover timing with a real agent on a real phone, outside the office if you can.
- Days 6 to 7: extend to sales pages if the first five days were clean, and set a weekly review for 20 random transcripts.
A small illustration of why the daily read matters: an agency noticed on day two that visitors kept asking "Is it still available?" about a property that had gone under offer. The listing feed refreshed overnight, so the bot kept saying yes during the day. The fix was an hourly refresh and a rule that any status other than "available" triggers a handover. None of that would have shown up in a vendor demo.
Scoring two vendors on the same sheet
Here is an illustrative score for the six-agent agency. Weight each Must at 3 and each Should at 1. Score each item 2 for a full pass, 1 for partial and 0 for a fail, then multiply by the weight.
| Item | Weight | Vendor A | Vendor B |
|---|---|---|---|
| Editable capture fields | 3 | 2 | 2 |
| Consent recorded | 3 | 2 | 0 |
| Live calendar booking | 3 | 1 | 2 |
| CRM field map and duplicates | 3 | 2 | 1 |
| Editable handover triggers | 3 | 2 | 2 |
| "I don't know" behaviour | 3 | 1 | 2 |
| Reminders and rescheduling | 1 | 2 | 1 |
| Multi-language | 1 | 0 | 2 |
| Weighted total (out of 40) | 32 | 30 |
Vendor A scores 32 against Vendor B's 30, which is close, and notice that Vendor B fails the consent Must outright. A rule that any failed Must is a stop makes Vendor B a "no" whatever the total, which is why you keep the pass marks separate from the score.
What the first-year value looks like
Back to the agency: 400 enquiries a month, 35% out of hours is 140. Suppose the bot turns 10% of those into booked viewings, which is 14 a month, and 1 in 10 viewings leads to an instruction or a sale for a $2,500 fee. That is 1.4 deals and $3,500 a month. These are assumptions to check against your own conversion, not benchmarks. The calculation shows what the chatbot has to do to cover its fee, and the checklist shows whether it can. If the break-even is two extra viewings a month, almost any bot that passes the Musts will clear it; if it is 20, look elsewhere.
Where checklists stop helping
A checklist cannot tell you how your agents will treat the leads. If handovers are ignored, the best bot will not fix that. Agree a response rule first: every handover is answered within 15 minutes in office hours and by 10 am the next day otherwise. Put the rule in the CRM task, and read a sample of 20 handovers weekly for the first month.
If you want a second opinion on your shortlist, that is part of what my AI implementation consultation covers; start from the contact page. Before then, if your questions need a written source, the tutorial on building the FAQ an AI chatbot needs is the place to start.
Questions agencies ask when choosing a chatbot
Which feature is the most common deal-breaker?
Handover. A chatbot that captures a lead but cannot pass the conversation, with its history, to a named agent in real time leaves the buyer repeating themselves. Test it during the demo with a request the bot should not handle, and watch what the agent actually receives.
Do I need CRM sync on day one?
If leads land in an inbox nobody checks, yes. Even a basic sync that creates a contact with the enquiry source, property and message saves retyping and stops duplicates. If your CRM is a spreadsheet, a daily export may be enough for the first month.
Can the chatbot answer questions about a specific property?
Only from data you give it. Check the source: your listing feed, a document upload or the website. It should say it does not know rather than guess, and it should show when the listing was last refreshed so withdrawn properties are not described as available.
Does a chatbot have to say it is an AI?
Often, yes. If you have customers in the EU, transparency duties under the AI Act require telling people they are talking to a chatbot, and many platforms expect it. Put a plain disclosure in the opening message and ask a qualified adviser about wording for your market.
Further reads
- How Estate Agents Use AI to Prepare Valuation Appointments — Prepares valuation appointments once the lead arrives.
- How to Set Up Missed Call Text Back for a Service Business — Catches the enquiries that arrive by phone instead.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: the EU AI Act Article 50 transparency duties (applying since 2 August 2026) and Meta's WhatsApp platform terms as summarised in the series fact sheet; the agency figures in the examples are illustrative.