A small estate agency should start an AI chatbot with three jobs: capture and route enquiries, collect viewing requests, and arrange valuation callbacks. Each gives the team a useful record without asking the bot to judge a buyer or invent property facts. An agent must take over for uncertain answers, access promises, negotiations and any valuation opinion.
These jobs can be launched in sequence. Begin with enquiry capture because it needs the least live integration; add viewing requests when property status and calendars are dependable; add valuation requests when the branch has a clear callback owner. A bot that answers beautifully but sends no usable record to an agent is an extra inbox, not a service.
Design the record an agent needs at the end of every chat
Before choosing software, write down the handover record. It needs the visitor's contact method and permission to use it, the property or service requested, the exact question, the bot's answer or uncertainty, the time of enquiry and the named person or queue responsible for the next action. Keep the original conversation alongside a short summary. The summary may omit a qualification or turn a tentative preference into a firm instruction.
Do not ask for more personal information than the task requires. A viewing request may need a name, contact route, property ID and preferred time. It does not need a guessed income band or a narrative about family circumstances. A valuation callback needs a property address and contact route, but the agent can request sensitive supporting documents later through an approved channel. Make the bot explain why it is collecting a field and let the visitor switch to a person.
An illustrative handover for a website enquiry reads: “Property P-17; visitor asks whether the current service charge includes parking; email supplied; bot did not answer because the approved listing has no parking figure; follow-up due next working period; owner: listing agent.” That is a better result than “warm lead interested in parking”, which loses the question. The agent can check the source and answer it accurately.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowFirst flow: capture an enquiry and route the exact question
The chatbot should greet a visitor, ask which property or service they mean, capture the question in their own words and offer a contact route. It can answer a small set of stable, approved facts: opening hours, the agency's general process, or a property fact read directly from a current listing record. It should not pretend to know a missing service charge, lease term or availability status. “I can ask the agent and get back to you” is a useful answer when the source is uncertain.
Start with five intent labels: property question, viewing, seller valuation, existing transaction and other. These are routing labels, not lead quality scores. The bot can ask one clarifying question if the property ID is missing, then pass the conversation on. Do not force a visitor through six questions to reach a person. A question about a complaint or a transaction delay should go to a team member even if the visitor also mentions a property.
For example, a visitor writes, “Is the flat with the balcony still available, and are pets allowed?” Two live listings may have balconies. The bot asks for the listing link or address. If the visitor cannot provide it, the bot records the words and routes them to the desk rather than choosing the first match. If the pet policy is not an approved listing field, it marks that part for an agent. The sample outcome is one enquiry with two unresolved questions, not two invented answers.
A separate illustrative error is a bot that replies “available” from yesterday's cached listing after the property is withdrawn. The correction is operational: update the approved data source, define how quickly changes propagate and test withdrawal with a dummy listing. In the live chat, the bot should say that availability is subject to confirmation unless the underlying system truly reflects current status. The agent takes ownership of a promised follow-up.
Measure this first flow by usable handovers, not messages sent. In a two-week pilot, count how many enquiries contain a property reference, exact question, valid contact route and named owner; count how many people ask for an agent and how quickly they reach one. Review a sample of complete transcripts for missing context. If the bot reduces obvious unanswered questions but creates vague CRM notes, it needs a better handover schema. For the wider routing design, see chatbot lead capture with a human hand-off.
Second flow: request a viewing, then confirm it properly
A viewing chatbot can show suitable request windows, collect preferences and create a task. It can confirm an appointment only when the calendar, property access instructions, occupant agreement and agent assignment are all current and connected. Most small agencies should begin by saying “request received” until a coordinator confirms. That phrasing avoids the bot offering a real-looking appointment that nobody can honour.
Use four statuses in the CRM: requested, proposed, confirmed and cancelled. The status should be visible to the agent and reflected in the message to the visitor. A proposed slot is not a booked viewing. Each property also needs a simple access rule: notice period, available windows, key holder, and whether another party must agree. The bot should not expose an occupant's details or the agent's whole diary.
Consider an illustrative visitor requesting 16:00 tomorrow for listing P-22. The agent calendar is clear, but the occupant requires a day's notice and has not yet agreed. The bot records 16:00 as a preferred slot and offers an alternative. Its output is “We have your request for 16:00. A coordinator will confirm access before the appointment is booked.” The coordinator checks with the occupant, changes the CRM status to confirmed and sends the final address and instructions. If access is refused, the coordinator offers the alternative without the bot silently moving the appointment.
Test the edge cases before launch: two visitors request the same slot, a listing is withdrawn, the assigned agent is absent, an occupant changes the access window, or the visitor asks to cancel. In a good cancellation flow, the bot records the change and notifies the agent; it does not simply produce a polite reply while the calendar appointment remains. Include a clear human contact path for same-day changes.
One filled test can be recorded like this: “P-22, 16:00 requested, access not confirmed; expected bot status: requested; expected message: no promise; expected task: coordinator queue; result: pass.” Another test uses a cancelled listing and expects no booking options. If the bot still offers one, disable booking until the listing feed is fixed. Out-of-hours valuation booking explores a related appointment flow; viewing access needs its own checks.
Third flow: take a valuation request for an agent
For valuation enquiries, the bot's job is to arrange a sensible next step, not to appraise the property. It can ask for the property address, type, contact details, preferred callback time and whether the visitor wants a sale or letting discussion. It can acknowledge any details the visitor volunteers, but should not use them to calculate a price in conversation. The agent checks the property information, comparable evidence and purpose before giving an opinion.
An illustrative seller says, “We extended the kitchen and need an estimate this week.” The bot takes the address, asks for a convenient contact period and records “extension mentioned; details unverified”. It should not respond with “the extension adds 15%” or invent nearby transactions. The callback owner can ask about measurements, permissions, condition and timing. This preserves the useful context while leaving judgement with someone who can inspect the evidence.
Valuation flow quality depends on ownership. Set a maximum internal time to assign a callback and a method for reassigning it if the agent is away. Do not promise a response time in the bot copy unless the branch can meet it. If the visitor asks for an instant number, explain that the chatbot can arrange an assessment and that any figure needs an agent's review. If the branch uses a separate automated valuation product, present it as a distinct service with its own data and disclosures, not as a hidden capability of a general chat widget.
For an illustrative seven-day trial, ten valuation requests arrive through chat. Eight have usable contact details, one gives an incomplete address and one asks for a number without a callback. The bot should create eight ready tasks and two exceptions for a person, not report ten “qualified valuations”. The branch checks whether the contact details are accurate and whether each visitor got the promised follow-up. That is a meaningful success measure.
Give the bot a small, maintained answer library
A chatbot needs an answer source as much as it needs a conversation script. Make a table with question, approved answer, source owner, last review date and escalation condition. Start with a dozen questions the front desk actually receives. Some will be agency-process questions that change rarely; others, such as property price and availability, need a direct live record or a clear “please confirm with an agent” response. Do not upload a folder of old brochures and assume the bot will choose the latest one.
A filled illustrative entry reads: “Question: is P-22 available for viewings? Source: active listing status and access note; owner: listing agent; last checked: current record; response: request a slot, subject to access confirmation; escalate when status is withdrawn or access note is missing.” Another entry is deliberately outside scope: “Question: will the seller accept an offer below the list price? Response: an agent will discuss your offer; never predict acceptance.” These two rows show why an answer library is a policy as well as a technical data feed.
Ask the bot to show its source reference to staff in the transcript, even if the visitor sees only a simple answer. When a wrong answer is reported, staff can find the record to correct. Give one person authority to retire an answer immediately. A nightly data refresh may be acceptable for stable process wording but too slow for withdrawn listings. Test the actual update delay with a dummy change. If correction is difficult, narrow the bot's allowed answers and rely more on human callbacks.
The assistant can help prepare drafts for this library, but an agent must approve the content. For instance, a generic model might write “all properties have flexible viewing times” because that sounds helpful. The branch's actual instruction may be “appointments depend on access and occupant agreement”. Publish the latter, and use the same language across website, chat and desk scripts. Inconsistency between channels creates more work than the bot saves.
Useful later flows, once the first three are reliable
Approved property questions
After enquiry routing works, the bot may answer a narrow set of property questions from fields the agency maintains: room count, listed amenities, published opening arrangements for a development, or a link to the current brochure. Each answer needs a source and a last-updated time. If a visitor asks whether a loft is legally usable as a bedroom, a brochure label is not enough; route the question to an agent. Avoid converting a seller's unverified statement into a published answer.
For example, a listing form records two parking spaces while a newer agent note says one space is blocked by a shared access agreement. The bot should stop answering parking-capacity questions until the discrepancy is resolved. An explicit “under review” status is more useful than letting it pick whichever record was indexed first. A small, trustworthy answer set is better than a large knowledge base nobody maintains.
Existing-transaction status
A buyer or seller may want to know what happens next. A bot can point to generic process information or pass a status request into the transaction queue. It should not reveal private file details until identity and access are verified, and it should not interpret a contract or say that a condition has been met. If the transaction system has approved status messages, an agent can choose which are safe to share through the authenticated channel.
Suppose a visitor writes “Has the other side signed?” from a new email address. The bot should not match the surname and reveal a status. It can acknowledge the question and route it to the assigned coordinator, who verifies the person and checks the document. This is a practical reason to keep transaction status out of the initial website bot even though it appears to be a simple FAQ.
Maintenance and tenant messages
If the agency also manages properties, a bot can collect a maintenance description, property reference, contact route and urgency indicators, then route the request. It should not diagnose safety issues or promise a contractor attendance time that has not been booked. Separate a routine query from an urgent situation with a clear path to immediate human contact. The property management team must decide what can be said from an approved policy and what needs a professional assessment.
An illustrative message says “water is entering through the ceiling”. A chatbot that suggests ordinary booking slots could delay attention. The safer flow gives the agency's verified urgent contact route and creates an immediate task. The bot can preserve the visitor's exact words and time, but a person decides the response. This later use case needs a different escalation design from sales enquiries.
Write handover rules as visible conditions
Make the stop conditions part of the bot specification. Handover is required when the visitor asks for a person; when the answer is absent, stale or contradictory; when a complaint or urgent issue appears; when a price, legal status or financing question needs judgement; when a property is withdrawn; or when the bot has failed to understand after one clarification. Give the visitor an immediate acknowledgement and give the team a task with the transcript attached.
Do not use a vague “confidence score” as the only safeguard. A bot can be highly confident in an outdated property fact. The safer test is whether the answer is tied to an approved, current record and whether its topic is within the allowed scope. Record the rule that caused handover. This lets the agency distinguish a good escalation from a broken conversation. Human handover for AI chatbots covers the general service design.
A filled illustrative escalation rule reads: “If the property price in the chat source differs from the current listing record, give no price, create a priority listing-data task, and send the visitor a neutral follow-up message.” Another reads: “If a visitor asks whether their family is suitable for a property, provide the same factual application process offered to everyone and route any individual decision to an agent.” These rules are testable; “be fair” alone is not.
Use a queue that shows who owes the next action
Many chatbots appear successful in the chat window while follow-ups vanish in an unowned email box. Set one CRM task for each unresolved conversation, with a deadline, owner, property reference and the visitor's exact question. The agent should be able to mark answered, waiting for a third party, duplicate or closed at the visitor's request. A closed task needs a reason. A bot saying “someone will contact you” should create a visible obligation, not a pleasant ending to a transcript.
For a small branch, a coordinator can review the queue twice a day. Imagine 24 overnight chats: nine property questions, seven viewing requests, four valuation requests, two existing-transaction queries and two unrelated messages. The coordinator checks that the seven viewing tasks are still requests, routes the transaction queries to the correct people and assigns each valuation callback. One property question concerns a withdrawn listing, so it becomes a data correction as well as a response. These numbers are illustrative; their value is showing what the morning work actually contains.
Use a simple breach check at closing time: list every unresolved task older than the branch's own response target and ask why it is still open. A missing owner suggests a routing fault. Repeated “waiting for information” suggests the bot is promising answers the data cannot supply. If visitors keep asking for a person after receiving a canned answer, shorten the bot flow. Let the operational evidence change the design rather than adding more conversational text to cover a broken hand-off.
Guard against unequal access and misleading property claims
Housing advertising and access rules deserve explicit review. Ad targeting can push property information towards some people and away from others even when nobody writes discriminatory copy. A chatbot can also create unequal service by prioritising people based on inferred characteristics, language style or a guessed budget. Use neutral routing categories, provide the same factual information to comparable enquiries and audit response times. Ask a qualified adviser to check the requirements that apply to the agency's market and business model.
The source-of-truth problem is more common than a dramatic chatbot failure. A price changes, a listing is withdrawn, an amenity is corrected or a viewing restriction is added. Decide who updates the underlying record and when the bot refreshes it. Then test a change end to end. If the bot cannot say when its property data was last updated, do not allow it to present time-sensitive facts as current. The broader eight-job brokerage map helps decide whether chat is even the right first AI project.
Run a 20-conversation test before public launch
Write test chats that match the branch's real work: eight ordinary property questions, four viewing requests, three valuation requests, two ambiguous property references, one withdrawn listing, one complaint and one request for a person. Include short messages, spelling errors and a visitor who changes their mind. For each, state the expected answer, CRM fields, owner and handover condition before running the test. A pleasant tone is not a pass if the record is incomplete.
For an illustrative agency, the test passes 16 conversations, fails two because an old listing remains searchable and fails two because the booking status says “confirmed” before access approval. The agency fixes the feed and booking language, then reruns those four and several neighbouring cases. It does not launch on the strength of 16 good conversations. Once live, review the first 30 real transcripts and keep a visible route to a person while the team learns what visitors actually ask.
Measure captured enquiries that received an accurate human response, viewing requests that became confirmed appointments, valuation callbacks completed and cases escalated correctly. Also count wrong property facts, missing transcript fields and visitors who gave up before a handover. These measures reveal whether the bot is making the branch more responsive. The aim is a short, dependable first conversation that gives an agent enough context to do the work only a person can do.
Questions before putting the bot on the site
Should the chatbot answer property questions from its own knowledge?
No. Connect it to a maintained, approved property record or keep it to collecting the question for an agent. If a fact such as price, availability, access or charges is absent or stale, it should say it cannot confirm it. Test how quickly a withdrawn listing disappears from its answers.
Can a chatbot confirm a viewing automatically?
Only if the booking system has current property access, agent availability and any required occupant confirmation. Otherwise the bot should collect preferred slots and label them as requests. A human confirms the appointment and sends the final details. Count requested and confirmed viewings separately in your reports.
Can it give a valuation range?
A small agency should begin by taking the valuation request and arranging a qualified agent's assessment. Do not let a general chatbot invent comparable transactions or promise a price. If you later use a specialist valuation system, verify its data, assumptions, disclosures and applicable professional requirements before presenting any figure.
Further reads
- Real Estate AI Chatbot Features: A Checklist Before You Buy — Check the features needed for these three flows.
- Real Estate AI Chatbot Demo: A Test Checklist Before You Sign — Run a difficult-question demo before signing.
- Real Estate Chatbot Pricing: What Agents Pay for AI Lead Capture — Budget after agreeing the scope.
- AI Receptionist for Estate Agents: Cost, Call Flow and Handover — Compare chat with telephone cover.
- AI Automation Agency for Real Estate: Scope, Fees and Red Flags — Typical scope, retainer versus project fees, red flags and an in-house alternative for a real estate business weighing an AI automation agency.
- ChatGPT for Commercial Real Estate: Source-Checked Reports and Comps — A commercial broker's method for ChatGPT market reports, comps summaries and owner emails, with a check on every figure before it reaches a client.
- Real Estate AI Chatbot Examples: Five Flows to Copy for Leads — Five property enquiry scripts show what the bot should ask, what it should record and when an agent must take over.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.