Residential Realtor Playbook: Lead Replies, Showings and Client Memory

An AI agent can help a residential realtor answer new inquiries within minutes, send showing confirmations and follow-ups, and keep a record of what each client asked for. The safe starting point is narrow: one role, one channel, a short list of approved facts, and firm rules for anything involving price, offers or fair housing. Client memory should sit in systems the agent reads through permissioned tools, so the agent or the broker can inspect, correct and delete it.

Key takeaways

• Give the agent one role first, either first replies to new inquiries or follow-up after showings, never both on day one.
• Limit what it can state to facts you have already published, such as listing details, showing windows and your contact rules.
• Keep pricing opinions, offers, negotiation and any fair housing question with a licensed person who reviews the handoff.
• Store client memory where you can read, correct, export and cut it off, and write down who can see each record.
• Define one observable success criterion, such as every inquiry having a logged first reply inside the response window you set.

Why lead replies come first

Inquiries arrive from listing portals, sign calls, website forms and messaging apps, often outside office hours. Each delayed reply gives the prospect a reason to contact the next agent on their list. An agent with a narrow job can handle the repetitive part: acknowledging the message, confirming the area or property of interest, asking the questions you already ask on a first call, and offering showing times from a calendar you approve.

It should not improvise about price, whether a home will sell, neighborhood safety claims, or anything that could be read as steering a buyer by protected characteristics. Those messages go to a person with a short handoff note. The handoff note matters more than the reply itself. A useful note records the source, the property or area, the stated timeline, the questions already answered, and what the prospect is waiting for. When the agent cannot answer, it says so and names when a human will reply, and that promise should be one you can keep.

Showing follow-up that keeps you in charge

Showings create small obligations: confirmations, directions, reminders, a post-visit note and a check on next steps. These are where leads quietly go cold. A narrow agent can run this sequence from templates you wrote and a calendar you control.

• Step 1: Confirm the showing as soon as it is booked, with the address, arrival instructions and the name of the person meeting the client.
• Step 2: Send one reminder before the showing, with a single question if the client needs to reschedule.
• Step 3: Within the window you set, ask for feedback using three fixed questions: what stood out, what did not fit, and what would make a second visit worthwhile.
• Step 4: Write the feedback to the client record, tagged by property, and flag anything that changes the search criteria for your review.
• Step 5: Draft the next step, such as a shortlist or a request for a decision date, and send it only from an approved template or after your review.
• Step 6: Close the thread when the client asks to stop or when the sequence ends without a reply, and record the reason.

Keep the templates short and specific. A reminder that says "Your showing is at 4 pm at the side entrance, and Dana will meet you there" is easier to verify than a friendly paragraph that adds details nobody approved.

What client memory should hold, and what it should not

Memory is what makes follow-up feel informed, and it is also the largest risk in the setup. Keep what the client gave you directly and what the job requires: contact preferences and consent, stated search criteria, properties viewed, feedback given, dates of past contact and open tasks. Be cautious with stated price ranges. Store them only when the client offered them and the next step needs them, and show them only to people who need them.

Do not let the agent infer sensitive traits from names, neighborhoods or the tone of a message, and do not let it pull outside data to build profiles. Set a retention rule, such as removing inactive inquiry records after a period you choose with your counsel, and test that deletion works in every store where memory actually lives. The [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) gives a practical checklist for documenting these choices: what data is used, who can change it, and how errors are found and fixed.

Permissions, and why inbound messages are not instructions

A lead can type anything into a message, including text that reads like a command to the agent, such as a request to ignore its rules and send a seller's contact details. The agent must treat inbound text as content to read, never as an instruction that changes what it is allowed to do. This class of failure is known as prompt injection, and the [OWASP guidance on prompt injection prevention](https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html) describes defenses such as separating instructions from user content and limiting what each tool can do.

Give each tool the smallest scope that does the job. A reply tool sends only from approved templates. A calendar tool can offer and hold slots but cannot cancel other bookings. The agent has no access to offer documents, seller contacts or listing records it does not need. Log every action with the time, the input it used and the tool it called, so you can review it later. The [OWASP authorization guidance](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html) is a sound reference for checking who may read and act on each record.

An illustrative week with one solo agent

Illustrative scenario, not a real account: a solo agent has an agent handling first replies from her website form and showing confirmations. On Tuesday evening a prospect asks whether the seller would accept a lower offer. The agent does not answer. It tags the message as price and offers, tells the prospect that a licensed agent will reply the next morning, and places the note at the top of her queue.

On Wednesday, after a showing, the agent sends the three feedback questions and records two comments about noise and storage. The client also mentions that a smaller floor plan would suit them better, so the agent flags the change to the search criteria instead of updating them on its own. The agent drafts a shortlist, and the agent approves the draft with one edit.

At the end of the week, her log review shows that every inquiry received a logged reply and that one template promised a callback time she could not keep. She changes the template that same day and reruns the sample of past inquiries before the next week starts.

How Be Human runs this as a managed service

Be Human starts with one useful role, usually lead replies or showing follow-up. The first engagement defines the memory and tools that role needs, the approval boundaries a licensed agent must hold, and one success criterion that can be observed in the logs. Managed operations then keep the setup current: monitoring of replies and failed actions, recurring evaluations against past inquiries, maintenance of calendar and messaging integrations, periodic permission reviews, incident response when something goes wrong, and controlled changes to templates and rules. The work is remote, in English or Portuguese, and the client keeps the authority to approve, correct or stop the agent at any point.

Scale only after the first role is dull

Add a second role only after the first has run for several weeks without surprises in your own review. Good candidates are nurture messages to past clients or status updates for transaction files. Each addition needs its own memory rules and its own test set, because a message that is harmless in one workflow can create a record problem in another. Before scaling, rerun the first role against a fixed set of past inquiries, including the awkward ones, and confirm that the answers still follow your rules. The marketplace agent notes on trust and disputes cover a similar discipline for records that many parties can see, and the property management tenant workflows show how maintenance and follow-up share the same approval questions. If you want to find the first task worth handing to an agent in your business, take the free Be Human AI diagnostic.

Frequently asked questions

Can an AI agent reply to real estate leads without a person reviewing each message? For a narrow set of routine replies, yes, but only if the facts and templates were approved in advance. Anything about price, offers, legal questions or fair housing should route to a licensed person with a handoff note. Review a sample of replies every week during the first period so you can see what the agent actually sends.

Where should client memory live? Keep it in systems you already own, such as your CRM or calendar, and let the agent read and write through scoped tools. You need an export path and a deletion path so you can correct or remove records on request. Avoid copying memory into a private store that you cannot inspect.

How do you know whether the agent is working? Measure the success criterion you chose before starting, using the logs over several weeks rather than impressions. Each time you change a template or a rule, rerun the same set of past inquiries and compare the results. A change that improves one reply but breaks the approval boundaries is not an improvement.

What should a first engagement include? A good first engagement defines one role, the memory and tools that role needs, the approval boundaries, and an observable success criterion. Managed operations can follow with monitoring, recurring evaluations and permission reviews. You should receive a written scope before any system access is granted.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.