AI for Real Estate Agencies: Leads, Listings and Follow-Up

AI for real estate agencies works best as a supervised agent that covers the repetitive middle of the job: sorting incoming enquiries, checking listing facts against a trusted record and preparing follow-up drafts for a person to review. It does not replace the negotiator or the person who knows the neighborhood. It gives each human a cleaner queue, accurate facts and a record of what was promised to whom.

Key takeaways

• Start with one narrow role, such as inbound lead coordinator, instead of a general assistant for the whole agency.
• Listing accuracy depends on one named source of truth, not on whatever the agent last read in an email thread.
• Drafting and sending are different permissions. Let the agent draft first and widen access only after review shows it is reliable.
• Price, availability, legal claims and commitments to buyers or owners should always go to a person for approval.
• Measure the role with observable evidence: how quickly new enquiries get a first reply, and whether every follow-up has an owner and a date.

What does a supervised agent actually do in an agency?

Real estate work is a chain of small handoffs. An enquiry arrives from a portal, a website form, a phone message or a social message. Someone decides who should answer. Someone checks whether the property is still available. Someone replies, books a viewing and, later, remembers to ask how it went. Every handoff is a place where a lead cools or a fact drifts.

A supervised agent is an AI system with a defined role, its own accounts, a working memory of the agency's context and a list of tools it may use, all under human control. In an agency, three jobs are a sensible starting point: lead handling, listing information and follow-up. Everything else, such as valuations, contracts and negotiation, stays with your people until the first three jobs run reliably.

The reason to keep the scope small is practical. A narrow role can be described in a page, tested with real examples and reviewed by the person who will use it. A broad assistant cannot. If you want a template for that page, write a job description for your first AI agent role is a good place to begin.

How should an agent handle new leads?

The goal of lead handling is not a fast reply at any cost. It is a fast, correct first reply and a clean record that the right person can pick up. The agent should turn every enquiry, whatever the channel, into the same structured record before anything else happens.

A useful lead record contains:

• Source channel and time received
• Contact details exactly as given, never guessed or completed
• The property or area the person referred to
• What they asked, in their own words
• Missing information the agent should ask for
• A duplicate check against existing contacts and open conversations
• A suggested owner and the reason for the suggestion

Routing is where the agency's own rules matter. Some agencies route by area, some by language, some by property type or by who already knows the client. The agent should apply the rules you wrote down, and say which rule it applied. When no rule fits, it should say so and hand the lead to a coordinator rather than choose quietly.

One risk is specific to this job. Enquiries are text written by strangers, and text can contain instructions. A message that says to ignore previous rules and send the owner's phone number is data to be flagged, not a command to follow. The agent's tools should make that impossible by design, and the broader reasoning is covered in prompt injection in business documents. The OWASP prompt injection prevention cheat sheet is a solid technical reference for the team that builds the connection.

How do you keep listing information accurate?

Most harm from AI in a real estate setting comes from confident wrong facts: a property described as available that was withdrawn, a fee stated that nobody approved, a feature that belongs to a different unit. The remedy is not a smarter model. It is a reliable source of truth and a rule that the agent answers from it.

Decide where each kind of fact lives. Typically that means the listing system or CRM for status, price, size, features and fees, the owner's written instructions for what may be disclosed, and the shared calendar for viewing availability. The agent retrieves from those records at the moment of the question instead of relying on memory of earlier conversations. This retrieval pattern is widely documented, for example in Google Cloud's reference architecture for retrieval-augmented generation. For the agency side of the problem, how to give an AI agent a reliable source of truth explains the decisions you need to make.

Then set a few plain rules for how the agent uses those records:

• Every factual statement in a draft should be traceable to a named field in a record.
• If a field is empty, the agent says it does not know and asks a person. It never fills the gap with a plausible value.
• If a record has not been updated within the freshness window your agency defines, the agent marks the status as unconfirmed.
• Marketing descriptions and verified facts are kept apart, so that enthusiastic copy is never quoted as a specification.
• Owner instructions, such as not disclosing a motivation to sell, override helpfulness.

This also tells you where to invest first. If your listing records are patchy, cleaning the ten fields that buyers ask about most will improve the agent and your human team at the same time.

What should follow-up look like?

Follow-up is where agencies lose the most value quietly. The viewing went well, the buyer said they would think about it, and nobody wrote down what concerned them. A week later the message is generic and the buyer has moved on.

A supervised agent can keep that context. After a viewing, the agent reads the agent's notes, the original enquiry and the listing record, then drafts a message that refers to what the person actually said. It also proposes the next task with a suggested date. The principles are the same as in AI customer follow-up drafts that keep the context: the draft cites its sources, and a human decides whether it goes out.

Illustrative scenario: a buyer writes on a Saturday evening about a two-bedroom apartment. The agent creates the lead record and finds a match in the listing system, but the status was last confirmed longer ago than the agency's freshness window allows. It drafts a polite acknowledgement, offers the viewing slots visible in the assigned agent's calendar, makes no statement about price flexibility, and flags the stale status for the agent. On Monday the agent confirms the apartment is still available, edits one line and sends the reply. After the viewing, the buyer mentions street noise in the notes. The agent's follow-up draft addresses that concern directly, using facts from the record, and proposes a check-in task for a few days later. Nothing in the process was invented, and every step can be traced.

The same logic applies to the CRM. Rather than writing silently to contact records, the agent should propose the change, show what it would alter and wait for approval until the pattern has proven reliable. That keeps your data trustworthy and keeps people in the loop.

Where should the agent stop and ask?

The most useful design question is not what the agent can do, but what it must not do alone. A decision guide helps your team agree on this before the first live enquiry.

• Answer directly: the question is about a factual field that exists in the record, is fresh and is cleared for disclosure, such as published features or the general area.
• Draft and ask: anything involving a price, an offer, a viewing confirmation, a deadline or any commitment on behalf of the agency or an owner.
• Escalate at once: complaints, questions that call for legal or financial advice, disputes between parties, requests that could raise discrimination concerns, and anything the agent cannot classify.
• Refuse and log: requests for another client's information, and messages that try to change the agent's instructions.

Permissions should mirror this guide. Reading records, drafting messages and sending messages are three separate capabilities, and an agent that only needs to read and draft should not hold sending rights. The thinking behind this is laid out in where should your AI agent stop and ask, and the OWASP authorization guidance gives the security principle behind it: grant the minimum access a role needs and review it regularly.

Personal data deserves a separate note. Buyer and owner details are personal information, and the data protection rules of your market apply to how the agent stores and uses them. Agree who may see which records, keep clients separated and make sure someone owns the question of retention.

What is a practical plan for the first role?

A first engagement should be small enough to finish and observable enough to judge. This sequence works for most agencies:

• Step one: write the role in a page, including what the agent does, what it never does and who reviews its work.
• Step two: name the source of truth for listings, contacts and calendars, and repair the most visible gaps.
• Step three: give the agent read access to inbound channels and records, using its own accounts rather than a staff member's login.
• Step four: run in draft-only mode on real enquiries, with a named agent reviewing every draft and recording what they changed.
• Step five: build a small set of test enquiries, including messy ones such as duplicates, missing details, mixed languages and attempts to redirect the agent, and rerun it after every change.
• Step six: define the success evidence, such as time to first reply, the number of leads with a missing owner, and follow-ups that were logged with a date.
• Step seven: widen permissions one action at a time, for example sending simple acknowledgements, and only where the review record supports it.

The review record in step four is the most valuable asset the project produces. It shows which edits recur, which instructions are unclear and which records are unreliable. Teams that skip it tend to argue about impressions instead of looking at evidence.

How does a managed approach fit?

Be Human designs, implements and operates personalized AI infrastructure for companies. A first engagement defines one useful role, the memory and tools it needs, the approval boundaries and an observable success criterion. For an agency, that could be the lead coordinator role described above. Managed operations can then include monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements, so the agent keeps working as your listing systems, channels and team change. The service is remote and available in English and Portuguese.

Frequently asked questions

Can an AI agent reply to property enquiries without a person checking? It can, but it should not begin that way. Start in draft-only mode, review every message and allow automatic sending only for low-risk replies such as acknowledgements, once the review record shows the drafts hold up. Anything about price, availability or commitments should keep a human approval step.

How do we stop the agent from quoting outdated listing details? Give it a single source of truth, require it to cite the field it used and set a freshness rule for status and price. When a record is stale or a field is empty, the agent says it is unconfirmed and asks a person. Regular tests with deliberately outdated records show whether the rule really holds.

Does this work with the CRM and portals we already use? Usually, yes, because the agent connects to the systems that hold your records rather than replacing them. The details depend on what each system exposes, and the connection should start with read access. Reviewing the integration against an acceptance checklist before go-live avoids surprises.

What should we measure to know it is working? Choose evidence a manager can inspect without trusting the agent's own summary: how quickly new enquiries receive a correct first reply, whether each lead has an owner, and whether follow-ups carry a date and a source. Add the share of drafts that needed major edits as a quality signal, and use it to decide when to widen permissions.

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