Superintelligence in Real Estate Agencies: Leads, Viewings, Contracts

Superintelligence in Real Estate Agencies: Leads, Viewings, Contracts

Superintelligence in real estate agencies, in practice, is not a machine that decides on its own. It is a layer of AI agents with business memory, limited tools and human approval, designed to absorb each new model without rebuilding the process. The broker stays responsible for price, negotiation, exceptions and the relationship with whoever is buying, selling or renting.

Key takeaways

• The gain comes from the layer around the model (memory, rules, tools, logs and tests), not from picking the model of the moment.
• Agents handle the repetitive work of leads, viewings and documents. Pricing, negotiation and signature stay with people.
• Human approval should be designed by type of action, not applied to everything, so it neither stalls the operation nor lets sensitive actions through.
• Every new model enters only after passing the same test cases as the previous one, using the agency's own records as the reference.
• Personal data from leads, owners and tenants requires a legal basis, minimization and traceability under LGPD, Brazil's data protection law.

What does superintelligence mean in a real estate agency?

The term usually points to a philosophical debate about machines more capable than humans. For someone running an agency, what matters is more concrete: a set of agents that, together, consistently carry out tasks that today depend on individual memory, spreadsheets and scattered messages. If you want to separate the theoretical debate from what your company can actually run, read the difference between AGI, ASI and AI agents.

The central point is that model capability changes over time, and your agency's knowledge does not. How you handle a rental lead, which questions qualify a buyer, which documents each contract type requires, what an owner has authorized: that is your asset. If this knowledge lives inside a prompt pasted into one tool, it is stuck in that tool. If it lives in structured memory, with reviewed rules and examples, any better model can use it.

This separation is what makes the gain grow with each new model. The structure, block by block, is described in company superintelligence layer architecture. Here the focus is applying it to the daily work of leads, viewings and contracts.

Where do agents work across leads, viewings and contracts?

The safest design splits the work into three fronts, each with a narrow-scope agent. One agent that does everything is harder to test and harder to limit. Anthropic itself recommends starting with the simplest pattern that solves the problem and adding complexity only when needed, as described in its engineering article Building effective agents.

On the leads front, the agent replies on WhatsApp, identifies the interest (buying, renting, selling their own property), collects the minimum needed to qualify and records everything in the CRM. It does not promise commercial terms. If the conversation crosses an agreed line, such as asking for a discount or mentioning a competitor, it hands over to the broker with a summary ready.

On the viewings front, the agent proposes times based on the broker's calendar, confirms, sends a reminder and reschedules. After the viewing, it asks the lead what they thought and organizes the feedback for the broker. For messages outside an active conversation, the platform rules must be respected, as explained in the article on the WhatsApp 24 hour customer service window.

On the contracts front, the agent checks the document list, flags what is missing, prepares the draft from a template approved by legal and points out mismatches, such as a name that differs between a document and the registration. It does not change clauses and does not send anything for signature without approval.

How does memory make the gain grow with each new model?

Memory here is not just conversation history. There are four layers worth keeping separate:

• Business memory: the agency's rules, service tone, viewing policies, qualification criteria and what is never promised.
• Property memory: the current listing record, status, owner restrictions and required documents.
• Contact memory: interaction history, stated preferences and consents, with the minimum of personal data.
• Decision memory: corrections made by brokers, approved and rejected cases, and the reason.

The fourth layer is the most valuable and the most ignored. Every time a broker corrects an answer or rejects a draft, that correction becomes a test case. Over time, the agency builds its own set of real examples, reviewed by people who understand the business.

When a new model is released, you do not have to hope for the best. You run the set of cases with the current model and with the candidate, compare where each succeeds and fails, and decide whether to switch. If the new model is better at understanding ambiguous messages, the leads agent improves without anyone rewriting rules. If it is worse at some specific point, you find out before it reaches a customer. Anthropic details this evaluation reasoning in Demystifying evals for AI agents, and the article on mock tools and sandboxes for testing agents shows how to test without touching real data or customers.

There is also a context concern. Dumping the full history of the property and the contact into every call makes answers worse and raises cost. The better approach is to retrieve only what the task needs, a practice described in Anthropic's Effective context engineering for AI agents.

Which decisions stay with the broker?

A good test is to ask whether the decision involves judgment, relationship or responsibility. If it does, it stays with a person. Use this list as a decision guide:

• Pricing and valuation: the agent can gather comparison data, but the recommendation to the owner is the broker's.
• Negotiation: counteroffers, discounts, deadlines and payment terms are handled by people.
• Reading a sensitive situation: separation, inheritance, financial urgency or conflict between parties call for a human listener.
• Contract exceptions: any clause outside the approved template goes to the broker and to legal.
• Application or credit approval: decisions that affect someone's access to a property require human review and a recorded reason. See what LGPD provides in automated decisions and the right to review.
• Promises to the customer: key handover dates, renovations, neighborhood and any guarantee are given only by an authorized person.
• Closing and signature: the final acceptance is human, always.

The agent does the preparation and follow-up. The broker decides and takes responsibility. This division also protects the agency: it is clear, in the records, who approved what.

How do you design human approval without stalling the operation?

Requiring approval for everything creates a queue, and a queue makes brokers approve out of habit. Requiring it for nothing creates risk. The middle ground is to classify actions by impact:

• Free action: answering a question with information from the listing record, confirming a time already agreed, sending a viewing reminder.
• Quick approval action: sending an alternative time proposal, updating a lead's status, sending a customized document list.
• Mandatory approval action: sending a contract draft, changing contract data, sharing owner data, any message containing a commercial condition.
• Prohibited action: negotiating price, accepting an offer, promising a deadline or guarantee, signing.

Each action needs explicit permission in the agent's tools. The leads agent, for example, should not even have write access to the contracts module. Least-privilege authorization principles are covered in OWASP's authorization guidance, and tools should be written with few parameters and clear descriptions, as Anthropic advises in Writing effective tools for agents.

Another precaution: messages from leads are untrusted input. A text can try to instruct the agent to ignore its rules. Limiting tools and validating outputs reduces this risk, and OWASP's guidance on preventing prompt injection helps in thinking it through.

Illustrative scenario: from listing to contract

The scenario below is illustrative and does not represent any real client.

A mid-sized agency receives a WhatsApp message from someone interested in renting an advertised apartment. The leads agent checks the property record, answers about the rent and the conditions already published, asks about the move-in timeline and whether there are pets, and logs everything in the CRM. The person asks for a discount. The agent does not answer that, tells them a broker will continue the conversation and hands the broker a summary: interest, questions asked, discount request.

The broker negotiates personally. Once the viewing is agreed, the viewings agent proposes times from the calendar, confirms and sends a reminder. After the viewing, the tenant decides to rent. The contracts agent lists the documents, checks the ones received and prepares the draft from the approved template. The broker reviews, adjusts one date at the owner's request and approves sending. The parties sign.

At the end, the date correction goes into decision memory as a case for review. Months later, when the agency tests a new model, that case is in the evaluation set along with the others.

How to start: map, limit, pilot and hand over

Juan Carlo's working method follows four moves, applicable to an agency without exaggeration:

• Map: describe the current flow of one type of service (for example, residential rentals), including where leads arrive, who replies, where there is waiting and where there is rework. Record the starting point so you can compare later.
• Limit: define which data the agent sees, which actions it can take and which require approval. Include the legal basis for processing personal data and consult guidance from ANPD, Brazil's data protection authority, when in doubt.
• Pilot: choose one real flow, with records of every decision and a success metric agreed beforehand, such as time to first reply or confirmed viewings that actually happen. Run it with a few brokers.
• Hand over: deliver documentation, the set of test cases and a review routine. Then decide, based on the records, whether to expand, adjust or stop.

For those who prefer managed operation, this follow-up can be done with Be Human (be-human.ai). In any format, the agency should own the rules, the memory and the test cases.

A final checklist before switching on the pilot:

• Is there a written list of what the agent can and cannot do?
• Does each tool have minimum permission and a usage log?
• Is a broker responsible for reviewing pending approvals?
• Do leads know they are talking to an assistant and how to reach a person?
• Is there a set of real test cases, without unnecessary personal data?
• Do we know how to pause the agent quickly?

Frequently asked questions

Does superintelligence in real estate agencies replace the broker? No. It takes over the repetitive work of qualifying, scheduling, reminding and checking documents, and frees time for negotiation and relationships. Price, exceptions and closing stay with the person.

Why does the gain grow with each new model? Because the agency's knowledge lives in memory, rules and test cases, not tied to one specific model. When a better model appears, you test it with the same cases and, if it passes, the agent improves without rewriting the process.

Does LGPD limit the use of agents with data from leads and owners? Yes, and it should be part of the design from the start. You need a legal basis, collect the minimum necessary, inform the data subject, record what the agent did and provide for human review when a decision affects the person. If data is sent outside Brazil, also check the rules on international data transfer under LGPD.

Where should a small agency start? With a single flow, such as the first rental inquiry on WhatsApp, with human approval for any commercial condition and a record of each answer. After a few weeks of real use, the records show whether it is worth expanding to viewings and contracts.

Sources

• Anthropic: Building effective agents
• Anthropic: Demystifying evals for AI agents
• Anthropic: Effective context engineering for AI agents
• ANPD: Autoridade Nacional de Proteção de Dados